Computer-implemented method and forecasting system for generating forecasts regarding whether peptides of interest will have targeted interaction with immunoglobulin complex, as well as associated machine-readable storage medium

Abstract

FIELD: biotechnology.SUBSTANCE: discloses the use of an attention-based machine learning model for generating a result that includes at least one of an interaction prediction for a target interaction, an affinity interaction prediction, or an immunogenicity prediction relative to a target interaction for a corresponding immunoglobulin peptide-complex (IPC) combination. The target interaction may occur between the peptide and an immunogenicity complex (IPC), such as the major histocompatibility complex (MHC), the T cell receptor (TCR), or both.EFFECT: expands the arsenal of means for identifying, producing and / or using a pharmaceutical composition that contains one or more peptides for which one or more target interactions are predicted to be more likely, as well as for determining and / or using treatment methods that include the administration of such a pharmaceutical composition.22 cl, 24 dwg

Description

CROSS-REFERENCE TO A RELATED APPLICATION

[0001] This application claims priority to U.S. Provisional Patent Application No. 63 / 053,307, filed July 17, 2020, entitled "Attention-Based Neural Network for Predicting Peptide Binding, Presentation, and Immunogenicity." This application is related to its co-filed U.S. Patent Application No. 17 / 378,651, entitled "Attention-Based Neural Network for Predicting Peptide Binding, Presentation, and Immunogenicity." Both of these applications are incorporated by reference into the specification of this application in their entireties. TECHNICAL FIELD

[0002] This present disclosure generally relates to the use of machine learning models (e.g., that include an attention mechanism) to generate predictions regarding whether peptides of interest (e.g., a mutant peptide) will experience target interaction(s) with an immunoglobulin complex (IPC) (e.g., bind to an MHC molecule represented by an MHC molecule, bind to a TCR, etc.), the affinity associated with such target interaction(s), and / or the ability of the peptides to activate an immune response. This present disclosure also relates to compositions that include, and methods of using, certain mutant peptides (or related precursors or sequences) selected based on such predictions for treatment.BACKGROUND

[0003] Neoantigen vaccines are a relatively new approach to providing personalized cancer treatment. Neoantigens are tumor-specific antigens derived from somatic mutations in tumors and presented by the subject's cancer cells and antigen-presenting cells.

[0004] Neoantigen vaccines can prime a subject's T cells to recognize and fight cancer cells expressing one or more specific tumor neoantigens. This approach generates a tumor-specific immune response that bypasses healthy cells while targeting tumor cells. However, there is high variability between subjects regarding the neoantigens produced by a subject's tumor cells and expressed by the subject's major histocompatibility complex (MHC) molecules. Thus, a personalized vaccine can be developed and administered to a specific subject. A personalized vaccine can be developed or selected based on a subject's specific tumor profile.A tumor profile can be determined by determining DNA and / or RNA sequences from a subject's tumor cells and using the sequences to identify antigens that are present on tumor cells but absent from normal cells.

[0005] In many cases, the bulk of mutant sequences found in tumor cells correspond to neoantigens that are not actually present on the surface of tumor cells. Such neoantigens would be poor candidates for a personalized vaccine. For example, a detected peptide sequence may identify amino acids in a mutant peptide that is produced intracellularly but does not bind to and / or is not presented (on the cell surface) by an MHC-I or MHC-II molecule. Alternatively, a mutant peptide capable of being presented by an MHC-I or MHC-II molecule may not be produced intracellularly. In either case, the mutant peptide will not activate an immunological response, for example, by a CD8+ cytotoxic T lymphocyte in the case of an MHC-I molecule, or by a CD4+ helper T cell in the case of an MHC-II molecule.

[0006] Thus, sequence analysis for identifying neoantigen candidates for a vaccine that focuses solely on detecting mutant peptide sequences or predicting which mutant peptide sequences will trigger a single biological interaction (e.g., whether a peptide will bind to a molecule) can generate numerous false positives. This type of sequence analysis will be ineffective in developing personalized vaccines designed to prime immunological responses.

[0007] Thus, it may be desirable to predict which neoantigens are presented by given tumor cells of a subject and / or for which a vaccine containing the neoantigen will activate a strong immunological response.SUMMARY OF THE INVENTION

[0008] The invention provides a computer-implemented method for generating predictions regarding whether peptides of interest will experience a target interaction with an immunoglobulin complex (IPC). The method according to the invention includes accessing a set of peptide sequences characterizing a set of peptides, wherein each peptide sequence from the set of peptide sequences was identified by processing a disease sample from a subject. The method also includes accessing an immunoglobulin complex (IPC) sequence identified for an immunoglobulin complex (IPC) of the subject, wherein the IPC comprises a major histocompatibility complex (MHC).The method also includes processing a set of peptide representations that represents a set of peptide sequences using a first attention unit in an initial attention subsystem of an attention-based machine learning model, and an immunoglobulin complex (IPC) representation that represents an IPC sequence using a second attention unit in the initial attention subsystem, to generate a result, wherein the result includes at least one of an interaction prediction or an interaction affinity prediction for a corresponding peptide-IPC combination, wherein the corresponding peptide-IPC combination includes a peptide from the set of peptides and an MHC, the interaction affinity prediction for the corresponding peptide-IPC combination predicts the binding affinity between the peptide and the MHC, and the interaction prediction for the corresponding peptide-IPC combination predicts whether the MHC will present the peptide on the cell surface.Finally, the method includes generating a report based on the result.

[0009] Another object of the invention is a prediction system for generating predictions regarding whether peptides of interest will experience a target interaction with an immunoglobulin complex (IPC). The system according to the invention comprises one or more data processing devices and a non-transitory computer-readable storage medium containing instructions, the execution of which on one or more data processing devices causes the one or more data processing devices to perform the method according to the invention or other methods described below.

[0010] The subject of the invention is also a machine-readable information carrier containing instructions, the execution of which on one or more data processing devices causes one or more data processing devices to perform the method proposed in the invention or other methods described below.

[0011] The invention can also be implemented in a vaccine that comprises one or more peptides; a plurality of nucleic acids that encode one or more peptides; or a plurality of cells expressing one or more peptides. The one or more peptides are selected from a set of peptides based on a report generated by part or all of one or more methods disclosed herein. The one or more peptides are a partial subset of the set of peptides.

[0012] The invention can also be implemented in a method for producing a vaccine. The method includes producing a vaccine comprising: one or more peptides; a plurality of nucleic acids that encode one or more peptides; or a plurality of cells expressing one or more peptides. The one or more peptides are selected from a set of peptides based on a report generated by part or all of one or more methods disclosed herein. The one or more peptides are a partial subset of the set of peptides.

[0013] The invention can also be implemented in a pharmaceutical composition that comprises one or more peptides selected from a set of peptides based on a report generated by part or all of one or more of the methods disclosed herein. The one or more peptides are a partial subset of the set of peptides.

[0014] The invention can also be implemented in a pharmaceutical composition that contains a nucleic acid sequence that encodes one or more peptides that were selected from a set of peptides based on a report generated by part or all of one or more of the methods disclosed herein. The one or more peptides are a partial subset of the set of peptides.

[0015] The invention may also be embodied in an immunogenic peptide that is identified based on a report generated by part or all of one or more of the methods disclosed herein.

[0016] The invention may also be embodied in a nucleic acid sequence that is identified based on a report generated by part or all of one or more of the methods disclosed herein.

[0017] The invention can also be implemented in a method for treating a subject. The method comprises administering at least one of one or more peptides, one or more pharmaceutical compositions, or one or more nucleic acid sequences identified based on a report generated by part or all of one or more of the methods disclosed herein.

[0018] The invention can also be implemented in a method comprising processing a set of biological samples obtained from a subject to generate a set of peptide sequences characterizing the set of peptides. The method comprises processing a set of biological samples obtained from a subject to generate an immunoglobulin complex (IPC) sequence identified for the subject's immunoglobulin complex (IPC). The method comprises generating a set of peptide representations that represents the set of peptide sequences using a first attention unit in an initial attention subsystem of an attention-based machine learning model. The method comprises generating a representation of the immunoglobulin complex (IPC) that represents the IPC sequence using a second attention unit in the initial attention subsystem.The method includes processing a set of peptide representations and an IPC representation to generate a result, wherein the result includes at least one of an interaction prediction, an interaction affinity prediction, or an immunogenicity prediction for a corresponding peptide-IPC combination, a corresponding peptide-IPC combination containing a peptide from the set of peptides.

[0019] The invention can also be implemented in a method comprising receiving a request for the development of a personalized vaccine on a user device. The method comprises transmitting a communication from the user device to a remote system, wherein the communication contains a subject identifier.The remote system is configured to: access a set of peptide sequences characterizing a set of peptides, wherein each peptide sequence from the set of peptide sequences was determined by processing a disease sample from a subject; access an immunoglobulin complex (IPC) sequence identified for the subject's immunoglobulin complex (IPC); and process a set of peptide representations that represents the set of peptide sequences using a first attention unit in an initial attention subsystem of an attention-based machine learning model, and a representation of the immunoglobulin complex (IPC) that represents the IPC sequence using a second attention unit in the initial attention subsystem, to generate a result. The result comprises at least one of an interaction prediction, an interaction affinity prediction, or an immunogenicity prediction for the corresponding peptide-IPC combination.The remote system is configured to generate a report based on the result and transmit the report to the user device. The method includes receiving the report on the user device.

[0020] The invention can also be implemented in a method for producing a means for treating a subject. The method includes receiving a report from a computing device.The computing device is configured to access a set of peptide sequences characterizing a set of peptides, wherein each peptide sequence from the set of peptide sequences was determined by processing a disease sample from a subject; access an immunoglobulin complex (IPC) sequence identified for the subject's immunoglobulin complex (IPC); and process a set of peptide representations that represents the set of peptide sequences using a first attention unit in an initial attention subsystem of an attention-based machine learning model, and a representation of the immunoglobulin complex (IPC) that represents the IPC sequence using a second attention unit in the initial attention subsystem, to generate a result. The result comprises at least one of an interaction prediction, an interaction affinity prediction, or an immunogenicity prediction for the corresponding peptide-IPC combination.The computing device is configured to generate a report based on the result. The method further comprises generating a treatment agent production plan for producing the treatment agent based on the report.

[0021] The invention can also be implemented in a method comprising introducing a plurality of variant-coding sequences characterizing a plurality of mutant peptides into an attention-based machine learning model, wherein each variant-coding sequence of the plurality of variant-coding sequences was identified by processing a disease sample from a subject. The method comprises introducing an immunoglobulin complex (IPC) sequence identified for the subject's immunoglobulin complex (IPC) into the attention-based machine learning model.An attention-based machine learning model is configured to process a plurality of variant representations, which represent a plurality of variant-encoding sequences, using a first attention unit in an initial attention subsystem of the attention-based machine learning model, and an immunoglobulin complex (IPC) representation, which represents the IPC sequence, using a second attention unit in the initial attention subsystem, to generate a result. The result comprises at least one of an interaction prediction, an interaction affinity prediction, or an immunogenicity prediction for the corresponding mutant peptide-IPC combination. The method includes receiving a report generated based on the result; and selecting, based on the report, a subset of the plurality of mutant peptides for use in treating a subject.

[0022] The invention can also be implemented in a method comprising obtaining a peptide sequence that characterizes a mutant peptide, wherein the peptide sequence comprises a variant relative to a corresponding reference sequence; obtaining an MHC sequence identified for a major histocompatibility complex (MHC); processing the peptide sequence and the MHC sequence using different processing paths in an attention-based machine learning model to generate a result, wherein the result provides information about immunological activity relative to both the mutant peptide and the MHC; and generating a report based on the result.

[0023] The invention can also be implemented in a method comprising obtaining a peptide sequence that characterizes a mutant peptide, wherein the peptide sequence comprises a variant relative to a corresponding reference sequence; obtaining a TCR sequence identified for a T-cell receptor (TCR); processing the peptide sequence and the TCR sequence using different processing paths in an attention-based machine learning model to generate a result, wherein the result provides information about immunological activity relative to both the mutant peptide and the TCR; and generating a report based on the result.

[0024] Some embodiments of the present disclosure include a system comprising one or more data processing devices. In some embodiments, the system comprises a non-transitory computer-readable storage medium containing instructions that, when executed on one or more data processing devices, cause the one or more data processing devices to perform a part or all of one or more methods and / or a part or all of one or more techniques disclosed herein. Some embodiments of the present disclosure include a computer program product that is tangibly embodied in a non-transitory computer-readable storage medium containing instructions configured to cause one or more data processing devices to perform a part or all of one or more methods and / or a part or all of one or more techniques disclosed herein.

[0025] The terms and expressions that have been used are used as terms of description and not of limitation, and there is no intention in the use of such terms and expressions to exclude any equivalents of the features shown and described or parts thereof, but it is recognized that various modifications are possible within the scope of the claimed invention. It should therefore be understood that although the present invention, as claimed, has been specifically disclosed by means of embodiments and optional features, modifications and variations of the concepts disclosed herein may be resorted to by those skilled in the art, and that such modifications and variations are considered to be within the scope of the present invention as defined in the appended claims. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The present disclosure is described in conjunction with the attached figures:

[0027] Fig. 1 is a block diagram of a prediction system according to various embodiments.

[0028] Fig. 2 is a flow chart of a method for generating predictions using a machine learning model according to one or more embodiments.

[0029] Fig. 3 is a schematic diagram of one configuration for the machine learning model of Fig. 1 according to one or more embodiments.

[0030] Fig. 4A is a schematic diagram of a machine learning model 400 according to one or more embodiments.

[0031] Fig. 4B is a schematic diagram of another configuration for a machine learning model 400 according to one or more embodiments.

[0032] Fig. 4B is a schematic diagram of another configuration for a machine learning model 400 according to one or more embodiments.

[0033] Fig. 5 is a schematic diagram of an attention unit 500 according to one or more embodiments.

[0034] Fig. 6 is a flow chart of a method for processing a sequence representation using a typical self-attention layer according to one or more embodiments.

[0035] Fig. 7 is a schematic diagram showing the method 600 described in Fig. 6 above, according to one or more embodiments.

[0036] Fig. 8 is a flow chart of a method for generating information on the immunological activity of various peptides.

[0037] Fig. 9 is a flow chart of a method for generating information on the immunological activity of various peptides.

[0038] Fig. 10 is a flow chart of a method for training a machine learning model and using the trained machine learning model to generate predictions regarding peptides and MHC, according to one or more embodiments.

[0039] Fig. 11 is an illustration that contains a training data table according to one or more embodiments.

[0040] Fig. 12 is an illustration of a candidate neoantigen and corresponding possible candidate neoepitopes according to one or more embodiments.

[0041] Fig. 13 is a flow chart of a method for training a machine learning model and using the trained machine learning model to generate predictions regarding peptides and TCRs, according to one or more embodiments.

[0042] Figs. 14A, 14B, and 14C are graphs of exemplary precision-response (PR) curves according to one or more embodiments.

[0043] Fig. 15 is a graph comparing exemplary mean accuracy values ​​of the elution ligand results of model A and model P-MHC-I for each allele in a test data set according to one or more embodiments.

[0044] Fig. 16A and 16B are graphs that compare the performance of the P-MHC-I model on a human data set with the performance of the P-MHC-I model on a mouse data set according to one or more embodiments.

[0045] Figs. 17A and 17B are graphs that compare the performance of the R-MHC-II model to the C model with respect to representation data according to one or more embodiments.

[0046] Figs. 18A and 18B are graphs that compare the performance of the R-MNS-P model to the C model, respectively, on a set of evasion data according to one or more embodiments.

[0047] Fig. 19 is a graph showing a comparison of the average accuracy per genotype for the P-MHC-II model with the C model on a test data set according to one or more embodiments.

[0048] Fig. 20 is a graph of receiver operating characteristic (ROC) curves that shows the performance of the P-MHC-I model (EL result), model A (EL result), and model B (BA result) relative to the CD 8 multimer assay data (first set of immunogenicity test data) according to one or more embodiments.

[0049] Fig. 21A-G are graphs that show the performance of the P-MHC-I model (E1 result), model A (EL result), and model B (BA result) relative to ELISpot assays (first set of immunogenicity test data) according to one or more embodiments.

[0050] Fig. 22A-G are graphs that show the performance of model A (BA result), model A (EL result), model C (BA result), and model P-MHC-I (EL result), respectively, according to one or more embodiments.

[0051] Fig. 23 is an illustration of a graph comparing ROC curves for model A (EL result), model B (BA result), and model P-MHC-I (EL result) using TESLA multimer analysis data according to one or more embodiments. Fig. 24 is a block diagram of a computer system in accordance with various embodiments of the invention

[0052] In the attached figures, similar components and / or features may have the same reference label. In addition, different components of the same type may be distinguished by following the reference label along the dotted line and a second label that differs among similar components. If only the first reference label is used in the description, the description is applicable to any of the similar components having the same first reference label, regardless of the second reference label. DETAILED DESCRIPTION OF THE SUMMARY OF THE INVENTION I. Overview

[0053] Recognizing the importance of being able to predict which mutant peptides (e.g., neoantigens) should be selected as candidates for a personalized vaccine, the embodiments described herein provide methods and systems for obtaining such predictions more accurately than various currently available methods and systems. The embodiments described herein utilize machine learning methods and systems to improve prediction performance by, for example, but not limited to, reducing the number of false signal recognitions generated during the analysis of sequences that characterize mutant peptides to determine the viability of these mutant peptides as vaccine candidates.

[0054] For example, embodiments described herein provide a machine learning model and various methods for applying the machine learning model and / or a result generated by the machine learning model to analyze sequences identified from a disease sample from a subject. To predict whether a mutant peptide detected in the disease sample interacts with a major histocompatibility complex (MHC) molecule (e.g., MHC-I, MHC-II), to predict the extent to which the mutant peptide interacts with the MHC molecule, or both, the machine learning model initially processes a representation of a sequence characterizing the mutant peptide separately from processing a representation of the MHC sequence corresponding to the MHC molecule. The sequence characterizing the mutant peptide may be referred to as a variant-coding sequence.The MHC sequence may consist of at least a portion of the complete sequence of an MHC molecule (e.g., the complete sequence, a pseudo-sequence of an MHC molecule that is the portion that interacts with the peptide-binding pocket, some other portion that contains the pseudo-sequence, etc.).

[0055] A machine learning model contains various processing subsystems. A machine learning model may contain, for example, a representation subsystem, a presentation attention subsystem, a composite subsystem, a composite attention subsystem, and a result subsystem. Each "subsystem" may consist of one or more blocks, with each block consisting of one or more subblocks and / or layers. A subblock may consist of any number of layers (or elements).

[0056] The representation subsystem can be used to generate a peptide representation of a peptide sequence (which may contain a variant coding sequence) and an MHC representation of the MHC sequence. The representation attention subsystem is used to process the peptide sequence representation independently or separately from (e.g., in parallel) the MHC sequence representation. These two parallel processing paths can be configured similarly or differently, but each contains at least one attention mechanism. Processing the peptide and MHC sequence representations through these parallel processing paths improves the predictive performance of the machine learning model.

[0057] Furthermore, the embodiments described herein recognize and take into account that training a model corresponding to a number of biological events may require significantly more data than training a model corresponding to a single biological event. Training a model for sequence analysis can be particularly challenging due to the large number of potentially observed sequences. Not only these millions of potential neoantigens, but also the genes encoding proteins for MHC class I molecules, for example, are also highly polymorphic: there are approximately 20,000 alleles for human MHC class I. Thus, the embodiments described herein provide methods and systems for training a machine learning model that reduce training complexity and improve training performance.For example, variant-coding sequences used for training can be selected and / or adjusted so that training is performed using variant-coding sequences with amino acid lengths at or below a threshold amino acid length (e.g., 14 amino acids). Generating a training dataset that contains variant-coding sequences with lengths equal to or shorter than the threshold amino acid length can reduce the overall training complexity and improve training and / or prediction performance (e.g., by reducing the variability in parameter performance at a given time, thereby improving prediction performance).

[0058] Therefore, the techniques disclosed herein include machine learning-based approaches for generating predictions regarding the immunological activity associated with a peptide, such as a mutant peptide. A machine learning model is provided that generates a result containing one or more predictions. The result may, for example, generate one or more interaction predictions, one or more interaction affinity predictions, one or more immunogenicity predictions, or a combination thereof. An interaction prediction may comprise a prediction regarding whether a peptide (e.g., a mutant peptide, including a given ordered set of amino acids, as identified by a given variant coding sequence) experiences one or more target interactions.A target interaction may be, for example, binding to an IPC (e.g., an MHC molecule, a TCR) presented by an MHC molecule on the cell surface, or another type of target interaction. An interaction affinity prediction may include an affinity prediction for one or more target interactions. For example, an interaction affinity prediction may indicate the binding affinity relative to peptide-MHC binding. The affinity of an interaction (e.g., binding) may be determined based on the direction, strength, and / or stability of the interaction (e.g., binding).

[0059] Furthermore, the result may include or indicate the immunogenicity of the peptide. For example, the result may predict whether the peptide will activate an immune response in a specific subject or group of subjects. These predictions can be generated for each of a plurality of mutant peptides, and the predictions can be used to select one or more mutant peptides for inclusion in a vaccine and / or use in treatment. For example, but not limited to, mutant peptides associated with high predicted binding affinity, high probability of presentation on tumor cell surfaces, and / or high predicted immunogenicity can be selected for inclusion in a vaccine or use in treatment.

[0060] Embodiments described herein provide methods and systems for using an attention-based machine learning model to generate predictions about immunological activity regarding peptides and immunoglobulin complexes (IPCs). An IPC can be an MHC or TCR. A set of peptide sequences characterizing a set of peptides can be accessed, wherein each peptide sequence from the set of peptide sequences was determined by processing a disease sample from a subject. An immunoglobulin complex (IPC) sequence for the subject's immunoglobulin complex (IPC) can be identified.A set of peptide representations, representing a set of peptide sequences, is processed using the first attention unit in the initial attention subsystem of the attention-based machine learning model and the immunoglobulin complex (IPC) representation, which represents the IPC sequence, using the second attention unit in the initial attention subsystem to generate a result. The result contains at least one of an interaction prediction, an interaction affinity prediction, or an immunogenicity prediction for the corresponding peptide-IPC combination. A report is generated based on the result.

[0061] The description below provides illustrative implementations of these methods and systems and pathways in which the report that is generated can be used to plan, design, and / or manufacture a treatment.II. Predictions regarding immunological activity involving mutant peptides using attention-based machine learning modelsII.A Overview

[0062] Referring now to the figures, Fig. 1 is a block diagram of a prediction system 100 according to various embodiments. The prediction system 100 is used to generate predictions regarding the immunological activity of peptides and, in particular, mutant peptides. The prediction system 100 comprises a computing platform 102, a data array 104, and a display system 106. The computing platform 102 can take various forms. In one or more embodiments, the computing platform 102 comprises a single computer (or computer system) or a plurality of computers in communication with each other. In other examples, the computing platform 102 takes the form of a cloud computing platform.

[0063] Each of the data array 104 and the display system 106 is in communication with the computing platform 102. In some examples, the data array 104, the display system 106, or both can be considered as part of, or otherwise integrated into, the computing platform 102. Thus, in some examples, the computing platform 102, the data array 104, and the display system 106 can be separate components in communication with each other, and in other examples, some combination of these components can be combined together. Communication between the various components can be implemented using any number of wired communication lines, wireless communication lines, optical communication lines, or a combination thereof.

[0064] The prediction system 100 comprises a sequence analyzer 108, which may be implemented using hardware, software, firmware, or a combination thereof. In one or more embodiments, the sequence analyzer 108 is implemented in the computing platform 102. The sequence analyzer 108 receives sequence data 110 for processing. For example, the sequence data 110 may be sent as input data to the sequence analyzer 108, obtained from a data array 104 or some other type of storage (e.g., cloud storage), obtained from a cloud storage, or obtained in some other way. In some cases, the sequence data 110 may be obtained from the data array 104 in response to receiving user input data entered by the user through an input device.

[0065] The sequence data 110 may be generated from processing a set of samples 112. The set of samples 112 may take the form of one or more biological samples from one or more subjects (e.g., samples from a subject with a disease, a sample from a healthy subject, a combination thereof). The set of samples 112 may comprise a sample obtained from a tumor of the subject. The tumor may be a manifestation of, for example, lung cancer, melanoma, breast cancer, ovarian cancer, prostate cancer, kidney cancer, gastrointestinal cancer, colon cancer, testicular cancer, head and neck cancer, pancreatic cancer, brain cancer, B-cell lymphoma, acute myeloid leukemia, chronic myelogenous leukemia, chronic lymphocytic leukemia, T-cell lymphocytic leukemia, non-small cell lung cancer, small cell lung cancer, or a combination thereof.

[0066] A sample in sample set 112 may contain, for example, various immunoglobulin complex (IPC) molecules and various peptides, or a combination thereof. When sample set 112 contains a disease sample, the peptides may contain one or more mutant peptides (e.g., neoantigens). The IPC molecules may include, for example, various MHC molecules, various TCR molecules, or a combination thereof.

[0067] In one or more embodiments, the sample set 112 includes an immunoglobulin complex (IPC) 114 (e.g., an MHC class I molecule, an MHC class II molecule, a TCR, etc.) and an amino acid chain 116. The amino acid chain 116 may be an amino acid chain that comprises a peptide 118, an N-flank 120, and a C-flank 122. The peptide 118 may be defined as including or excluding the N-terminus between the peptide 118 and the N-flank 120 and as including or excluding the C-terminus between the peptide 118 and the C-flank 122. The peptide 118 is considered to be a mutant peptide when the peptide 118 contains one or more variants (e.g., one or more sequence variants) compared to the corresponding reference sequence. In some embodiments, the sample set 112 also comprises an immunoglobulin complex 123 (e.g., an MHC class I molecule, an MHC class II molecule, a TCR, etc.).

[0068] The set of samples 112 can be processed to generate sequence data 110. In some embodiments, in which multiple samples in the set of samples 112 can be processed at different times. In some embodiments, the prediction system 110 includes a sample analyzer that is used in processing the set of samples 112 to generate sequence data 110. The sequence data 110 comprises, for example, at least one immunoglobulin complex (IPC) sequence 124 (for example, one IPC sequence 124 corresponding to the immunoglobulin complex 114) and at least one peptide sequence 126 (for example, one peptide sequence 126 corresponding to the peptide 118).The sequence data 110 may also include at least one N-flanking sequence 128 (e.g., one N-flanking sequence 128 corresponding to N-flank 120), at least one C-flanking sequence 130 (e.g., one C-flanking sequence 130 corresponding to C-flank 122), or both, which correspond to the corresponding peptide sequence 126.

[0069] When the immunoglobulin complex 114 takes the form of an MHC, the IPC sequence 124 can be, for example, an MHC sequence that characterizes at least a portion of the MHC.When the immunoglobulin complex 114 takes the form of a TCR, the IPC sequence 124 can be, for example, a TCR sequence that characterizes at least a portion of the TCR.In still other embodiments, the IPC sequence 124 can include both a TCR sequence and an MHC sequence that characterize at least a portion of a TCR molecule and at least a portion of an MHC molecule, which can represent a peptide for the TCR molecule, respectively.In some embodiments, the sequence data 110 may include an IPC sequence 124 as an MHC sequence characterizing at least a portion of an immunoglobulin complex 114 as an MHC, as well as a separate TCR sequence 131 characterizing at least a portion of a TCR (e.g., an immunoglobulin complex 123) in the set of samples 112.

[0070] Peptide sequence 126 characterizes at least a portion of peptide 118. N-flank sequence 128 characterizes at least a portion of N-flank 120. For example, since the number of amino acids (or amino acid residues) above the N-terminus may be large, the corresponding sequence for N-flank 120 can be adjusted to generate N-flank sequence 128. C-flank sequence 130 characterizes at least a portion of C-flank 122. In some cases, when the number of amino acids (or amino acid residues) below the C-terminus is large, the corresponding sequence for C-flank 122 can be adjusted to generate C-flank sequence 130.

[0071] The sequence analyzer 108 receives the sequence data 110 as input for processing. The sequence analyzer 108 includes a machine learning model 132 that processes the sequence data 110. In some embodiments, the sequence analyzer 108 is sent directly to the machine learning model 132 for processing. In other embodiments, the sequence analyzer 108 pre-processes the sequence data 110 before sending the sequence data 110 to the machine learning model 132 for processing.

[0072] The machine learning model 132 may be implemented in any number of different ways. In one or more embodiments, the machine learning model 132 takes the form of an attention-based machine learning model. The machine learning model 132 may be used in either a training mode or a prediction mode. In the training mode, the machine learning model 132 is trained using a training data set 133. Examples of data that may form a training data set are described further below in Section II.D. The machine learning model 132 is trained so that it can be used in a prediction mode.

[0073] Machine learning model 132 processes IPC sequence 124 via IPC processing pathway 134 and peptide sequence 126 via peptide processing pathway 136. Separating these two pathways for IPC and peptide provides improved prediction performance of machine learning model 132. In some embodiments, machine learning model 132 additionally processes N-flank sequence 128 via N-flank processing pathway 138, C-flank sequence 130 via C-flank processing pathway 140, or both.

[0074] The IPC processing pathway 134 may consist of one or more different pathways. For example, in some cases, the IPC processing pathway 134 takes the form of an MHC processing pathway for processing, for example, the IPC sequence 124 as an MHC sequence. In other cases, the IPC processing pathway 134 comprises a TCR processing pathway for processing, for example, the IPC sequence 124 as a TCR sequence. In still other cases, the IPC processing pathway 134 includes a processing pathway for processing the IPC sequence 124, which contains both an MHC sequence and a TCR sequence. In some embodiments, when the IPC processing pathway 134 takes the form of an MHC processing pathway, the machine learning model 132 also comprises a TCR processing pathway 142 for processing, for example, the TCR sequence 131. Examples of implementations of these different processing pathways are described in more detail below.

[0075] The machine learning model 132 processes the sequence data 110 to generate a result, which is used to generate the report 144. The report 144 may contain the exact result of the machine learning model 132, may contain a transformed or filtered version of the result, or both. In some cases, the sequence analyzer 108 may generate notifications, recommendations, alerts, or other information based on the result of the machine learning model 132, and this additional information is included in the report 144.

[0076] Report 144 may be a result that includes, for example, information on the immunological activity of interest regarding one or more peptides (e.g., one or more mutant peptides). For example, report 144 may contain information on the immunological activity regarding peptide 118 and immunoglobulin complex 114 (e.g., MHC), peptide and immunoglobulin complex 123 (e.g., TCR), or both. Report 144 may contain, for example, interaction information 146 and immunogenicity information 148, or both. Interaction information 1346 may provide predictions regarding a selected set of interactions between peptide 118 and immunoglobulin complex 114, between peptide 118 and immunoglobulin complex 123, or both. The immunogenicity information 148 can provide predictions regarding the immunogenicity of peptide 118.

[0077] In one or more embodiments, the report 144 may be displayed on the user graphical interface 150 on the display system 106. The user may view the report 144 and / or interact with the report 144 via the user graphical interface 150 and use the report 144 to make a decision regarding the treatment of the subject from whom at least one set of samples 112 was obtained (or collected).

[0078] In some embodiments, the prediction system 100 sends the report 144 to a remote system 152 (e.g., wirelessly). The remote system 152 may be a cloud computing platform, cloud storage, another computing system, a user device (e.g., a smartphone, tablet, laptop, etc.), or some other type of platform. In some embodiments, the remote system 152 may be a treatment manufacturing system (or machine) or a part thereof.

[0079] Fig. 2 is a flow chart of a method for generating predictions using a machine learning model according to one or more embodiments. Method 200 can be implemented using the prediction system 100 described in Fig. 1. For example, method 200 can be implemented using sequence analyzer 108 and machine learning model 132 in Fig. 1.

[0080] The method 200 may include, for example, step 202. Step 202 includes training an attention-based machine learning model using a training data set that contains training peptide sequence data, training immunoglobulin complex (IPC) data, and training immunological activity data.

[0081] Step 204 includes accessing a set of peptide sequences characterizing a set of peptides, wherein each peptide sequence in the set of peptide sequences was determined by processing a disease sample from a subject.

[0082] Step 206 involves accessing the sequence of the immunoglobulin complex (IPC) identified for the immunoglobulin complex (IPC) of the subject.

[0083] Step 208 includes processing a set of peptide representations that represents a set of peptide sequences using the first attention block in the initial attention subsystem of the attention-based machine learning model, and a representation of an immunoglobulin complex (IPC) that represents an IPC sequence using the second attention block in the initial attention subsystem, to generate a result, wherein the result includes at least one of an interaction prediction, an interaction affinity prediction, or an immunogenicity prediction for the corresponding peptide-IPC combination. The first attention block is independent of the second attention block.

[0084] Step 210 involves generating a report based on the result. The report can be used to facilitate the design and / or production of a treatment and / or treatment plan. For example, the report can identify a subset of peptides from a set of peptides or provide guidance on which of the subset of peptides to select for use in creating a treatment for a subject. The treatment can be, for example, a subset of peptides, a precursor for each of the subset of peptides, or some other form. II.B. Typical Machine Learning Model Architecture II.B. 1 General Characteristics and Implementation Considerations

[0085] As described above, in various embodiments, the machine learning model of the embodiments described herein, for example, the machine learning model 132, may be an attention-based machine learning model (e.g., which comprises one or more attention layers). The machine learning model 132 may implement, for example, one or more self-attention layers. The machine learning model 132 may use a self-attention mechanism, a global attention mechanism, a soft attention mechanism, a local attention mechanism, and / or a hard attention mechanism.

[0086] In some cases, an attention-based machine learning model may be configured to learn alignments (e.g., between a peptide sequence and an MHC sequence). Alignments may be learned and performed using an attention-based alignment metric function, such as, for example, a content-based function, an additive function, a location-based function, a dot product function, and / or a scaled dot product function. Machine learning model 132 may comprise one or more encoders, one or more transformers, and / or one or more encoders of transformers. In some embodiments, machine learning model 132 may use one or more features (such as, for example, one or more encoders), as described in Vaswani, A, et al., "Attention is All You Need." 31 st Conference on Neural Information Systems, http: / / papers.nips.cc / paper / 7181-attention-is-all-you-need.pdf, 2017.

[0087] The machine learning model 132 may comprise one or more encoders configured to, for example, transform input data (e.g., a sequence representation representing, for example, an amino acid sequence, a nucleic acid sequence, a codon sequence, etc.) into a more multidimensional space. The encoder may be a transformer encoder. The encoder may be configured to implement an attention-based technique and / or to include one or more attention layers (e.g., one or more self-attention layers).

[0088] In some embodiments, the machine learning model 132 may use or may exclude a convolutional layer, a long-short-term memory unit, a recurrent structure, and / or a recurrent component. For example, in some cases, the machine learning model 132 does not contain any convolutional layer, any recurrent structure, any long-short-term memory (LSTM) unit, and / or any recurrent component. In some cases, the machine learning model 132 is not a recurrent machine learning model and / or does not contain a recurrent neural network. In some cases, the machine learning model contains a recurrent neural network and / or may use position encoding to provide temporal information for one or more sequences. In some cases, the machine learning model 132 is not a convolutional machine learning model and / or does not contain a convolutional neural network. Machine learning model 132 may comprise a plurality of subsystems (or subnetworks). Each of the plurality of subsystems may include an encoder, a transformer encoder, one or more attention layers, and / or one or more self-attention layers. Machine learning model 132 may comprise attention units, with a first attention unit used to process the peptide representation independently of a second attention unit used to process the IPC representation of a portion or the entire IPC sequence (e.g., an MHC pseudosequence). The independence of these attention units may facilitate parallel processing when using the machine learning model. Furthermore, independence may improve the performance (e.g., prediction accuracy) of machine learning model 132.

[0090] In the machine learning model 132, the attention-based mechanisms can be configured such that the value of the output at any given layer depends not only on the corresponding value of the input data, but also on one, several, or all other values ​​of the input data. Thus, the machine learning model 132, the loss function, and / or the optimization function can be configured to optimize the output corresponding to one position representing the degree to which a given MHC molecule (represented by the corresponding input data) will bind to and / or represent a given peptide (represented by other corresponding input data) and / or cause immunogenicity in response to a given peptide. In some cases, any of the plurality of outputs of the converter encoders can represent such a probability of occurrence and / or the model can be trained accordingly.In some cases, an endpoint (e.g., a redundant endpoint), such as a sequence start element, may represent (in response to training) the probability of binding, representation, and / or immunogenicity. The resulting summaries may, for example, be fed to another layer and / or another attention subsystem or unit (e.g., one that contains an attention layer and / or a self-attention layer and / or one that is a transducer encoder and / or encoder).

[0091] In some cases, one, two, or all of the output dimensions from another layer and / or from another attention subsystem or unit have the same dimension as the input data fed to the other layer and / or another attention subsystem or unit. In some cases, the input data fed to this other layer and / or another attention subsystem or unit has a length along one axis that is greater than or equal to the sum of the number of amino acids in the IPC sequence, the number of amino acids in the peptide sequence, and the potential number of amino acids in one or more of the N-flank and C-flank. In some cases, the length of the input data dimension is longer than the total number of amino acids. The length of the input data along one axis may exceed the summed amino acid count when, for example, an additional feature vector (e.g., a sequence start feature vector) is appended to the amino acid-specific feature values.A different input size may include a number of features (e.g., defined via a hyperparameter). The output generated by another layer and / or another subsystem or attention unit may have the same size as the input.

[0092] A subset of the output values ​​generated by another layer and / or another subnetwork may be further processed by another neural network (e.g., a fully connected feedforward network). The subset of values ​​may include a 1-dimensional vector of values, which may correspond to a single set of feature values. The 1-dimensional vector may correspond to the feature values ​​associated with the sequence onset feature vector.

[0093] The neural network in the machine learning model 132 can be configured to produce one or more results. The one or more results can include, for example, a numerical result, a binary result, and / or a categorical result. Each of the one or more results can predict whether and / or the extent to which the IPC and the peptide undergo a particular type of reaction (e.g., are bound together). The machine learning model 132 can comprise one or more activation layers to produce a target time result (e.g., to convert an intermediate real number value into a binary and / or categorical result). The machine learning model 132 can be trained to generate multiple types of predictions (e.g., interaction predictions, interaction affinity predictions, and / or immunogenicity predictions). In some cases, the prediction can be binary or categorical. Other predictions can be non-binary or non-categorical.For example, the forecast can be scalar.

[0094] The machine learning model 132 may include and / or be contained in a complex model. The complex model may contain multiple (e.g., identical) submodels that can be trained using different portions of the training dataset. II.B.2. Typical Configurations for a Machine Learning Model

[0095] Fig. 3 is a schematic diagram of one configuration for the machine learning model 132 of Fig. 1 according to one or more embodiments. The machine learning model 132 is described by continuing reference to Fig. 1. The machine learning model 132 has a configuration 300. With the configuration 300, the machine learning model 132 comprises a representation subsystem 302, a source attention subsystem 304, a composite subsystem 306, a composite attention subsystem 308, and a result subsystem 310. Each "subsystem" in the machine learning model 132 may consist of one or more blocks, one or more sub-blocks, one or more layers, or a combination thereof. Each "block" in the machine learning model 132 may consist of one or more sub-blocks, one or more layers, or a combination thereof. Each "sub-block" of the machine learning model 132 may consist of one or more layers (or elements).

[0096] The representation subsystem 302 receives the sequence data of the software as input and generates representations for various sequences in the sequence data 110. A "representation" may include, for example, a set of elements (e.g., each element containing one or more values), where each element represents or identifies one or more amino acids or one or more nucleic acids in the original sequence of the representation. For example, each amino acid in the original sequence may be represented by a unique binary string and / or vector of values ​​that differs from the binary string and / or vector representing another amino acid.

[0097] The initial attention subsystem 304 receives these representations as input, processes these representations, and generates transformed representations, which are sent to the composite subsystem 306. The initial attention subsystem 304 consists of various attention units, each of which includes at least one self-attention layer.

[0098] In one or more embodiments, the presentation subsystem 302 can process the peptide sequence 126 to generate a representation 312 of the peptide, which is then processed by the attention unit 314 in the source attention subsystem 304 to generate a transformed representation 316 of the peptide. This processing can form at least a part of the peptide processing path 136 in Fig. 1. In addition, the presentation subsystem 302 can process the IPC sequence 124 to generate an IPC representation 318, which is then processed by the attention unit 320 in the source attention subsystem 304 to generate a transformed IPC representation 322. This processing can form at least a part of the IPC processing path 134 in Fig. 1. When the IPC sequence 124 is an MHC sequence, the IPC representation 318 is called an MHC representation, and the transformed IPC representation 322 is called a transformed MHC representation.When the IPC sequence 124 is a TCR sequence, the IPC representation 318 is called the TCR representation and the converted IPC representation 322 is called the converted TCR representation.

[0099] In some embodiments, the presentation subsystem 302 may process the N-flank sequence 128 to generate the N-flank representation 324, which is then processed by the attention unit 326 in the original attention subsystem 304 to generate the inverted N-flank representation 328. This processing may form at least a portion of the N-flank processing path 138 in Fig. 1. In some embodiments, the presentation subsystem 302 may process the C-flank sequence 130 to generate the C-flank representation 330, which is then processed by the attention unit 332 in the original attention subsystem 304 to generate the inverted C-flank representation 334. This processing may form at least a portion of the C-flank processing path 140 in Fig. 1.

[0100] When the machine learning model 132 also comprises a TCR processing path 142, the representation subsystem 302 can process the TCR sequence 131 to generate a TCR representation 336, which is then processed by the attention unit 338 in the original attention subsystem 304 to generate a transformed TCR representation 340. This processing can form at least a part of the TCR processing path 142 in Fig. 1.

[0101] The composite subsystem 306 receives the transformed representations (for example, the transformed peptide representation 316, the transformed IPC representation 322, the transformed N-flank representation 328, the transformed C-flank representation 334, the transformed TCR representation 340, or a combination thereof), which are the result from the original attention subsystem 304 and performs one or more operations to generate the composite representation 342. The composite representation 342 can be, for example, a collection of transformed representations that are the result of the original attention subsystem 304. In one or more embodiments, the composite representation can include a linkage layer that links the transformed representations that are the result of the original attention subsystem 304.In some embodiments, the composite representation 342 comprises one or more additional feature vectors (e.g., which may be added at the beginning or end of the transformed representation). The additional feature vector may have, for example, a length equal to the number of features corresponding to each individual amino acid present in the corresponding original sequence. An additional feature may include, for example, a sequence start element (SSE).

[0102] The composite representation 342 is sent as input to the composite attention subsystem 308. The composite attention subsystem 308 comprises one or more attention units for processing the composite representation 342. For example, the composite attention subsystem 308 may comprise an attention unit 344 (which may be referred to as a composite attention unit), which receives and processes the composite representation 342. The result of the composite attention subsystem 308 is sent to the result subsystem 310 for processing, which generates a report 144, as described above in Fig. 1.

[0103] Fig. 4A-4B are schematic diagrams of various configurations for a machine learning model 400 according to one or more embodiments.

[0104] Fig. 4A is a schematic diagram of a machine learning model 400 according to one or more embodiments. The machine learning model 400 represents one example of an implementation for the machine learning model 132 in Figs. 1 and 3. The machine learning model 400 is an attention-based machine learning model. The machine learning model 400 comprises a representation subsystem 401, a source attention subsystem 403, a composite subsystem 405, a composite attention subsystem 407, and a result subsystem 409, which are example implementations for the representation subsystem 302, the source attention subsystem 304, the composite subsystem 306, the composite attention subsystem 308, and the result subsystem 310, respectively, in Fig. 3.

[0105] The presentation subsystem 401 comprises a peptide presentation unit 402 and an IPC presentation unit 404. In some embodiments, the presentation subsystem 401 further comprises an N-flank presentation unit 406, a C-flank presentation unit 408, or both. In some embodiments, when the IPC presentation unit 404 corresponds to an MHC and is used as an MHC presentation unit, the presentation subsystem 401 may also include a TCR presentation unit 410. Each of these different presentation units comprises at least one insertion layer and may include, for example, a position encoder.

[0106] The embedding layer can provide sequence embedding by, for example, converting an original non-numerical representation (e.g., a string of amino acid identifiers) into a numerical representation to generate an embedded representation. The embedding can be performed using, for example, one-hot encoding, evolutionary-motivational encodings such as BLOSUM, randomly or pseudo-randomly initialized trained embeddings, or a combination thereof. The embedded representation can be positionally encoded to generate an encoded representation. The sequence representation obtained by the representation unit can be the encoded representation or an aggregation (e.g., concatenation or sum) of the encoded representation and the embedded representation.

[0107] In some cases, various attention mechanisms may be unable to detect potential information conveyed by a series of values ​​in the input data set. Position encoders can be used and added to the embedded representation, where the position encoding uses an encoding algorithm that is trained or fixed. For example, a fixed position encoding can be defined using a sine and / or cosine function (e.g., having position within the sequence and / or size as independent variables). The position encoding can have the same size as the encoded representation. The position encodings can be summed with the embedded representation to obtain a position-indicative embedded representation of the sequence that was fed to the original attention subsystem 403.

[0108] For example, the peptide representation unit 402 may comprise an insertion layer 412 that provides insertion of a peptide sequence (for example, peptide sequence 126 in Fig. 1) to generate an embedded peptide representation, and a position encoder 414 that encodes the positionally embedded peptide representation to generate a peptide representation (for example, peptide representation 312 in Fig. 3) that represents the peptide sequence. The IPC representation unit 404 may comprise an insertion layer 416 that provides insertion of an IPC sequence (for example, IPC sequence 124 in Fig. 1) to generate an inserted IPC representation, and a position encoder 418 that encodes the positionally inserted IPC representation to generate an IPC representation (for example, IPC representation 318 in Fig. 3) that represents the IPC sequence.

[0109] Furthermore, the N-flank representation unit 406 may comprise an insertion layer 420 that inserts an N-flank sequence (for example, an N-flank sequence 128 in Fig. 1) to generate an inserted N-flank representation, and a position encoder 422 that encodes the positionally inserted N-flank representation to generate an N-flank representation (for example, an N-flank representation 324 in Fig. 3) that represents the N-flank sequence. The C-flank representation 408 may comprise an insertion layer 424 that inserts a C-flank sequence (e.g., C-flank sequence 130 in Fig. 1) to generate an inserted C-flank representation, and a position encoder 426 that encodes the positionally inserted C-flank representation to generate a C-flank representation (e.g., C-flank representation 330 in Fig. 3) that represents the C-flank sequence.

[0110] Furthermore, the TCR representation unit 410 may comprise an insertion layer 428 that inserts a TCR sequence (for example, TCR sequence 131 in Fig. 1) to generate an inserted TCR representation, and a position encoder 430 that encodes the positionally inserted TCR representation to generate a TCR representation (for example, TCR representation 336 in Fig. 3) that represents the TCR sequence.

[0111] Sequence insertion may involve, for example, converting the original non-numerical representation (e.g., one containing a series of amino acid identifiers) into a numerical representation. The insertion may include one-hot encoding, evolutionary-motivational encodings such as BLOSUM, or randomly or pseudo-randomly initialized trained insertions. The representation may contain a sum and / or aggregation (e.g., linkage) of the encoding of the sequence positions and the inserted sequence.

[0112] The representations generated by the representation subsystem 401 are sent as input to the source attention subsystem 403 for processing. The source attention subsystem 403 may comprise a different internal attention mechanism that determines, for each one, several, or all positions in the representation, an attention weight (e.g., indicating how much attention to pay) to the value of each one or more other positions. The attention weights can then be used to generate a transformed value for the position.

[0113] The original attention subsystem 401 comprises an attention unit 432 and an attention unit 434. The original attention subsystem 401 may also comprise, in some embodiments, an attention unit 436, an attention unit 438, an attention unit 440, or a combination thereof. The attention unit 432 receives the peptide representation from the peptide representation unit 402 and processes the peptide representation using a set of attention subunits 442 to generate a transformed peptide representation (e.g., transformed peptide representation 316 in Fig. 3). One implementation example for the attention subunit is described in more detail in Fig. 6 below. The attention unit 434 receives the IPC representation from the IPC representation unit 404 and processes the IPC representation using a set of attention subunits 444 to generate a transformed IPC representation (e.g., transformed IPC representation 322 in Fig. 3).

[0114] Furthermore, when turned on, the attention unit 436 receives the N-flank representation from the N-flank representation unit 406 and processes the N-flank representation using the set of attention subunits 446 to generate a transformed N-flank representation (for example, the transformed N-flank representation 328 in Fig. 3). The attention unit 438 receives the C-flank representation from the C-flank representation unit 408 and processes the C-flank representation using the set of attention subunits 448 to generate a transformed C-flank representation (for example, the transformed C-flank representation 334 in Fig. 3). The attention unit 440 receives the TCR representation from the TCR representation unit 410 and processes the TCR representation using the set of attention subunits 450 to generate a transformed TCR representation (for example, the transformed TCR representation 340 in Fig. 3).

[0115] The result of the transformed representations from the original attention subsystem 403 is sent to the composite subsystem 405 for processing. The composite subsystem 405 includes a composite unit 452. The composite unit 452 can form a composite representation (e.g., composite representation 342 in Fig. 3) using the result of the transformed representations from the original attention subsystem 403. For example, the composite unit 452 can aggregate, concatenate, or otherwise combine the transformed representations to form an original composite representation. In some cases, the composite unit 452 also adds one or more additional feature vectors (e.g., a NP vector) to the original composite representation.

[0116] In some embodiments, the composite subsystem 405 may also include a position encoder 454. The position encoder 454 encodes the initial composite representation positionally to generate a composite representation that is the result of the composite attention subsystem 407. When the position encoder 454 is not present in the composite subsystem 405, the initial composite representation generated by the composite unit 452 may be the result of the composite representation for the composite attention subsystem 407.

[0117] The composite attention subsystem 407 may include an attention unit 456 (which may also be referred to as a composite attention unit). The attention unit 456 comprises a set of attention subunits 458. The attention unit 456 receives the composite representation generated by the composite subsystem 405 and processes the composite representation using the set of attention subunits 458 to generate a transformed composite representation. This transformed composite representation is then output to the results subsystem 409 for processing.

[0118] The size of the result generated by the composite attention subsystem 407 or the attention subunit in the composite attention subsystem 407 may be equal to the size of the input data supplied to the composite attention subsystem 407 or the attention subunit in the composite attention subsystem 407. The size may be, for example, m × n, where m is equal to the total number of amino acids, considered as 1 (for example, for the beginning of the sequence representation), and n is equal to the number of features (a predetermined value). One column (having n values) may be selected for additional processing. One column may be the first column and / or a column associated with the representation of the beginning of the sequence.In cases where only a portion of the result for the composite attention subsystem 407 or an attention subunit in the composite attention subsystem 407 is fed to the result subsystem 409, training the machine learning model 400 can yield learned parameter values ​​that convey relevant information about both the IPC sequence and the peptide-associated sequence(s) and peptide-IPC interactions, to be represented in the sequence start representation. In other cases, the aggregated representation can be combined after the result from the composite attention subsystem 407 to produce a single vector output, which can then be fed to the result subsystem 409.

[0119] The result subsystem 409 may comprise various blocks, sub-blocks, layers, or a combination thereof for generating the final result. In one or more embodiments, the result subsystem 409 comprises a dropout block 460, a fully connected block 462, and a result block 464. The dropout block 460 may comprise, for example, one or more dropout layers. The fully connected block 462 may comprise, for example, one or more fully connected layers. The result block 464 may comprise, for example, one or more layers for filtering, selecting, transforming, or otherwise generating the result. For example, the result block 464 may comprise at least one maximum layer 465, which is configured to select a subset of the input data received in the result block 464 based on, for example, selected threshold values ​​or ranges.

[0120] In some cases, the transformed representation of the composite is received and processed by the dropout unit 460 to generate a first result, which is received by the fully connected unit 462. The fully connected unit 462 can receive and process this first result to generate a second result, at least a portion of which is received by the results unit 464. The results unit 464 receives and processes its input data to generate the interaction result 466, the immunogenicity result 468, or both of them.

[0121] In some embodiments, the fully connected unit 462 may be configured to generate one or more results having a dimension that is less than the dimension supplied to the fully connected unit 462 (e.g., less than a predetermined number of features). For example, the result of the fully connected unit 462 may contain one value, two values, or three values ​​each corresponding to a prediction belonging to a target interaction or immune response. The fully connected unit 462 may contain, for example, one hidden layer, two hidden layers, or three or more hidden layers. The number of nodes in the initial hidden layer may be greater than the number of nodes in the subsequent hidden layer. For example, the first hidden layer may contain 256 nodes, while the second hidden layer may contain 126 nodes.In various embodiments, each output from fully connected block 462 may contain a real number indicator, which may, for example, be converted into a binary and / or categorical output (e.g., using a trained activation function) and / or converted into a scaled number. For example, the scaled number may include a probability on a scale from 0 to 1.

[0122] The interaction result 466 may comprise, for example, a set of interaction predictions 470, a set of interaction affinity predictions 472, or both with respect to one or more target interactions. An interaction prediction may comprise, for example, a prediction for a corresponding peptide-IPC (e.g., peptide-MHC, peptide-TCR) combination, whether the IPC (e.g., MHC, TCR) will bind to the peptide. The interaction prediction may include, for example, a prediction for a corresponding peptide-IPC (e.g., peptide-MHC) combination, whether the IPC (e.g., MHC) will present the peptide on the cell surface. In addition, the interaction affinity prediction may include, for example, a prediction of the affinity for the target interaction for the corresponding peptide-IPC (e.g., peptide-MHC, peptide-TCR) combination. The target interaction may be, for example, the binding of a peptide and an IPC.The affinity for the target interaction, which may be, for example, the binding affinity, indicates the strength, tendency, and / or stability of binding between the peptide and the IPC.

[0123] Immunogenicity result 466 contains a set of immunogenicity predictions. An immunogenicity prediction may include, for example, a prediction of the immunogenicity relative to the corresponding peptide-IPC combination. For example, an immunogenicity prediction may indicate the ability of a peptide to provoke an immune response relative to a specific IPC of interest (e.g., a TCR or MHC-TCR complex).

[0124] In some cases, the first part of the result from the fully connected block 462 is sent to the result block 464, while the second part of the result from the fully connected block 462 is in its final form and is used as the set of interaction affinity predictions 472.

[0125] In other embodiments, the transformed composite representation received by the result subsystem 409 is received and processed by a fully connected unit 462, which processes the transformed composite representation to generate a first result, which is sent to the dropout unit 460. The result of the dropout unit 460 or a portion thereof may then be sent to the result unit 464 for processing.

[0126] In some embodiments, the result from the result subsystem 409 may include a plurality of results that include, for each IPC allele (e.g., MHC), a prediction regarding whether and / or the probability of binding of a peptide to the IPC allele. Allele-specific predictions may be the result, or in some case, the maximum layer 465 may be used to determine the maximum of allele-specific predictions, and the maximum may be the result.

[0127] In this method, the result subsystem 409 can be implemented in any of a number of different ways, with any number of different blocks, sub-blocks, and / or layers that provide for the generation of the interaction result 466, the immunogenicity result 468, or both. Processing of peptide sequences separately from the processing of IPC sequences (e.g., MHC sequences, TCR sequences, combined MHC-TCR sequences, etc.) before enhancing the predicted performance of the machine learning model 400 by the composite subsystem 405.For example, generating a transformed peptide representation using peptide representation unit 402 and attention unit 432 in a way that is separate from generating a transformed IPC representation using IPC representation unit 404 and attention unit 434 (and, if necessary, separate from generating a transformed TCR representation using TCR representation unit 410 and attention unit 440) before generating a composite representation improves the accuracy of the result generated by result subsystem 409. Moreover, such processing can provide efficient processing (e.g., using reduced computing resources, faster processing, etc.), since multiple peptide-IPC (and peptide-TCR) combinations can be considered in a modular way.

[0128] In various embodiments, the machine learning model 400 can facilitate the automatic determination of which specific IPC allele is predicted to bind and present a peptide. For example, if an MHC molecule contains 6 MHC alleles (as is the case for humans), 6 repetitions of at least a portion of the neural network processing can be performed (e.g., in parallel) - one for each allele. Each processing can use as input an MHC representation of the MHC sequence for the MHC allele and a peptide representation of at least a portion of the peptide sequence. Each processing can generate a result corresponding to the prediction of whether the peptide will bind to and / or be presented by the MHC allele.It may be concluded that the peptide associated with the highest prediction value (e.g., indicating the most likely binding and / or presentation prediction) across alleles is the one to which the peptide will bind and the one that will present the peptide.

[0129] In some cases, for 6 MHC alleles, six composite representations can be created by running 6 different MHC allele sequences through the same IPC representation block 404 and generating a composite representation for each allele-peptide combination. In some embodiments, each of the six composite representations can be aggregated (e.g., linked) together, along with a sequence start marker (vector) inserted with the insertion layer. Each of the six composite representations can then be fed through the composite subsystem 407, as described above.

[0130] In some embodiments, the processed sequence start marker can be extracted and fed to a fully connected result block 462 directly into the terminal node of the machine learning model 400. This NP marker can represent the node representation probability. In some cases, each fully connected subblock in the fully connected block 462 can have dropout applied and is followed by a batch normalization layer. In some embodiments, the result block 464 is used for inverse convolution, such that ~6 pairwise peptide-MHC interactions will be matched for one selected MHC allele by applying an activation function (e.g., via a max layer 465, which can contain a softmax function) on ~6 representation predictions.During training, the selected peptide-MHC interaction result can be normalized as a value between 0 and 1 and can be compared with the true representation value using a loss function (e.g., binary loss function) to generate an error to tune the model parameters.

[0131] In yet other embodiments, one or more attention units or attention subunits included in the machine learning model 400 may be replaced by another type of network and / or processing unit for transforming the representation of one or more sequences. The transformation may represent the degree to which different amino acids (at specific positions) are predicted to influence the binding affinity and / or the likelihood of the representation and / or the degree to which different specific combinations of amino acids (at specific positions) occurring in one sequence or in sequences are predicted to influence the binding affinity and / or the representation. For example, one or more attention subunits may be replaced by one or more adjustable repeating elements.

[0132] Fig. 4B is a schematic diagram of another configuration for a machine learning model 400 according to one or more embodiments. With the configuration shown in Fig. 4B, the representation subsystem 401 comprises an aggregate representation unit 480. The aggregate representation unit 480 receives an aggregate sequence, such as, for example, an aggregate of a peptide sequence (e.g., peptide sequence 126 in Fig. 1) and an N-flanking sequence (e.g., N-flanking sequence 128 in Fig. 1) and / or a C-flanking sequence (e.g., C-flanking sequence 130 in Fig. 1).

[0133] The aggregate representation unit 480 may include, for example, an insertion layer 482 that processes the aggregate sequence to obtain an inserted aggregate representation that can be received by a position encoder 483 that positionally encodes the inserted aggregate representation 484. Thus, the aggregate representation 484 may include a peptide representation 485 of the original peptide sequence and an N-flank representation 486 of the original N-flank sequence and / or a C-flank representation 487 of the original C-flank sequence.

[0134] Aggregate representation 484 represents the result of aggregate representation block 480 and is sent to attention block 488 in the original attention subsystem 403 for processing. Attention block 488 contains a set of attention subblocks 489 that process aggregate representation 484 to generate a transformed aggregate representation, which is sent to composite block 452 for processing.

[0135] In some embodiments, if the aggregate sequence sent to the aggregate representation unit 480 contains either an N-flank sequence or a C-flank sequence, but not the other, then the machine learning model 400 may also contain a corresponding representation unit (e.g., an N-flank representation unit 406 or a C-flank representation unit 408) and a corresponding attention unit (e.g., an attention unit 436 or an attention unit 438, respectively) for a sequence not included in the aggregate sequence.

[0136] Fig. 4B is a schematic diagram of another configuration for a machine learning model 400 according to one or more embodiments. With the configuration shown in Fig. 4B, the peptide representation and the N-flank representation and, optionally, the C-flank representation generated by the representation subsystem 401 are sent to the aggregate unit 490. The aggregate unit 490 can aggregate (e.g., concatenate) these representations to form an aggregate representation, which is sent to the attention unit 492. The attention unit 492 comprises a set of attention subunits 494 that process the aggregate representation to generate a transformed aggregate representation, which is sent to the composite unit 452 for processing.

[0137] As shown in Figs. 4A-4B, the machine learning model 400 can be implemented in a number of different ways, using any number of combinations of blocks, sub-blocks, and / or layers in different subsystems. Thus, the machine learning model 400 is modular and can be adapted to a given task.

[0138] Fig. 5 is a schematic diagram of an attention unit 500 according to one or more embodiments. The attention unit 500 may be one example of an implementation for the attention unit in the original attention subsystem 304 in Fig. 3, the attention subsystem 308 of the composite in Fig. 3, or the original attention subsystem 403 in Figs. 4A-B. In addition, the attention unit 500 may be an example of an implementation for the attention unit 456 in Figs. 4A-4B.

[0139] The attention unit 500 comprises one or more attention sub-units. For example, the attention unit 500 may comprise an attention sub-unit 1 501 and, optionally, one or more other attention sub-units up to an attention sub-unit n 504. When multiple attention sub-units are present in the attention unit 500, these attention sub-units may be connected in series (e.g., connected in a daisy chain to obtain the final result).

[0140] The attention sub-unit 1 501 can be implemented in various ways. In one or more embodiments, the attention sub-unit 1 501 comprises, for example, an intra-attention layer 506, an addition and normalization layer 508, a feedforward layer 510, and an addition and normalization layer 512. With this configuration for the attention sub-unit 501, the attention sub-unit 1 501 can also be called a transformer encoder. The intra-attention layer 506 can be implemented using, for example, a single-head attention unit or a multi-head attention unit. If present, one or more other attention sub-units in the attention unit 500 up to the attention sub-unit n can be implemented in a manner similar to the attention sub-unit 1 501.

[0141] In the addition and normalization layer, the transformed representation can be added to the position-indicative embedded sequence representation (via residual connection), and the resulting representation can be normalized. The normalized data can be fed to a corresponding feedforward layer 510 (e.g., a fully connected feedforward network). The feedforward network can influence (for example) for each position one, two, three or more linear transformations and / or can include an activation (e.g., ReLU activation) between each linear transformation. For example, a feedforward layer can be represented: FF(x)=max(0,xW1+b1)W2+b2, where x is the input data to the layer, W1 and W2 are the slopes of the linear transformations, and b1 and b2 are the interrupts of the linear transformation.The output dimension of a particular feedforward layer of an attention subunit may be the same as the input dimension of the feedforward layer of the attention subunit. Thus, in some cases, to preserve representations of different types of information, the input and output may be summed and normalized (e.g., via another residual connection through another summation and normalization layer). II.B.3 A Typical Self-Attention Mechanism.

[0142] Fig. 6 is a flow chart of a method for processing a sequence representation using a typical self-attention layer according to one or more embodiments. The method 600 can be used, for example, by one or more attention units present in the machine learning model 132 in Figs. 1 and 3, one or more attention units present in the machine learning model 400 in Figs. 4A-4B, and / or the attention unit 500 in Fig. 5.

[0143] Step 602 includes obtaining a sequence representation that contains a plurality of elements. The sequence representation represents an amino acid sequence, or a genetic sequence of a nucleic acid, or a codon sequence in a genetic sequence. In one or more embodiments, each element of the plurality of elements in the sequence representation represents an amino acid (or amino acid residue), a nucleic acid, a codon, etc. Furthermore, each element is associated with a unique position in the sequence.

[0144] The sequence representation may be, for example, a peptide representation, an IPC representation, an N-flank representation, a C-flank representation, an MHC representation, a TCR representation, an aggregate representation, or another type of representation. For example, the sequence representation may represent part or all of: a variant-encoding sequence, part or all of a sequence that encodes a wild-type or mutant peptide, an epitope sequence (e.g., which contains a variant), a candidate neoepitope sequence, part or all of a neoantigen sequence, a sequence that begins or ends at the end of a peptide (e.g., the N-flank or C-flank), an MHC sequence (e.g., a pseudo-MHC sequence). The sequence representation may, for example, be generated using the representation subsystem 302 in Fig. 3 or the representation subsystem 401 in Figs. 4A-4B. Step 304 comprises determining a key vector, a value vector, and a query vector for each element in the sequence representation using a set of key weights, a set of value weights, and a set of query weights, respectively. For example, if the sequence represented in the sequence representation contains, for example, 20 amino acids, 20 key vectors, 20 value vectors, and 20 query vectors can be generated. An element in the sequence representation can correspond, for example, to a row or column in a 2-dimensional sequence representation (e.g., where the first dimension represents different amino acids in the sequence, and the second dimension represents, for example, different components characterizing individual amino acids).

[0146] In some embodiments, the set of key weights is in the form of a key weight matrix. The key weight matrix for a particular element may have a size equal to the length of the element by the length that the key vector must have. For example, an element may have a length of 20 (e.g., each value corresponding to a binary indication regarding whether an amino acid in the sequence is the same as a specific 1 of 21 amino acids), and if the length of the key vector is 5 (e.g., representing 5 components or features), the key weight matrix may have a size of [5, 21]. The key weight matrix may be learned during training (e.g., and randomly initialized at the beginning of training).

[0147] The value vector for an element may have the same size as the key vector for the element. The value vector may be defined using a set of value weights, which may be learned during training and included in the value weight matrix. The value weight matrix for a given element may have the same size as the key weight matrix and / or may have a size determined based on the length of the element and the length of the value vector.

[0148] The query vector for an element may have the same size as the key vector and / or value vector for the element. The query vector may be defined using a set of query weights that can be learned during training and which can be included in the query weight matrix. The query weight matrix for an element may have the size of the key weight matrix and / or the value weight matrix and / or may have a size determined based on the element's length and the length that the query vector can have.

[0149] Step 606 includes generating a set of element-oriented attention scores for each element in the sequence representation using a query vector of elements (generated using query weights and a sequence representation) and key vectors of a plurality of elements (generated using key weights and a sequence representation). For a given element, the set of element-oriented attention scores may indicate what weight to give to the value vector of the given element. The plurality of elements for which the key vectors are used in generating the set of element-oriented attention scores for a selected element in the sequence representation may include some or all elements in the sequence representation (e.g., representations of some or all represented amino acids).A set of elements may include an attentional element (e.g., a particular amino acid for which a set of element-oriented attentional metrics is defined).

[0150] A set of item-oriented attention scores is generated by generating, for each item, a sequence representation of the attention score for each pair of the attention item (the first item) with the same or different item (the second item). The attention score for this pair can be defined as the product of the query vector of the first item and the key vector of the second item.

[0151] In some cases, step 606 may include the implementation of an activation function and / or normalization. Normalization may be based on the dimensionality of the underlying vector (or query vector). For example, normalization may be defined as the square root of the length of the key vector. The activation function may contain a softmax function. In some cases, normalization is applied before the activation function.

[0152] Step 608 includes performing a transformation of a plurality of elements to form a plurality of modified elements, wherein the transformation is performed using a set of element-oriented attention scores generated for each of the plurality of elements and a value vector defined for each of the plurality of elements. For example, if a sequence representation includes 11 elements (for example, a representation of 11 amino acids), and if attention scores are determined for all pairwise combinations of elements, a modified sequence representation containing a plurality of modified elements is generated, in which a modified element is defined as a weighted average of the value vectors of all elements (using attention scores for weighting).

[0153] Step 610 includes generating a sequence encoding using a transformed sequence representation, an original sequence representation, and a feedforward neural network. For example, the transformed sequence representation and the original sequence representation can be summed. This result can still include multiple elements (e.g., each updated through transformation, summation, and normalization). The feedforward neural network can then process the summed representations (e.g., by performing one, two, or more linear transformations and / or implementing one or more activation functions). Summing the representations can reintroduce position information that may be hidden in the transformed sequence representation (due to attention to the values ​​of other elements when generating the transformed value vector for a given element).

[0154] A feedforward neural network can be configured to process each of the updated set of elements separately (e.g., using the same technique and / or the same set of parameters). Thus, the input to the network with an event prediction mechanism may include a vector that corresponds to a single element, a single amino acid, and / or a single sequence position. The feedforward network can be configured such that the output of the feedforward network has the same size as the input to the feedforward network. In some cases, instead of processing the transformed sequence representation and the original sequence representation using the feedforward network, a convolution (e.g., a 1-dimensional convolution) is used to perform a localized transformation that operates identically across positions / elements.1-D convolution can be used as another way to interpret the functioning of a neural network with an event prediction mechanism.

[0155] The technique shown in Fig. 6 applies to unidirectional attention (where key vectors, value vectors, and query vectors are used to calculate attention scores). Multidirectional attention can be used alternatively. Each direction of attention in multidirectional attention can be associated with its own set of key weights, its own set of value weights, and its own set of query weights. Each direction of attention in multidirectional attention can then produce a separate key vector, a separate value vector, and a separate query vector. Each direction of attention in multidirectional attention can use these separate vectors to obtain attention scores and transformed values ​​for each element. The transformed values ​​can be concatenated and reflected.

[0156] It should also be understood that, although Fig. 6 refers to the calculation and application of various vectors, matrix representations can be used instead. Matrix representations can facilitate efficient element-by-element calculations, as opposed to iteratively calculating various vectors individually.

[0157] Fig. 7 is a schematic diagram showing the method 600 described in Fig. 6 above, according to one or more embodiments. In Fig. 7, the method 700 for representation and attention receives a sequence 702 as input. The sequence 702 can be, for example, an amino acid sequence.

[0158] In the illustrative example in Fig. 7, sequence 702 contains a plurality of amino acids 704 (4 amino acids: x 1 -X 4 ). Representation 706 of a sequence containing a set of elements a 1 -A 4, is generated by insertion and, in some embodiments, position encoding. Each element a i may be, for example, a numerical vector. Sequence representation 706 may be one example of a sequence representation obtained in step 602 in Fig. 6.

[0159] Vectors 708 (for example, vector q i queries, vector k i keys and vector v i values) can be generated for each element a i . Vectors 708 may be example implementations for the vectors generated in step 604 of Fig. 6. An illustrative example corresponds to the generation of selected element-oriented attention indicators 710, with orientation to the first element, al. The element-oriented attention metrics 710 are an example of one set of element-oriented attention metrics generated for a particular element in step 606 of Fig. 6. Each of the element-oriented attention metrics is defined as the scalar product q 1 with k i . Weighted sum of vectors of values ​​v i and the weights were set on were calculated to perform the transformation that generated the modified element 712, b 1 . The modified element 712 is one example of a modified element generated in step 608 of Fig. 6. Similar transformations can be performed for other elements of the sequence representation 706. II.B Typical techniques using a machine learning model

[0160] The machine learning model 132 in Figs. 1 and 3 and the machine learning model 400 in Figs. 4A-4B can be used in various ways to generate predictions about the immunological activity (e.g., predicted binding, binding affinity, predicted presentation phenomenon, immunogenicity, etc.) associated with various peptides, including mutant peptides (e.g., neoantigens).

[0161] Fig. 8 is a flow chart of a method for generating information on the immunological activity of various peptides. At least part of the method 800 can be implemented using, for example, but not limited to, the prediction system 100 described in Fig. 1. For example, at least part of the method 800 can be implemented using, for example, but not limited to, the machine learning model 132 of Figs. 1 and 3 or the machine learning model 400 of Figs. 4A-4B.

[0162] Step 802 includes obtaining a peptide sequence that characterizes a mutant peptide, wherein the peptide sequence comprises a variant relative to a corresponding reference sequence. The peptide sequence characterizes the mutant peptide by determining the characteristics of at least a portion of the mutant peptide. The mutant peptide can be, for example, a neoantigen. Step 802 can be performed, for example, by extracting the peptide sequence from a data array (e.g., data array 104 in Fig. 1, cloud storage, a server or server system, etc.). In some embodiments, the peptide sequence can be one of a plurality of peptide sequences that are processed by a machine learning model.

[0163] Step 804 includes obtaining an immunoglobulin complex (IPC) sequence identified for an immunoglobulin complex (IPC). An IPC can be, for example, an MHC, a TCR, or an MHC-TCR complex. Thus, the IPC sequence can be an MHC sequence, a TCR sequence, or an MHC-TCR sequence. The IPC sequence characterizes the IPC by determining the characteristics of at least a portion of the IPC. Step 802 can be performed, for example, by retrieving the IPC sequence from a data array (e.g., data array 104 in Fig. 1, cloud storage, a server or server system, etc.). In some embodiments, the IPC sequence can be one of a plurality of IPC sequences that are processed by a machine learning model.

[0164] Step 806 includes processing the peptide sequence and the IPC sequence using different processing paths in the attention-based machine learning model to generate a result, wherein the result provides information about the immunological activity regarding both the mutant peptide and the IPC. Step 806 includes, for example, processing the peptide sequence through a corresponding representation unit to generate a peptide representation, which is processed through a corresponding attention unit to generate a transformed peptide representation that represents the peptide sequence.This peptide processing pathway is separate from the IPC processing pathway, in which an IPC sequence is processed by an appropriate representation unit to generate an IPC representation (e.g., an MHC representation, a TCR representation, an MHC-TCR representation), which is processed by an appropriate attention unit to generate a converted IPC representation (e.g., a converted MHC representation, a converted TCR representation, a converted MHC-TCR representation) that represents the IPC sequence.

[0165] In some embodiments, the peptide representation is part of an aggregate representation that also comprises an N-flank representation for the N-flank sequence and / or a C-flank representation for the C-flank sequence. In such embodiments, the aggregate processing pathway (which will by definition include the peptide processing pathway) remains separate from the IPC processing pathway.

[0166] In various embodiments, at step 806, the transformed peptide representation and the transformed IPC representation are used to form a composite representation, which is then further processed to generate a result. For example, the composite representation can be transformed using an attention block to generate a transformed composite representation, which is then processed to generate a result. The result can include, for example, but not limited to, a set of interaction predictions, a set of interaction affinity predictions, a set of immunogenicity predictions, or a combination thereof.

[0167] Step 808 includes generating a report based on the result. The report may include the result. In other embodiments, the report includes a transformed or filtered version of the result. In yet other embodiments, the report includes a summary, synopsis, or visual representation of the result.

[0168] In some embodiments, method 800 further includes step 810. Step 810 includes performing a set of actions based on the report. The set of actions may include various actions regarding the design and / or production of the treatment agent based on the report.

[0169] Fig. 9 is a flow chart of a method for generating information on the immunological activity of various peptides. At least part of the method 900 can be implemented using, for example, but not limited to, the prediction system 100 described in Fig. 1. For example, at least part of the method 900 can be implemented using, for example, but not limited to, the machine learning model 132 from Figs. 1 and 3 or the machine learning model 400 from Figs. 4A-4B.

[0170] Step 902 includes obtaining sequence data that contains a plurality of peptide sequences and a plurality of IPC sequences.

[0171] Step 904 involves generating multiple peptide-IPC combinations using peptide and IPC sequences. Each peptide-IPC combination is unique.

[0172] Step 906 includes inputting, for each peptide-IPC combination, a peptide sequence corresponding to the peptide-IPC combination into the peptide processing path of the machine learning model and an IPC sequence corresponding to the peptide-IPC combination into the IPC processing path of the machine learning model.

[0173] Step 908 includes processing, for each peptide-IPC combination, a peptide representation from the peptide sequence using a first attention block, and processing an IPC representation of the IPC sequence using a second attention block, to generate a transformed peptide representation and a transformed IPC representation, respectively.

[0174] Step 910 includes generating, for each peptide-IPC combination, a composite representation using the transformed peptide representation and the transformed IPC representation.

[0175] Step 912 comprises processing for each combination of peptide-IPC representation of the composite using the third attention block to generate a transformed representation of the composite.

[0176] Step 914 includes generating a result based on the transformed composite representations. The result may provide an indication of which of the peptide sequences can be used to generate a therapeutic agent. For example, the result may provide an indication of which of the peptide sequences (and the peptide that contains this peptide sequence) has a high probability of binding to MHC, a high probability of being presented by MHC, a high affinity interaction for peptide-MHC binding, and / or a high probability of being immunogenic, thereby activating an immune response. II.B. 1 Exemplary Methodologies: Peptides and MHC

[0177] Fig. 10 is a flow chart of a method for training a machine learning model and using the trained machine learning model to generate predictions regarding peptides and MHC, according to one or more embodiments. The method 1000 can be performed using the prediction system 100 in Fig. 1. For example, the method 1000 can be implemented using the machine learning model 132 in Figs. 1 and 3 or the machine learning model 400 in Figs. 4A-4B. In some cases, part or all of the method 1000 can be performed on a remote computing system that is remote from the user device and / or laboratory. The remote computing system can be a cloud computing system.

[0178] Step 1002 includes accessing a training data set with training elements identifying training peptide sequence data, training MHC sequence data, and training immunological activity data. The training data set may be one example of implementation for training data 133 in Fig. 1. The training data on immunological activity may include, for example, interaction readings.

[0179] The peptide sequence training data may include, for example, one or more peptide sequences (which may contain variant-encoding sequences) for training. The peptide sequence may identify an ordered set of amino acids in a peptide (e.g., a neoantigen). The peptide sequence may identify amino acids in an epitope (e.g., which contains a variant and / or which contains or which is a neoepitope) of the peptide. In some embodiments, the peptide sequence is in an aggregate sequence that also contains an N-flanking sequence (e.g., characterized by a chain of amino acids at the N-terminus of the corresponding peptide) or a C-flanking sequence (e.g., characterized by a chain of amino acids at the C-terminus of the corresponding peptide). Neither the N-flank nor the C-flank binds to an MHC molecule, although each can influence whether it is presented by an MHC molecule.

[0180] The MHC sequence training data may include one or more MHC sequences for training. An MHC sequence may, for example, identify amino acids in part or all of an MHC molecule (e.g., an MHC-I molecule or an MHC-II molecule). The MHC sequence may comprise a pseudo-MHC sequence (e.g., one that contains 34 amino acids). The MHC sequence may identify amino acids in, for example, 1, 2, 3, 4, 5, or 6 MHC alleles for MHC-I or 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, or 12 MHC allotypes for MHC-II. The MHC sequence may identify amino acids constituting part or all of an HLA molecule.

[0181] The training data for immunological activity may include, for example, one or more interaction readouts for one or more peptide-MHC combinations. For example, the training data set may include training units, in which each training unit contains a peptide sequence and an MHC sequence for training, as well as one or more interaction readouts for the corresponding peptide-MHC combination. An interaction readout may indicate whether a target interaction (e.g., binding of the peptide to the MHC, presentation of the peptide on the cell surface by the MHC) occurs between the peptide and the MHC or the affinity for the target interaction.

[0182] The interaction indicator may be, for example, a marker. A negative interaction marker may indicate that the peptide does not bind to and / or is not presented by the MHC molecule. A positive interaction marker may indicate that the peptide binds to and / or is presented by the MHC molecule. Furthermore, the interaction marker may indicate the probability that the peptide binds to the MHC molecule, the probability that the MHC molecule presents the peptide on the cell surface, the binding affinity of the peptide-MHC combination, the strength of binding between the peptide and the MHC molecule, the stability of binding between the peptide and the MHC molecule, the tendency of the peptide to bind to the MHC, or another measure or characteristic related to the interaction between the MHC and the peptide.

[0183] The training dataset could have been generated through, for example, in vitro or in vivo experiments and / or based on medical records. The training data could have been generated based on one or more of the techniques disclosed in Section II.D below.

[0184] Accessing a training data set may include, for example, retrieving the training data set from a local or remote storage, downloading the training data set, and / or requesting (and receiving) a portion or all of the training data set from one or more data sets (e.g., a cloud data store, a server system, or some other data source).

[0185] In some cases, the original training data set (e.g., containing variant-encoding sequences) may contain predominantly negative data, in which a relatively small portion of sequence combinations (e.g., peptide-MHC combinations) are found to be associated with the actual target interaction. The training data set can be designed to include negative training data elements. In some embodiments, a negative training data element can be defined to identify amino acids in a pseudo-randomly selected fragment of the origin protein in the positive set (corresponding to the observed representation). For example, a negative training data element can be simulated based on the positive set. The fragment can be selected as having a length in a predetermined range (e.g., 8-14 amino acids for MHC-I and 8-30 amino acids for MHC-II, using a standard probability).The N-terminal and C-terminal flanking sequences can be retained in the negative training data element, potentially yielding a maximum length (e.g., 10 amino acids). Any peptide fragment (e.g., at least a 9-mer) that overlaps with the positive peptide can be discarded from the negative training data.

[0186] In various embodiments, negative training data elements are simulated based on positive data elements. Furthermore, the training data is selected such that a different set of negative training data elements is used at each moment of the training period. For example, for each moment, a different "negative subset" of negative peptide sequences can be selected from the total space of available negative peptide sequences identified based on the positive set of peptide sequences. The negative subset selected for each moment can be unique in that no negative peptide sequence is repeated in any of the negative subsets for the total number of moments.Thus, the training data used for each moment of the training period includes the same set of positive peptide sequences, but a completely different set of negative peptide sequences. This technique, which may be referred to as "negative set switching," can ensure the full rigidity of training and help ensure either a reduced number of false negatives (e.g., false negative reads / predictions) by the machine learning model or that a false negative is repeated more than once. Furthermore, using this technique, the machine learning model can be trained on a total number of negative peptide sequences equal to the number of positive peptide sequences multiplied by the number of moments in the training period.

[0187] Step 1004 includes training a machine learning model using a set of training data. The machine learning model may be, for example, the machine learning model 132 in Figs. 1 and 3, or the machine learning model may be, for example, the machine learning model 400 in Figs. 4A-4B.

[0188] The machine learning model 132 can be trained using a static or dynamic learning rate. The dynamic learning rate can be obtained by, for example, equalizing the learning rate. Training can be performed using, for example, a classification loss function and / or a regression loss function. The loss function can be based on, for example, the root mean square error, the root mean square error, the mean absolute error, the median absolute error, the entropy-based error, the cross-entropy-based error, and / or the binary error based on the cross-entropy. Validation data (for example, a separate subset of the training data set used to train the machine learning model 132) can be used to evaluate the performance of the machine learning model 132 when it is trained. Training can end if and / or when the target performance characteristics are obtained, and / or the maximum number of training repetitions has been completed.

[0189] Step 1006 includes accessing a subject-specific set of variant-encoding sequences corresponding to a set of mutant peptides. As described above, the variant-encoding sequence represents one example of a peptide sequence. The subject-specific set of variant-encoding sequences may correspond to a set of mutant peptides such that each of the subject-specific sets of variant-encoding sequences identifies amino acids in a corresponding mutant peptide from the set of mutant peptides and / or such that each of the subject-specific set of variant-encoding sequences identifies one or more amino acids in a mutation. Each of the subject-specific set of variant-encoding sequences may be associated with a specific subject (e.g., a human subject).A particular subject may be diagnosed as having and / or experiencing symptoms and / or having received test results related to a particular medical condition (e.g., cancer). For example, a subject-specific set of coding variant sequences could be identified by processing a tumor sample. The sample may be or may be included in, for example, sample set 112 in Fig. 1.

[0190] A subject-specific set of variant-encoding sequences may be identified using the techniques disclosed herein (e.g., in Section II.D). For example, a subject-specific set of variant-encoding sequences could be identified by performing a sequencing technique to identify peptides in a disease sample and comparing the identified peptides to those found in a healthy sample or a reference database to identify unique sequences. In some embodiments, if the unique sequences are nucleic acid sequences, each unique nucleic acid sequence can be converted into an amino acid sequence.

[0191] Each of the subject-specific sets of variant-encoding sequences may identify amino acids in a peptide (which may be amino acids in a neoepitope of a neoantigen). In some cases, each of one, several, or all of the subject-specific sets of variant-encoding sequences may be part of a corresponding aggregate sequence that further comprises a sequence on the N-flank of the peptide and / or a sequence on the C-flank of the peptide.

[0192] Accessing a subject-specific set of variant-encoding sequences may include, for example, obtaining a subject-specific set of variant-encoding sequences from a local or remote storage and / or receiving a request for a subject-specific set of variant-encoding sequences from another device. Accessing a subject-specific set of variant-encoding sequences may include and / or may be performed in combination with determining a subject-specific set of variant-encoding sequences.

[0193] A subject-specific set of variant-encoding sequences could be obtained by identifying peptide sequences in the subject's disease sample and determining which of the peptide sequences is not present in the reference, healthy sample, and / or wild-type sequence set. In cases in which a healthy sample is used for comparison, the healthy sample could (but does not necessarily) be collected from the subject.

[0194] Step 1008 includes accessing an MHC sequence corresponding to the MHC. The MHC sequence may include, for example, a pseudo-MHC sequence (e.g., an MHC molecule) in a sample collected from a subject. In some cases, the MHC sequence and the subject-specific set of variant-coding sequences are identified from the same sample from the subject or from multiple samples from the subject (e.g., a disease sample and a healthy sample). In some cases, the MHC sequence and the subject-specific set of variant-coding sequences are identified from samples from the subject and one or more other subjects. Thus, in some cases, the MHC sequence may be specific to the subject. The MHC sequence may be or may have been determined using, for example, sequencing techniques and / or mass spectrometry.

[0195] Accessing an MHC sequence may include, for example, retrieving an MHC sequence from a local or remote storage location and / or requesting a subject-specific MHC sequence from another device. Accessing an MHC sequence may include and / or be performed in combination with determining the MHC sequence.

[0196] Step 1010 includes, for example, processing a set of subject-specific variant-coding sequences and an MHC sequence using a trained machine learning model to generate a result. Step 1010 may include processing each unique combination (e.g., a variant-coding sequence-MHC combination or a peptide-MHC combination) of a subject-specific variant-coding sequence from the set of subject-specific variant-coding sequences and an MHC sequence to generate a result.

[0197] The output generated by the machine learning model may include the same or similar data type as the immunological activity training data used to train the machine learning model. For each unique combination, the machine learning model generates an output that includes at least one of the set of interaction predictions or the set of interaction affinity predictions.

[0198] An interaction prediction in an interaction prediction set includes a prediction of whether a target interaction will occur between a mutant peptide that contains a variant coding sequence and an MHC that contains an MHC sequence. For example, an interaction prediction may include a binary or categorical prediction regarding whether a mutant peptide with an amino acid structure as indicated by a subject-specific variant coding sequence will present to and / or bind to an MHC molecule with an amino acid structure as indicated by the MHC sequence. An interaction affinity prediction in an interaction affinity prediction set includes a prediction of the affinity for the target interaction. This affinity can be determined based on, for example, the strength, tendency, and / or stability of the target interaction.For example, a prediction of interaction affinity may include a predicted binding affinity as a real number associated with a mutant peptide that includes amino acids identified in a subject-specific variant coding sequence and an MHC molecule including amino acids as defined in the MHC sequence.

[0199] Step 1012 involves generating a report based on the result of the machine learning model. The report may be implemented as, for example, report 144 in Figures 1 and 3. The report may represent or may contain the result. In some cases, the report may be a transformed or filtered version of the result.

[0200] In one or more embodiments, a subject-specific set of variant-encoding sequences is filtered, ordered, and / or otherwise processed based on the result to generate information for inclusion in a report. For example, a subject-specific set of variant-encoding sequences can be filtered to exclude sequences for which the predicted interaction affinity (e.g., binding affinity) was below a predetermined threshold affinity and / or for which the target interaction (e.g., presentation by or binding to an MHC molecule) was predicted not to or likely not to occur. In some cases, filtering is performed to identify a predetermined number and / or fraction of the subject-specific set of variant-encoding sequences.For example, filtering may be performed to identify 10, 20, 40, 60, 80, 100, 500, or 1000 variant-coding sequences associated with relatively high predicted probabilities (e.g., relative to unselected variant-coding sequences in a subject-specific set of variant-coding sequences) of whether a mutant peptide will bind to and / or be presented by an MHC molecule.

[0201] The report may identify one or more variant-coding sequences (e.g., those not filtered from the set) and / or one or more mutant peptides (e.g., those related to the selected variant-coding sequences). A mutant peptide may be identified, for example, by its name, by its sequence, and / or by identifying both the corresponding wild-type sequence and the variant present in the variant-coding sequence.

[0202] The report may, in some embodiments, identify one or more predictions associated with one or more variant-encoding sequences or one or more mutant peptides. The report may include the name of the subject. The report may, for example, be presented locally (e.g., for display on a user device's display system, sent as a notification to the user device, etc.) and / or transmitted to another device (e.g., sent to a cloud computing system, sent to cloud storage, sent to a user device associated with a healthcare professional or laboratory worker, transmitted as an email, etc.).

[0203] Fig. 11 is an illustration that contains a training data table according to one or more embodiments. Table 1100 contains training data 1102 (e.g., a training data set). Training data 1102 may be one example of a portion of training data 133 in Fig. 1. Training data 1102 may be one example of a portion of a training data set, such as the training data set described in step 1002 in Fig. 11.

[0204] The training data 1102 includes an allele identifier 1106, a training N-flanking sequence 1108, a training peptide sequence 1110, a training C-flanking sequence 1112, a training MHC sequence 1114 (e.g., a pseudo-MHC sequence), a binding affinity 1116, and a presentation indication 1118. The binding affinity 1116 indicates the detected (e.g., observed) binding affinity for binding of the peptide characterized by the training peptide sequence 1110 and the corresponding MHC characterized by the training MHC sequence 1114. The presentation indication 1118 shows whether binding or presentation of the peptide by the MHC was detected (or observed).

[0205] Fig. 12 is an illustration of a candidate neoantigen and corresponding possible candidate neoepitopes according to one or more embodiments. When a method such as method 1000 is implemented, the mutant peptide can be a neoantigen.

[0206] For a relatively long mutant peptide, which is candidate neoantigen 1200, it is possible for multiple epitopes (called neoepitopes), all containing the same mutation or variant, to be presented by an MHC molecule. Thus, the immunogenicity of the candidate neoantigen can be predicted based on the predictions generated for each of the candidate neoepitopes 1202.

[0207] Immunogenicity can be predicted, for example, by generating a list of all possible neoepitopes that can be derived from a given neoantigen and providing a representation prediction for each of some or all candidate neoepitopes (flanked by the remaining amino acids, upstream of the N-terminus and downstream of the C-terminus of the epitope, up to 10 amino acids in length) in the list. From these representation predictions, the candidate neoepitope with the highest representation probability relative to MHC 1204 candidates is selected to represent the entire neoantigen. Alternatively, the aggregate representation of multiple candidate neoepitope-MHC pairs can be used to derive an aggregate score representing the neoantigen. Such an aggregate can be performed either by considering all candidate neoepitope-MHC pairs or by considering the best neoepitope on an MHC and then aggregate across all MHC molecules.Generalization can be performed by several mathematical functions, including, for example, taking the arithmetic mean or harmonic mean of the representation or binding affinity score of each candidate neoepitope-HLA pair.

[0208] Although Fig. 12 is described with respect to neoantigens and neoepitopes, a similar technique can be used for other types of relatively long mutant peptides containing a mutation or variant and with multiple possible epitope candidates. In some embodiments, this technique can be used in conjunction with antibody drug sequences. II.B.2 Exemplary Techniques: Peptides and TCRs

[0209] Fig. 13 is a flow chart of a method for training a machine learning model and using the trained machine learning model to generate predictions regarding peptides and TCRs, according to one or more embodiments. Method 1300 can be performed using prediction system 130 in Fig. 1. For example, method 1300 can be implemented using machine learning model 132 in Figs. 1 and 3 or machine learning model 400 in Figs. 4A-4B. In some cases, part or all of method 1300 can be performed on a remote computing system that is remote from the user device and / or laboratory. The remote computing system can be a cloud computing system. Steps 1302-1312 can be implemented in a manner that is similar to steps 1002-1012 in Fig. 10, but with respect to TCRs.

[0210] Step 1302 includes accessing a training data set with training elements identifying training peptide sequence data, training TCR sequence data, and training immunological activity data. The training TCR sequence data may include one or more TCR sequences for training. A TCR sequence may, for example, identify amino acids in part or all of a TCR molecule.

[0211] The training immunological activity data may include, for example, one or more interaction readouts for one or more peptide-TCR combinations and / or one or more immunogenicity predictions. An immunogenicity prediction may predict the immunogenicity of a peptide relative to a TCR. For example, the training data set may include an interaction marker that indicates whether a mutant peptide with amino acids, as identified by the variant coding sequence, activates an immunological response (e.g., whether the mutant peptide is immunogenic). Immunogenicity may indicate that the mutant peptide has activated a T cell receptor (e.g., a CD8+ cytotoxic T lymphocyte receptor or a CD4+ helper T cell receptor) and / or activated an immunological response.

[0212] The training dataset may be generated by, for example, expressing different mutant peptides in a sample (e.g., one or more dendritic cells) and / or administering different mutant peptides (e.g., to a sample or to a subject from whom the sample is subsequently collected) via immunization and / or vaccination. The mutant peptides could be expressed or administered individually (e.g., thus focusing each experiment on a single mutant peptide) or in groups.

[0213] Immunogenicity could be tested by, for example, analyzing tumor-infiltrating cells. It could be determined that a mutant peptide activates an immunological response (and is thus immunogenic) if, for example, epitopes of the mutant peptide are detected (e.g., at levels above a threshold), the measured level of interferon-gamma (IFN-γ) or T-cell immunoglobulin and mucin domain-3 (TIM-3) exceeds the relevant threshold, the detectable number of cytotoxic T cells (e.g., in total or cytotoxic T cells displaying an epitope corresponding to the mutant peptide) exceeds the relevant threshold, and / or at least a threshold degree of apoptosis is observed. As another example, the mutant peptide could be expressed in the sample (e.g., by one or more dendritic cells).A mutant peptide can be determined to activate an immunological response (and thus be immunogenic) if, for example, it is determined that the presented peptide is subsequently recognized by a T cell. It will be appreciated that some embodiments include collecting and / or determining at least a portion of a training data set (e.g., by performing one or more experiments and / or assays disclosed herein).

[0214] Accessing a training data set may include, for example, retrieving the training data set from a local or remote storage, downloading the training data set, and / or requesting (and receiving) a portion or all of the training data set from one or more data sets (e.g., a cloud data store, a server system, or some other data source).

[0215] Step 1304 includes training a machine learning model using a set of training data. The machine learning model may be, for example, the machine learning model 132 in Figs. 1 and 3, or the machine learning model may be, for example, the machine learning model 400 in Figs. 4A-4B.

[0216] Step 1306 involves accessing a subject-specific set of variant encoding sequences corresponding to a set of mutant peptides.

[0217] Step 1308 involves accessing a TCR sequence corresponding to the TCR. In some cases, the TCR sequence and the subject-specific set of variant coding sequences are identified from the same sample from the subject or from multiple samples from the subject (e.g., a disease sample and a healthy sample). In some cases, the TCR sequence and the subject-specific set of variant coding sequences are identified from samples from the subject and one or more other subjects. Thus, in some cases, the TCR sequence may be subject-specific. The TCR sequence may be or may have been determined using, for example, sequencing techniques and / or mass spectrometry.

[0218] Accessing a TCR sequence may include, for example, retrieving a TCR sequence from a local or remote repository and / or requesting a subject-specific TCR sequence from another device. Accessing a TCR sequence may include and / or be performed in combination with determining the TCR sequence.

[0219] Step 1310 includes, for example, processing a set of subject-specific variant-coding sequences and a TCR sequence using a trained machine learning model to generate a result. Step 1310 may include processing each unique combination (e.g., a variant-coding sequence-TCR combination or a peptide-TCR combination) of a subject-specific variant-coding sequence from the set of subject-specific variant-coding sequences and a TCR sequence to generate a result.

[0220] The output generated by the machine learning model may include the same or similar data type included in the immunological activity training data used to train the machine learning model. For each unique combination, the machine learning model generates an output containing a set of immunogenicity predictions. An immunogenicity prediction in the immunogenicity prediction set may indicate whether the mutant peptide has activated an immune response (and is thus immunogenic). In some cases, the immunogenicity prediction indicates the degree of immunogenicity (e.g., low, moderate, high, very high, etc.).

[0221] Step 1312 includes generating a report based on the result of the machine learning model. The report may be implemented as, for example, report 144 in Figs. 1 and 3. Step 1312 may be implemented in a manner similar to step 1012 in Fig. 10.II.B.3 Illustrative Techniques: Additional Considerations for Training and Prediction Using a Machine Learning Model

[0222] Thus, the embodiments described herein provide a machine learning model that can be used to generate predictions for the immunological activity associated with a peptide, which may be a mutant peptide. A peptide sequence that characterizes the mutant peptide—for example, a variant encoding sequence—can be analyzed by the machine learning model with an IPC sequence characterizing the IPC to generate one or more predictions about one or more target interactions (interactions of interest) between the peptide and the IPC and / or about the ability of the peptide to provoke an immune response. The output generated by the machine learning model can thus include one or more results that provide information about one or more target interactions and / or the immunogenicity of the peptide.

[0223] In some embodiments, one or more variant-encoding sequences can be selected from a subject-specific set of variant-encoding sequences based on the results of one or more machine-learning models described herein. The input data can include representations of the MHC sequence and the variant-encoding sequence that corresponds to the mutant peptide. The machine-learning model can be trained using binding affinity data and mass spectrometry elution data that indicate which peptides are rearranged by the MHC molecules.Binding affinity data may include qualitative data (e.g., as determined by ELISA, pull-down and / or gel shift assays, fluorescence resonance energy transfer assays, and mass spectrometry analyses) or quantitative data (e.g., using a biosensor-based technique such as surface plasmon resonance, isothermal titration colorimetry, biolayer interferometry, or microscale thermophoresis). In some cases, binding affinity data may include data from a competitive binding assay, data from an immune epitope database, and / or data of a type found in an immune epitope database. Elution data can be collected by peptide-MHC immunoprecipitation, followed by elution and detection of the presented MHC ligands by mass spectrometry.The training data includes "positive" cases (for which the mass spectrometry results indicate that the peptide was represented by an MHC molecule) and "negative" cases (corresponding to, for example, simulated length-matched n-mers (nmers)) from the same proteins as the positive cases, but which were not detected in the mass spectrometry assessments).

[0224] In some cases, the number of positive cases in the training data is equal to the number of negative cases in the training data. In some cases, the number of positive cases is less or greater than the number of negative cases. Each of one, several, or all negative cases in the training data may be the same length as a positive case in the training data. In some cases, all sequences in the training data have the same length.

[0225] Part or all of a sequence may be represented, for example, by data encoding. Encoding may be performed according to a known and / or statistical rule or technique and / or by a trained network. For example, encoding may include one-hot encoding, such that each encoded sequence indicates, for each sequence position and for each set (e.g., 21) of amino acids, whether a particular amino acid is present at that position. Alternatively, evolution-motivated encodings, such as BLOSUM, or learned encodings can be used to represent amino acids in a sequence. Encoding may include position encoding (e.g., learned or fixed encoding).

[0226] In some cases, the machine learning model includes one or more neural networks used to process sequences. The neural network(s) may additionally or alternatively include, for example, an encoder neural network and / or part or all of a transformer network.

[0227] The machine learning model may include an attention-based machine learning model that comprises one or more neural networks that are attention-based, do not have any convolutional layers, and / or do not have any recurrent layers. The attention-based machine learning model may (but does not necessarily) also include one or more other neural networks that are not attention-based, contain one or more convolutional layers, and / or contain one or more recurrent layers.

[0228] An attention-based network can use a set of query weights, a set of key weights, and a set of value weights to determine, for a given amino acid representation, the degree to which each of one or more other amino acid representations should be "attentioned" when processing the given amino acid representation. A self-attention layer can use keys, values, and queries from the same layer, so that, for example, an encoder or decoder can pay attention to all positions in the previous layer of the encoder or decoder.

[0229] When predicting whether a given mutant peptide will bind and / or be presented by a particular MHC molecule, one or more transducer encoders may separately process representations of different parts or the entire variant coding sequence and / or MHC sequence. Each transducer encoder may include an intrinsic attention layer and a feedforward layer. Each attention layer may also include one or more inserted components configured to, for example, perform positional and / or non-positional insertion. In some cases, the sequences of each of the N-flanking region of the mutant peptide, the epitope region of the mutant peptide, the C-flanking region of the mutant peptide, and the MHC molecule are separately processed by different repetitions of the transducer encoder. The encoded representation of the sequence may include, for each amino acid in the sequence, a feature vector representing the amino acid.The encoded sequence representations can then be concatenated and fed into another iteration of the transducer encoder. The concatenation can thus include a feature vector for each amino acid in part or all of the variant coding sequence and for all or part of the MHC sequence.

[0230] One or more additional feature vectors may be included in the concatenation. Each of the additional features may, for example, be assigned arbitrary or pseudo-arbitrary values ​​for the feature vector. The concatenated representation (e.g., containing the additional feature vector(s)) may be processed by an additional transformer encoder to generate an encoded concatenated representation. This encoded representation of the sequence combination may be processed by a feedforward network (e.g., a fully connected neural network), where dropout and / or batch normalization may be used. In some cases, the encoded representation(s) of the additional feature vector(s) are selectively fed to the feedforward network (e.g., while feature vectors corresponding to individual amino acids of the MHC molecule and / or mutant peptide are not).For example, suppose that a subsequence of an MHC molecule contains x1 amino acids, that a subsequence of a mutant peptide (e.g., and one or more flanks) contains x2 amino acids, and that the feature transformation identifies y feature values ​​to represent each amino acid. A concatenated representation containing 1 additional feature vector might thus have size [(x1 + x2 + 1), y]. The input to the event prediction network might have size [1, y], in the case where a single feature vector is selected for processing by the event prediction network. The advantage of using the additional feature vector approach is that the model can then handle sequences of varying lengths.

[0231] The results obtained by the feedforward network may correspond to predictions regarding the binding affinities between a mutant peptide and an MHC molecule (e.g., an MHC molecule of a subject) and / or whether the mutant peptide will be presented by the MHC molecule. The binding affinity prediction may be, for example, numerical (e.g., corresponding to the predicted probability that the mutant peptide will bind to the MHC molecule, the predicted binding strength, and / or the predicted binding stability), categorical (e.g., predicting no, low, or high binding stability between the mutant peptide and the MHC molecule), or binary (e.g., predicting whether the mutant peptide binds to the MHC molecule).

[0232] The presentation prediction generated in relation to the mutant peptide may be, for example, numerical (e.g., corresponding to the predicted probability that the MHC molecule of the subject presents the mutant peptide on the cell surface or the predicted fraction of tumor cells in the subject that present the mutant peptide), categorical (e.g., a prediction of the absence, absence of frequent, or frequent presentation of the mutant peptide by the MHC molecules of the subject), or binary (e.g., a prediction of whether the mutant peptide is expressed by the MHC molecules of the subject). The presentation prediction may (but does not necessarily) be normalized and / or represent a conditioned prediction. For example, the presentation prediction may correspond to a prediction regarding whether the MHC molecule of the subject presents the mutant peptide if the mutant peptide stably binds to the MHC molecule.

[0233] In some cases, a machine learning model generates predictions corresponding to one or more potential interactions between a mutant peptide and an MHC-I molecule. For example, a machine learning model may predict the binding affinity of an MHC-I molecule and a mutant peptide and / or whether an MHC-I molecule will express a mutant peptide. The machine learning model may accept as input and may process (e.g., using one or more self-attention layers) a sequence or subsequence of an MHC-I molecule and the coding variant sequence associated with the mutant peptide.

[0234] In some cases, a machine learning model generates predictions corresponding to one or more potential interactions between a mutant peptide and an MHC-II molecule. For example, a machine learning model may predict the binding affinity of an MHC-II molecule and a mutant peptide and / or whether an MHC-II molecule will express the mutant peptide. The machine learning model may accept as input and may process (e.g., using one or more self-attention layers) a sequence or subsequence of an MHC-II molecule and the coding sequence of a variant of the mutant peptide.

[0235] In some cases, the machine learning model generates predictions corresponding to one or more potential interactions between a mutant peptide, an MHC sequence, or a subsequence, and a T cell receptor (e.g., instead of or in addition to generating predictions corresponding to one or more potential interactions between the mutant peptide and an MHC molecule). The machine learning model can then predict, for example, the binding affinity between the mutant peptide and the T cell receptor and / or whether the mutant peptide will activate and / or elicit an immune response in a T cell. The machine learning model can accept as input and can process (e.g., using one or more self-attention layers) a T cell receptor sequence or subsequence, an MHC sequence or subsequence, and the coding sequence of the variant mutant peptide.

[0236] The immunogenicity of a mutant peptide (e.g., with respect to a particular subject) may be predicted based on one or more results generated by the machine learning model disclosed herein (e.g., an attention-based machine learning model). For example, it may be predicted that a neoantigen detected in a disease sample from a subject will not cause immunogenicity or will have low immunogenicity when the result of the machine learning model predicts that the mutant peptide will have low binding affinity to an MHC molecule; that the MHC molecule will not or is not likely to present the mutant peptide; and / or that the mutant peptide will not elicit an immune response via the T cell receptor.The immunogenicity prediction generated in relation to the mutant peptide may be, for example, numerical (e.g., corresponding to the predicted probability that an immunogenicity response will be elicited in response to the mutant peptide and / or corresponding to the predicted intensity of any immunogenicity response to the mutant peptide), categorical (e.g., predicting no, low, or high immune response), or binary (e.g., predicting whether the specified mutant peptide will elicit an immune response in a subject).

[0237] Predicted immunogenicity may be further based on predictions and / or experimental indications of one or more immunogenicity factors. Factors that indicate immunogenicity may include: i) the protein level of the mutant peptide precursor; ii) the expression level of the transcript encoding the mutant peptide precursor; iii) the efficiency of processing the mutant peptide precursor by the immunoproteasome; iv) the timing of expression of the transcript encoding the mutant peptide precursor; v) the binding affinity of the mutant peptide for the T cell receptor; vi) the position of the amino acid variant in the mutant peptide; vii) the solvent exposure of the mutant peptide upon binding to the MHC molecule; vii) the solvent exposure of the amino acid variant upon binding to the MHC molecule; x) the content of aromatic residues in the peptide; xi) the properties of the amino acid variant compared to the wild-type residue; and / or xii) the nature of the mutant peptide precursor;xiii) microbial similarity of the mutant peptide to known microbial peptides; xiv) self-similarity or dissimilarity of the mutant peptide to the wild-type proteome; xv) thymus expression of the wild-type peptide. Immunogenicity factors may additionally or in addition include: the protein sequence and / or length of the mutant peptide (e.g., as indicated by the number of amino acids identified in the variant coding sequence), and / or the level of expression of the MHC allele in the subject (e.g., as measured by RNA-Seq or mass spectrometry).

[0238] A binding affinity prediction and / or a prediction regarding whether (or the likelihood that) a mutant peptide will be presented (e.g., by one or more tumor cells and / or one or more MHC molecules in a subject) can be generated according to the technique disclosed herein (e.g., using an attention-based machine learning model) for each of a set of mutant peptides (e.g., that was detected in a disease sample from a subject). These predictions can be used to select a subset of the set (e.g., less than 50% of the set, less than 25% of the set, less than 10% of the set, less than 5% of the set, and / or less than 1% of the set).A subset can be selected using one or more appropriate cutoff values ​​(e.g., to identify mutant peptides in the set that have the most stable binding to MHC molecules and / or the highest probabilities of being represented relative to others in the group) or one or more absolute cutoff values. For example, each selected mutant peptide can have an MHC binding affinity with a relatively strong affinity value (e.g., within the best 50%, best 25%, best 10%, or best 5% of the affinity values ​​in the set) and / or an absolute strong affinity value (e.g., with an affinity value better than a predetermined cutoff value / cutoff, such as 5000 nM, 1000 nM, or 500 nM, in the case of IC50 values). A subset of a set may include 1, 2, 3, 4, 5, 6, 7, 8, 9, 10 or more mutant peptides, despite a predetermined affinity threshold / cutoff.A partial subset of a set may include 20 or more neoantigens or 30 or more mutant peptides.

[0239] Each selected mutant peptide may be produced, experimentally tested (e.g., to determine binding affinity, prevalence of presentation, and / or other immune factor), incorporated into a composition (e.g., a pharmaceutical composition such as a vaccine and / or treatment), and / or administered to a subject.

[0240] Each of the set of mutant peptides for each binding affinity prediction and presentation is generated and may include a mutant peptide associated with a specific subject (e.g., a specific human subject). Each of the set of mutant peptides may be a disease-specific, immune mutant peptide determined using a disease-specific sample from an individual. A distinct variant-encoding sequence can be identified by determining the genetic and / or nucleic acid sequence (e.g., DNA, RNA, and / or mRNA sequences) in the disease sample and comparing each identified genetic and / or nucleic acid sequence to the sequence of a reference sample. Codons in the genetic and / or nucleic acid sequence indicate the existence of the corresponding amino acid in the peptide.Namely, each of the many codons can code for a given amino acid, so that while a nucleic acid sequence may specify (e.g., deterministically) an amino acid sequence, the same amino acid sequence may be coded for by other nucleic acid sequences.

[0241] Some of the sequences identified in the disease sample may be non-disease-related sequences that correspond to non-disease-related peptides. To identify disease-specific nucleic acid sequences and / or disease-specific amino acid sequences, for each sequence detected by sequencing the disease-specific sample, it can be determined whether the sequence is also identified in a reference sequence data set. The reference sequence data set may include a set of reference sequences for which the sequence is known, implied, or suspected not to indicate or characterize a disease (e.g., any disease or the specified disease).A reference sequence data set may, for example, include sequences identified by sequencing one or more reference sample sequences collected from the same subject from whom the disease-specific sample was collected, by sequencing one or more reference sample sequences collected from one or more other subjects who have not been diagnosed with any disease or a disease corresponding to the disease-specific sample, and / or by sequencing one or more cell lines unrelated to a specific disease. In some cases, a reference sequence data set may include sequences selected from one or more reference data repositories.A sequence that is found in association with a disease-specific sample but that is not found (or is found at a frequency below a predetermined threshold) in a data set of reference sequences may be classified as a variant-coding sequence (e.g., in general or for the subject from whom the disease-specific sample was collected).

[0242] In some cases, a plurality of variant encoding sequences can be identified (e.g., each was detected in a disease sample but was not represented in the reference sample sequences), and the representation of each of the plurality of variant encoding sequences can be processed (e.g., individually, sequentially, and / or in parallel) using a machine learning model disclosed herein (e.g., an attention-based machine learning model) to predict binding affinity and / or predict representation.

[0243] A disease sample may contain, for example, tissue (e.g., a solid tumor), blood, and / or a collection of cells (e.g., cancer cells that may have been collected by fine needle aspiration or laparoscopy). A disease sample may include cancer cells collected from a subject who has been diagnosed with and / or has, for example, lung cancer, melanoma, breast cancer, ovarian cancer, prostate cancer, kidney cancer, gastrointestinal cancer, colon cancer, testicular cancer, head and neck cancer, pancreatic cancer, brain cancer, B-cell lymphoma, acute myeloid leukemia, chronic myelogenous leukemia, chronic lymphocytic leukemia, and T-cell lymphocytic leukemia, non-small cell lung cancer, or small cell lung cancer.

[0244] In some cases, the original sample was split into a disease sample and another remaining sample (e.g., one that could be discarded or used as a reference sample). The reference sample may contain a matched non-disease sample. Each disease sample and reference sample may be collected from the same subject and / or may include or be of the same or similar sample type (e.g., tissue type). In some cases, the disease sample is collected from a first subject (e.g., one diagnosed with a medical condition or disease), and the reference sample is collected from a second subject (e.g., one not diagnosed with a medical condition or disease). In some cases, the sequences of the reference samples were obtained from a database of known genes associated with the organism.

[0245] The training data may also include the sequences of one or more peptides, along with indications of whether each peptide was presented bound to an MHC molecule, the MHC molecule, and / or activated an immune response. To collect training data that relates sequence data to observed presentation and / or binding data, the disease sample (and potentially a reference sample) may be (separately) processed to isolate MHC / peptide complexes (e.g., by performing immunoprecipitation with an MHC-specific antibody) and / or by eluting (and sequencing) peptides from MHC molecules (e.g., using chromatography and / or mass spectrometry).In some cases, reference sample sequences were identified for use in generating presentation data by determining the sequence of one or more cell lines engineered to express one or more MHC alleles (e.g., those detected in a disease sample), which may include MHC class I and / or MHC class II alleles. The one or more cell lines may include one or more human cell lines derived or collected from one or more subjects. For the purposes of this description, peptide sequences identified using a disease sample but not represented in the set of reference sample sequences may be identified as encoding a sequence variant.

[0246] In some embodiments, the collection of measures indicative of immunogenicity for use in training may be based on HLA typing, which can identify a subject-specific profile of MHC molecules. When the subject is a human, this profile may be referred to as a human leukocyte antigen (HLA) profile, where the HLA complex is a gene complex encoding MHC proteins in humans. HLA typing can be performed using a sample (e.g., a normal tissue sample and / or a non-disease-related sample) from the subject. The profile can be determined using a sequencing technique, such as PCR-based sequencing, direct sequencing, and / or next-generation sequencing.HLA typing may include, for example, high-resolution typing (e.g., excluding null alleles that are not expressed on the cell surface) or allele-level typing (e.g., which refers to the precise determination of the HLA gene nucleotide sequence). HLA typing analysis may include low-resolution typing and / or HLA supertyping, which identifies broader allele families.

[0247] For any type of sequence determination (e.g., to identify sequences in a sample, peptides associated with the MHC molecule, or HLA typing), the result may identify one or more nucleic acid sequences or one or more amino acid sequences. Once nucleic acid sequences are identified and an attention-based model (or other processing) is configured to process the amino acid sequences, a technique (e.g., a lookup table) can be used to convert individual codons in the nucleic acid sequences to individual amino acids.

[0248] Some embodiments include synthesizing a peptide (e.g., using a nucleic acid sequence encoding a peptide, such as a selected peptide) or a precursor for the selected peptide. The synthesized peptide or precursor can then be used in an experiment to identify a corresponding representation and / or binding data (e.g., to validate the predicted representation and / or binding or to generate results for use in training). For example, the experiment can include assessing the binding affinity of the selected peptide to a specific MHC molecule using an ELISA pull-down assay, gel shift assays, or a biosensor-based technique.As another example, an experiment may involve collecting elution data indicative of whether a selected peptide was presented by an MHC molecule via peptide-MHC immunoprecipitation, followed by elution and detection of the MHC ligands present using mass spectrometry.

[0249] In addition to or instead of training or validation data indicating whether individual peptides bound and / or were presented by individual MHCs, the training or validation data may indicate whether individual peptides induced immunogenicity. Immunogenicity results may be determined using in vivo or in vitro testing. Testing of one or more selected peptides may be configured to examine one or more factors of immunogenicity (e.g., to determine whether and / or the extent to which a specified event occurred) and / or immunogenicity (e.g., to determine whether and / or the extent to which a peptide induced an immune response).Testing can be configured to investigate whether administration of a composition (e.g., a vaccine) containing one or more peptides to a specified subject (e.g., for whom the MHC sequence used to select the mutant peptide was identified) is effective in preventing or treating a medical condition (e.g., a tumor) or disease (e.g., cancer). The subject may be a human subject.

[0250] Some embodiments include producing a composition based on one or more selected mutant peptides (or a plurality of nucleic acids encoding one or more selected mutant peptides). For example, each of the one or more selected mutant peptides could be predicted to bind to and be presented by an MHC molecule of the subject (e.g., at least to a threshold extent).The composition may comprise each of one or more selected mutant peptides, one or more precursors for one or more selected mutant peptides, one or more polypeptide sequences corresponding to one or more selected mutant peptides, RNA (e.g., mRNA) corresponding to one or more selected mutant peptides, DNA corresponding to one or more selected mutant peptides, cells (e.g., antigen-presenting cells) including one or more selected mutant peptides, and / or nucleic acid(s) encoding such peptides, plasmids corresponding to one or more selected mutant peptides, and / or vectors corresponding to one or more selected mutant peptides.

[0251] The composition may also contain an adjuvant, an auxiliary substance, an immunomodulator, a checkpoint protein, a PD-1 antagonist (e.g., an anti-PD-1 antibody) and / or a PD-L1 antagonist (e.g., an anti-PD-L1 antibody). The composition may be a vaccine, such as an anti-tumor vaccine. The composition may be a personalized vaccine manufactured or selected for a specific subject.

[0252] The composition may comprise a polynucleotide construct (e.g., a DNA construct or an RNA construct). A polynucleotide construct is an artificially constructed segment of nucleic acid that can be "transplanted" into a target tissue or target cell. The polynucleotide construct comprises a DNA or RNA (e.g., mRNA) insert that contains a nucleotide sequence encoding one or more selected mutant peptides. To enhance antigen presentation (e.g., presentation of one or more selected mutant peptides by an MHC molecule), the polynucleotide construct may also contain a modification designed to improve antigen presentation and thus improve the immunogenicity of one or more selected mutant peptides.In some cases, the modification is the inclusion of a transmembrane region and a cytoplasmic region of a chain of an MHC molecule in a polynucleotide construct, as described in International Application Publication WO2005038030A1, which is incorporated herein by reference in its entirety for all purposes.

[0253] To provide an RNA insert with increased stability and translation efficiency, the polynucleotide construct may also comprise a modification designed to improve stability and translation, and thus improve immunogenicity, for one or more selected mutant peptides. In some cases, the modification is the inclusion of a nucleic acid sequence with at least two copies of the 3'-untranslated region of the human beta-globin gene in the polynucleotide construct, as described in International Patent Publication WO2007036366A2, which is incorporated herein by reference in its entirety for all purposes. In other cases, the modification is the inclusion of a nucleic acid sequence that encodes a 3'-untranslated region, such as the F1 3' UTR, described in International Patent Publication WO2017060314A3, which is incorporated herein by reference in its entirety for all purposes.

[0254] To provide an RNA insert with increased stability and expression, the polynucleotide construct may also contain a modification designed to improve stability and expression, and thus improve immunogenicity, for one or more selected mutant peptides. In some cases, the modification is the inclusion of a cap at the end of the RNA, such as a 5' cap structure. The cap structure may be the diastereomer D1 beta-S-ARCA, as described in International Application Publication WO2011015347A1, which is incorporated herein by reference in its entirety for all purposes.

[0255] To deliver the polynucleotide construct with high selectivity to antigen-presenting cells, the composition may also include cationic liposomes or a lipoplex for improved uptake of the polynucleotide construct, and thus improved immunogenicity for one or more selected mutant peptides. In some cases, the composition comprises nanoparticles containing the polynucleotide construct. The nanoparticles may be lipoplexes containing one or more lipids, such as DOTMA and DOPE, as described in International Application Publication WO2013143683A1, which is incorporated herein by reference in its entirety for all purposes.

[0256] Some embodiments include treating a medical condition (e.g., a tumor) or a disease (e.g., cancer) in an individual by administering to the individual an effective amount of a composition (e.g., a vaccine) comprising one or more selected mutant peptides. The individual may be the same individual from whom the disease sample was collected. In some cases, the vaccine is administered to a different individual relative to the individual from whom the disease sample was collected. The different individual may, for example, be a relative of the individual from whom the disease sample was collected, have a genetic risk for developing a particular type of cancer, and / or have MHC molecules that have one, several, or all alleles corresponding to a sequence that is the same (or similar) to one or more MHC alleles of the subject from whom the disease sample was collected.

[0257] In some embodiments, for each set of mutant peptides (e.g., detected in a sample from a subject), one or more techniques disclosed herein are used to predict whether the mutant peptide will bind to the subject's MHC molecule (or the strength, stability, and / or prevalence of such binding) and / or to predict whether the subject's MHC molecule will present the mutant peptide (and / or the prevalence of such presentation). The predictions can be used to select a subset of mutant peptides (e.g., for which the MHC presentation of the mutant peptide is predicted to be likely). The selection can include comparing, for each mutant peptide, a measure corresponding to a prediction measure for an absolute threshold value and / or a prediction measure of another measure of mutant peptides (e.g., thereby performing a relative comparison).Each selected mutant peptide can be identified as having: a high probability of being presented on the surface of a tumor cell; a high probability of being able to elicit a tumor-specific immune response; a high probability of being able to be presented to naive T cells by professional antigen-presenting cells (e.g., dendritic cells); a low probability of being subject to inhibition via central or peripheral tolerance; and / or a low probability of being able to induce an autoimmune response to normal tissue in a subject.

[0258] Some embodiments include generating and / or using a model to identify one or more peptides (e.g., mutant peptides) that are likely to bind to MHC molecules and will be presented by MHC molecules on the surfaces of tumor cells. More specifically, the training data set may comprise a set of data elements, wherein each data element includes: an epitope (or peptide) sequence (e.g., and potentially the sequences of the N-flank of the peptide and the C-flank of the peptide), a subsequence of the MHC molecule, and one or more experimental results relating to the peptide and the MHC molecule (e.g., binding affinity and / or eluted ligand presentation data).

[0259] An attention-based machine learning model can be trained using at least a portion of a training data set. The training data set can contain multiple training data items. Each training data item can contain a representation of a sequence and an outcome (e.g., indicating whether at least a portion of a peptide corresponding to the sequence is represented by an MHC molecule and / or whether it causes immunogenicity). Training data items for which a representation was not found can be computationally generated. For example, for each protein of origin from the positive set (corresponding to the positive representation data of the eluted ligand), one, several, or all possible peptide fragments (e.g., in a predetermined length range, e.g., from 8 to 11) can be generated, potentially with a uniform probability for each length.The N-terminal and C-terminal flanking sequences may be conserved (e.g., potentially with a maximum length of 10 amino acids). In some cases, peptide fragments (e.g., one, several, or all of lengths 8:11) may be generated for each allele represented in positive cases in the training data. Generation and / or subsequent selection may be conducted such that the probability of generating a sequence of a given length is uniform across lengths. The N-terminal and C-terminal flanking sequences may or may have been conserved with a specific maximum length (e.g., a maximum length of 10 amino acids).

[0260] The attention-based machine learning model may comprise 1, 2, 3, 4, 5, 6, 7, 8 or more transformer encoder networks (e.g., each comprising a unidirectional attention network and a feedforward network). For example, the attention-based machine learning model may comprise a plurality of first-level transformer encoders, including a transformer encoder configured to process a peptide representation, a transformer encoder configured to process a MHC molecule representation, potentially a transformer encoder configured to process a peptide N-flank representation, and potentially a transformer encoder configured to process a peptide C-flank representation. The attention-based machine learning model may further comprise a second-level transformer encoder configured to process aggregated (e.g., concatenated) results generated by the first-level transformer encoders.

[0261] The attention-based machine learning model may also comprise a feedforward network (e.g., a fully connected feedforward network with one, two, or more hidden layers) configured to process the output from the fifth encoder of the transformer (e.g., after applying dropout) to generate a predicted (e.g., real-valued) binding affinity and / or a predicted representation (e.g., as a binary prediction). The attention-based machine learning model may be one or multiple models (e.g., with the same configuration) in a model suite. The training dataset may be arbitrarily split, shuffled, and / or divided to train different models in the suite.The loss function can use an error term (e.g., root-mean-squared error or mean squared error) and / or an entropy term (e.g., cross-entropy or binary cross-entropy). Multi-task learning can be used, where the model is simultaneously trained to predict each of two different types of outcomes (e.g., binding affinity and representation emergence). A static or non-static learning rate can be used. For example, learning rate equalization (e.g., using stepwise equalization or cosine equalization) can be used to reduce the learning rate across iterations. Validation data evaluation can be used to potentially terminate training early (e.g., when determining that the performance goal has been achieved).

[0262] An MHC molecule comprises multiple alleles in vivo (e.g., 6 alleles per individual). Therefore, for this single MHC molecule, multiple sequence inputs can be generated (e.g., each representing one allele from the set of alleles). Each of the multiple sequence inputs can be individually processed by one or more neural networks (e.g., one or more transducer encoders) to generate a predicted neoantigen binding or presentation value associated with each of the alleles. A function (e.g., a softmax function) can identify which allele among the set of alleles is associated with the highest presentation prediction. During training, this maximum presentation prediction for this specific sequence input can then be compared to the true presentation value using a binary loss function to generate an error term for the tuning parameters.

[0263] In some cases, it is unknown how many amino acids from a flank (e.g., the N-flank) are used by peptidases to determine when to trim long peptides in the core of a presented peptide. To address this uncertainty in training data generation, flanks can then be trimmed to a length selected based on a technique (e.g., a pseudorandom selection technique), such as a length in a predetermined range (e.g., 1-10 amino acids). The selection technique can select the length using a distribution (e.g., a uniform or Gaussian distribution). In some cases, a flank that is below a threshold length (e.g., 10 amino acids) is not trimmed. In some cases, flank trimming is determined to preserve the C-side of the N-flank.

[0264] The trained model can then receive a set of input data that contains the representation(s) of one or more mutant peptide sequences (e.g., the N-flanking region, the candidate epitope region, and / or the C-flanking region) and a subsequence of an MHC molecule (bound to the subject), and generate a predicted binding affinity and / or a presentation prediction. If the mutant peptide is predicted to stably bind and be presented by the MHC molecule, the mutant peptide can be selected for inclusion in a composition (e.g., a vaccine) to be used to treat the subject. II.G. Typical identification of input data for a machine learning model

[0265] The typical input data identification methods and systems described herein can be used to identify input data for, for example, the machine learning model 132 of Figs. 1 and 3 and / or the machine learning model 132 described in Figs. 4A-4B.

[0266] Each of a set of mutant peptides associated with a specified subject may be analyzed using an attention-based machine learning model to generate one or more predictions regarding the binding affinity, presentation probability, and / or immunogenicity of the mutant peptide. To generate these predictions, the machine learning model may receive and process a peptide (e.g., coding) sequence corresponding to the mutant peptide and one or more other sequences or subsequences (e.g., corresponding to an MHC-I molecule, an MHC-II molecule, or a T-cell receptor). In some cases, predictions are generated for each of a set of peptide sequences (e.g., a set of variant coding sequences corresponding to a set of mutant peptides).The set of mutant peptides may correspond to peptides present in a disease sample collected from a subject, but which are not observed in one or more non-disease samples (e.g., from the subject or another subject).

[0267] Various methods are available to identify a set of mutant peptides associated with a given subject. The mutations may be present in the genome, transcriptome, proteome, or exome of disease-relevant cells of the subject, but not in a non-disease-relevant sample, such as a non-disease-relevant sample from the subject or from another subject. Mutations include, but are not limited to: (1) non-synonymous mutations resulting in different amino acids in the protein; (2) read-through mutations in which the stop codon is modified or deleted, resulting in translation of a longer protein with a new tumor-specific sequence at the C-terminus; (3) splice site mutations that result in the inclusion of an intron in the mature mRNA and thus a unique tumor-specific protein sequence; (4) chromosomal rearrangements that produce a chimeric protein with tumor-specific sequences at the junction of the 2 proteins (i.e.(5) frameshift insertions or deletions that result in a new open reading frame with a new tumor-specific protein sequence. Mutations may also include one or more of a non-frameshift insertion, a nonsense or nonsense substitution, a splice site change, a genomic rearrangement or gene fusion, or any genomic or expression alteration resulting in a neoORF.

[0268] Peptides with mutations or mutated polypeptides resulting from, for example, splice site mutations, frameshifts, read-through or gene fusions in disease-related cells can be identified by determining the sequence of DNA, RNA or protein in a disease-related sample and comparing the obtained sequences with sequences from a non-disease-related sample.

[0269] In some embodiments, whole genome sequencing (WGS) or whole exome sequencing (WES) data from a disease-related sample and a non-disease-related sample can be obtained and compared. After aligning the reads of the non-disease-related sample and the disease-related sample to a reference human genome, somatic variants that contain single nucleotide variants (SNVs), gene fusions, and insertion or deletion variants (insertions) can be detected using variant-based algorithms. One or more variant callers can be used to detect different somatic variant types (i.e., SNVs, gene fusions, or insertions) (see Xu et al. "A review of somatic single nucleotide variant calling algorithms for next-generation sequencing data." Comput. Struct. Biotechnol. J. 16: 15–24 (2018), which is hereby incorporated by reference in its entirety for all purposes).

[0270] In some examples, mutant peptides are identified based on transcriptome sequences in a disease sample from an individual. For example, all or partial transcriptome sequences (e.g., using methods such as RNA-Seq) can be obtained from disease-related tissue of an individual and subjected to sequencing analysis. Sequences obtained from the disease-related tissue sample can then be compared with those obtained from a reference sample. Optionally, the disease-related tissue sample is subjected to whole-transcriptome RNA-Seq. Optionally, transcriptome sequences are "enriched" for specific sequences before comparison with the reference sample. For example, specific probes can be designed to enrich for certain desired sequences (e.g., disease-specific sequences) before sequencing analysis.Methods for whole-transcriptome sequencing and targeted sequencing are known in the art and are cited, for example, in Tang, F. et al., "mRNA-Seq whole-transcriptome analysis of a single cell", Nature Methods, 2009, v. 6, 377-382; Ozsolak, F., "RNA sequencing: advances, challenges and opportunities", Nature Reviews, 2011, v. 12, 87-98; German, M. A et al., "Global identification of microRNA-target RNA pairs by parallel analysis of RNA ends", Nature Biotechnology, 2008, v. 26, 941-946; and Wang, Z. et al., "RNA-Seq: a revolutionary tool for transcriptomics", Nature Reviews, 2009, v. 10, pp. 57-63. Each of these references is hereby incorporated by reference in its entirety for all purposes.

[0271] In some embodiments, transcriptome sequencing techniques include, but are not limited to, poly(A) RNA libraries, microarray analysis, parallel sequencing, massively parallel sequencing, PCR, and RNA-Seq. RNA-Seq is a high-throughput technique for sequencing a portion or substantially the entire transcriptome. Briefly, an isolated population of transcriptome sequences is converted into a library of cDNA fragments with adaptors attached to one or both ends. With or without amplification, each cDNA molecule is then analyzed to obtain short stretches of sequence information, typically 30-400 base pairs. These fragments of sequence information are then aligned with a reference genome, reference transcripts, or assembled de novo to discover the structure of transcripts (i.e., transcription boundaries) and / or expression levels.

[0272] Once obtained, sequences in the disease-related sample can be compared with corresponding sequences in a reference sample. Sequence comparison can be performed at the nucleic acid level, by aligning nucleic acid sequences in the disease-related tissue with corresponding sequences in the reference sample. Variations in the genetic sequence that result in one or more changes in the encoded amino acids are then identified. Alternatively, sequence comparison can be performed at the amino acid level, namely, nucleic acid sequences are first converted to amino acid sequences in silico before comparison. Either an amino acid-based approach or a nucleic acid-based approach can be used to identify one or more mutations (e.g., single-point or multipoint mutations) in a peptide.For nucleic acid-based approaches, open variants can be used to identify one or more nucleic acid sequences (e.g., DNA sequences, RNA sequences, or mRNA sequences) that will result in a specified observed mutant protein (e.g., via a lookup table that associates individual peptide mutations with a set of codon variants).

[0273] In some embodiments, comparison of the sequence from the disease sample with that of the reference sample can be performed using techniques known in the art, such as manual alignment, FAST-A11 (FASTA), and Basic Local Alignment Search Tool (BLAST). In some embodiments, comparison of the sequence from the disease sample with that of the reference sample can be performed using a short-read aligner, such as GSNAP, BWA, and STAR.

[0274] In some embodiments, the reference sample is a matched disease-free sample. As used herein, a "matched" disease-free tissue sample is one that is selected from the same or similar sample, for example, a sample from the same or similar tissue type as the disease sample. In some embodiments, the matched disease-free tissue and the disease-related tissue can be obtained from the same individual. The reference sample described herein, in some embodiments, is a disease-free sample from the same individual. In some embodiments, the reference sample is a disease-free sample from another individual (e.g., an individual without the disease). In some embodiments, the reference sample is obtained from a population of different individuals.In some embodiments, the reference sample is a database of known genes associated with the organism. In some embodiments, the reference sample may be from a cell line. In some embodiments, the reference sample may be a combination of known genes associated with the organism and genomic information from a matched sample without the disease. In some embodiments, the coding sequence of the variant may contain a point mutation in the amino acid sequence. In some embodiments, the coding sequence of the variant may contain an amino acid deletion or insertion.

[0275] In some embodiments, a set of variant-coding sequences is first identified based on the genomic sequence and / or nucleic acid sequence. This initial set is then further filtered to obtain a narrower set of expressed variant-coding sequences based on the presence of the variant-coding sequences in a transcriptome sequencing database (and is thus considered "expressed"). In some embodiments, the set of variant-coding sequences is reduced by at least about 10, 20, 30, 40, 50, or more times by filtering through the transcriptome sequencing database.

[0276] Alternatively, protein mass spectrometry can be used to identify or validate the presence of mutant peptides, such as mutant MHC binding proteins on tumor cells. Peptides can be acid-eluted from disease-related cells, such as tumor cells or HLA molecules that immunoprecipitate from the tumor, and then identified using mass spectrometry.

[0277] The mutant peptide may have, for example, 5 or more, 8 or more, 11 or more, 15 or more, 20 or more, 40 or more, 80 or more, 100 or more, 120 or less, 100 or less, 80 or less, 60 or less, 50 or less, 40 or less, 30 or less, 25 or less, 20 or less, 18 or less, 15 or less, or 13 or less amino acids.

[0278] Tumor-specific T cell receptor sequences can also be identified, for example, by single-cell T cell receptor sequencing. See, for example, De Simone et al. "Single Cell T Cell Receptor Sequencing: Techniques and Future Challenges," Front. Immunol. 9: 1638 (2018); Zong et al. "Very rapid cloning, expression, and identifying specificity of T-cell receptors for T-cell engineering," PloS ONE 15(2):e0228112 (2020) (which is hereby incorporated by reference in its entirety for all purposes). High-throughput sequencing of T cell assemblies can additionally or alternatively be performed to identify tumor-specific signatures for a particular disease. See, for example, Wang et al. "High-throughput sequence of CD4+T cell repertoire reveals disease-specific signatures in IgG4-related disease," Arthritis Research & Therapy 21: 295 (2019) (which is hereby incorporated by reference in its entirety for all purposes).

[0279] MHC-I sequences and / or MHC-II sequences can be determined, for example, by HLA genotyping or mass spectroscopy (Caron et al., "Analysis of Major Histocompatibility Complex (MHC) Immunopeptides Using Mass Spectroscopy," Molecular and Cellular Proteomics 14(12): 3105–3117 (2015) (which is hereby incorporated by reference in its entirety for all purposes). II.D Typical identification of training data for a machine learning model

[0280] The typical methods and systems for identifying training data described in this document can be used to identify training data for, for example, the machine learning model 132 in Figs. 1 and 3 and / or the machine learning model 132 described in Figs. 4A-4B. For example, these methods and systems can be used to identify the training data 131 in Fig. 1.

[0281] The training set can be generated using data collected from a plurality of other samples (e.g., potentially related to one or more other subjects). Each of the plurality of other samples may include, for example, tissue (e.g., a biopsy), a single cell, a plurality of cells, cell fragments, or an aliquot of bodily fluid. In some cases, the plurality of other samples are selected from a different type of subject than the subject related to the input data for processing by the trained model. For example, a machine learning model can be trained using training data collected by processing samples from one or more cell lines, and the trained machine learning model can be used to process input data determined by processing one or more samples from a human subject.

[0282] The training data set may include a plurality of training elements. Each of the plurality of training elements may include input data that comprises a set of peptide sequences (which comprises a set of either wild-type or variant-encoding sequences), each of which encodes and / or represents any variant in the corresponding peptide, and a subsequence or pseudosequence of an MHC molecule. The input data may be collected according to one or more techniques disclosed herein (e.g., in Section II.D).

[0283] Each training element may also include one or more experimental outcomes. An experimental outcome may indicate whether and / or to what extent each of one or more specific types of interactions will occur between a wild-type or mutant peptide (bound to a variant-coding sequence in the training element) and an MHC molecule (bound to a subsequence of an MHC molecule in the training element). A specific type of interaction may include, for example, binding of a peptide to an MHC molecule and / or presentation of a peptide by an MHC molecule on the surface of a cell (e.g., a tumor cell).

[0284] The result may include the binding affinity between a peptide and an MHC molecule. The result may include or be based on quantitative data and / or qualitative data characterizing whether the specified peptide binds to the specified MHC molecule, the strength of such binding, the stability of such binding, and / or the tendency for such binding to occur. For example, a binary binding affinity indicator or a qualitative binary affinity result can be generated using ELISA, a pull-down assay, a gel shift assay, a biosensor-based technique such as surface plasmon resonance, isothermal titration colorimetry, biolayer interferometry, or microscale thermophoresis.

[0285] The result may, for example, additionally or alternatively characterize whether and / or the probability that the specified MHC molecule represents the specified peptide. MHC ligands may form an immunoprecipitate from the sample. Subsequent elution and mass spectrometry can be used to determine whether the MHC molecule represents the ligand.III. Pharmaceutically acceptable composition and manufacture

[0286] One or more variant-coding sequences may be selected from a set of subject-specific set of variant-coding sequences based on the results of one or more machine learning models described herein. For example, the selection may include identifying each of the set of subject-specific set of variant-coding sequences for which the predicted binding affinity is less than 500 nM, for which the MHC molecule is predicted to represent the mutant peptide identified by the variant-coding sequence, and / or for which the mutant peptide is predicted to elicit an immune response. It will be appreciated that the model results may be on different scales, such that 500 nM may correspond to, for example, another value (e.g., 0.42) on the [0,1] scale.

[0287] A pharmaceutically acceptable composition can be designed and / or manufactured using one, several, or all of the selected variant coding sequences. The composition can include mutant peptides corresponding to one variant coding sequence. The composition can include mutant peptides and / or mutant peptide precursors corresponding to multiple selected variant coding sequences. A subset of candidate peptides (e.g., those associated with 5, 10, 15, 20, 30, or any number in between, the highest representation predictions) can be used for further development of the precursor.

[0288] Each of one, more or all of the mutant peptides in the composition can have, for example, a length of from about 7 to about 40 amino acids (e.g., about any of 7, 8, 9, 10, 11, 12, 13, 14, 15, 17, 20, 22, 25, 30, 35, 40, 45, 50, 60 or 70 amino acids in length). In some embodiments, the length of each of one, more or all of the mutant peptides in the composition is within a predetermined range (e.g., 8-11 amino acids, 8-12 amino acids or 8-15 amino acids). In some embodiments, each of one, more or all of the mutant peptides in the composition is about 8-10 amino acids in length. Each of one, more or all of the mutant peptides in the compositions can be in their isolated form. Each of one, more than one, or all of the mutant peptides in the compositions may be a "long peptide" produced by adding one or more peptides to the end (or to each end) of a mutant peptide.Each of one, more than one, or all of the mutant peptides in the compositions may be labeled, may be a fusion protein, and / or may be a hybrid molecule.

[0289] A pharmaceutically acceptable composition can be designed and / or manufactured to include or by using one or more nucleic acids that encode, for each one, several, or all of the selected variant coding sequences, a peptide that comprises or is composed of amino acids as defined in the variant coding sequence. The nucleic acid(s) can include DNA, RNA, and / or mRNA. Provided that any of the plurality of codons can encode the specified amino acids, the codons can be selected to, for example, optimize or activate expression in the specified type of organism. Such selection can be based on the frequency of use of each of the plurality of potential codons by the specified type of organism, the translation efficiency of each of the plurality of potential codons in the specified type of organism, and / or the specified type of degree of bias of the organism in the direction of each of the plurality of potential codons.

[0290] In some cases, the composition may include nucleic acids encoding the mutant peptide(s) or precursor of the mutant peptide(s) described above. The nucleic acid may include sequences flanking the sequence encoding the mutant peptide (or its precursor). In some cases, the nucleic acid contains epitopes corresponding to more than one selected variant coding sequence. In some cases, the nucleic acid is DNA having a polynucleotide sequence encoding the mutant peptides or precursors described above.

[0291] In some cases, the nucleic acid is RNA. In some cases, the RNA is reverse transcribed from a DNA template with a polynucleotide sequence encoding the mutant peptides or precursors described above. In some cases, the RNA is mRNA. In some cases, the RNA is naked mRNA. In some cases, the RNA is modified mRNA (e.g., mRNA protected from degradation by protamine, mRNA containing a modified 5'CAP structure, or mRNA containing modified nucleotides). In some embodiments, the RNA is single-stranded mRNA.

[0292] The composition may comprise cells comprising the mutant peptide and / or nucleic acid(s) encoding the mutant peptide described above. The composition may also comprise one or more suitable vectors and / or one or more delivery systems for the mutant peptide and / or nucleic acid(s) encoding the mutant peptide. In some cases, the cells comprising the mutant peptide and / or nucleic acids encoding the mutant peptide are non-human cells, such as bacterial cells, protozoan cells, fungal cells, or non-human animal cells. In some cases, the cells comprising the mutant peptide and / or nucleic acids encoding the mutant peptide are human cells. In some cases, the human cells are immunocytes. In some cases, the immunocytes are antigen-presenting cells (APCs).In some cases, the APCs are professional APCs, such as macrophages, monocytes, dendritic cells, B cells, and microglia. In other cases, the professional APCs are macrophages or dendritic cells. In some cases, APCs containing mutant peptides and / or nucleic acid sequence(s) encoding the mutant peptide are used as a cellular vaccine, activating a CD4+ or CD8+ immune response. In other cases, the composition used as a cellular vaccine comprises T cells specific for the mutant peptide, primed by APCs containing the mutant peptide and / or nucleic acid sequence(s) encoding the mutant peptide.

[0293] The composition may contain a pharmaceutically acceptable adjuvant and / or a pharmaceutically acceptable excipient. Adjuvants refer to any substance whose addition to the composition modifies the immune response to the mutant peptide. Adjuvants can be combined using, for example, an immune stimulating agent. Excipients can increase the molecular weight of a particular mutant peptide to enhance activity or immunogenicity, impart stability, enhance biological activity, and / or increase serum half-life.

[0294] The pharmaceutically acceptable composition may be a vaccine, which may comprise a personalized vaccine that is specific to (e.g., and potentially developed for) a particular subject. For example, the MHC sequence may have been identified using a sample from a particular subject, and the composition may be developed for and / or used to treat a particular subject.

[0295] The vaccine may be a nucleic acid-based vaccine. The nucleic acid may encode a mutant peptide or a precursor of the mutant peptide. The nucleic acid-based vaccine may include sequences flanking the sequence encoding the mutant peptide (or its precursor). In some cases, the nucleic acid-based vaccine contains epitopes corresponding to more than one selected variant-coding sequence. In some cases, the nucleic acid-based vaccine is a DNA-based vaccine. In some cases, the nucleic acid-based vaccine is an RNA-based vaccine. In some cases, the RNA-based vaccine contains mRNA. In some cases, the RNA-based vaccine contains naked mRNA. In some cases, the RNA-based vaccine contains modified mRNA (e.g., mRNA protected from degradation using protamine).mRNA containing a modified 5'CAP structure, or mRNA containing modified nucleotides). In some embodiments, the RNA-based vaccine comprises single-stranded mRNA.

[0296] A nucleic acid-based vaccine may contain a personalized neoantigen-specific therapeutic tailored to a specific subject for use as part of next-generation immunotherapy. The personalized vaccine may be developed by first detecting mutant peptides in a specific subject's sample and then predicting, for each detected mutant peptide, whether and to what extent the peptide will bind to the specific subject's MHC, be presented by the MHC, bind to the specific subject's T-cell receptor, and / or induce an immune response. Based on these predictions, a subset of the detected mutant peptides can be selected (e.g., a subset with at least 1, at least 2, at least 3, at least 5, at least 8, at least 10, at least 12, at least 15, at least 18, up to 40, up to 30, up to 25, up to 20, up to 18, up to 15 and / or up to 10 mutant peptides).For each selected mutant peptide, a synthetic mRNA sequence encoding the mutant peptide can be identified. An mRNA-based vaccine can contain mRNA (encoding part or all of the mutant peptide) complexed with lipids to form an mRNA lipoplex. Administration of the mRNA lipoplex-containing vaccine can produce mRNA that stimulates TLR7 and TLR8, leading to T-cell activation by dendritic cells. Furthermore, administration can result in translation of the mRNA into the mutant peptide, which can then bind and be presented by MHC molecules, eliciting a T-cell response.

[0297] The composition may comprise substantially pure mutant peptides, substantially pure precursors thereof, and / or substantially pure nucleic acids encoding the mutant peptides or their precursors. The composition may comprise one or more suitable vectors and / or one or more delivery systems for containing the mutant peptides, their precursors, and / or nucleic acids encoding the mutant peptides or their precursors. Suitable vectors and delivery systems include viral ones, such as those based on adenovirus, vaccinia virus, retroviruses, herpes virus, adeno-associated virus, or hybrids containing elements of more than one virus. Non-viral delivery systems include cationic lipids and cationic polymers (e.g., cationic liposomes). In some embodiments, physical delivery may be used, for example, using a "gene gun."

[0298] In some embodiments, the RNA vaccine comprises an RNA molecule including, in 5'→3' direction: (1) a 5' cap; (2) a 5' untranslated region (UTR); (3) a polynucleotide sequence encoding a secretory signal peptide; (4) a polynucleotide sequence encoding one or more mutant peptides derived from cancer-specific somatic mutations present in a tumor sample; (5) a sequence encoding at least a portion of the transmembrane and cytoplasmic domain of a major histocompatibility complex (MHC) molecule; (6) a 3' UTR including: (a) the 3' untranslated region of the amino-terminal cleavage enhancer (ACTE) mRNA or a fragment thereof; and (b) non-coding RNA of mitochondrial encoded 12S RNA or a fragment thereof; and (7) a poly(A) sequence. This example RNA molecule was also used in the evaluation of a typical implementation of the attention-based prediction model, as discussed in relation to Section V below.

[0299] In some embodiments, the RNA molecule further comprises a polynucleotide sequence encoding an amino acid linker; wherein the polynucleotide sequences encoding the amino acid linker and the first of the one or more mutant peptides form a first module of the neoepitope linker; and wherein the polynucleotide sequences forming the first module of the neoepitope linker are located between the polynucleotide sequence encoding the secretory signal peptide and the polynucleotide sequence encoding at least a portion of the transmembrane and cytoplasmic domain of the MHC molecule in the 5'→3' direction. In some embodiments, the amino acid linker comprises the sequence GGSGGGGSGG. In some embodiments, the polynucleotide sequence encoding the amino acid linker comprises the sequence GGCGGCUCUGGAGGAGGCGGCUCCGGAGGC.

[0300] In some embodiments, the RNA molecule further comprises, in the 5'→3' direction: at least a second linker-epitope module, wherein the at least second linker-epitope module comprises a polynucleotide sequence encoding an amino acid linker and a polynucleotide sequence encoding a neoepitope; wherein the polynucleotide sequences forming the second linker-neoepitope module are located between the polynucleotide sequence encoding the neoepitope of the first linker-neoepitope module and the polynucleotide sequence encoding at least a portion of the transmembrane and cytoplasmic domain of the PSGS molecule in the 5'→3' direction, and wherein the neoepitope of the first linker-epitope module differs from the neoepitope of the second linker-epitope module. In some embodiments, the RNA molecule comprises 5 linker-epitope modules, wherein each of the 5 linker-epitope modules encodes different neoepitopes.In some embodiments, the RNE molecule comprises 10 linker-epitope modules, wherein each of the 10 linker-epitope modules encodes a different neoepitope. In some embodiments, the RNE molecule comprises 20 linker-epitope modules, wherein each of the 20 linker-epitope modules encodes a different neoepitope.

[0301] In some embodiments, the RNE molecule further comprises a second polynucleotide sequence encoding an amino acid linker, wherein the second polynucleotide sequence encoding the amino acid linker is located between the polynucleotide sequence encoding the neoepitope that is most distal in the 3' direction and the polynucleotide sequence encoding at least a portion of the transmembrane and cytoplasmic domain of the MHC molecule.

[0302] In some embodiments, the 5' cap comprises diastereoisomer D1 of the structure:

[0303] In some embodiments, the 5' UTR comprises the sequence UUCUUCUGGUCCCCACAGACUCAGAGAGAACCCGCCACC. In some embodiments, the 5' UTR comprises the sequence GGCGAACUAGUAUUCUUCUGGUCCCCACAGACUCAGAGAGAACCCGCCACC.

[0304] In some embodiments, the secretory signal peptide comprises the amino acid sequence MRVMAPRTLILLLSGALALTETWAGS. In some embodiments, the polynucleotide sequence encoding the secretory signal peptide comprises the sequence AUGAGAGUGAUGGCCCCCAGAACCCUGAUCCUGCUGCUGUCUGGCGCCCUGGCCCUGACAGAGACAUGGGCCGGAAGC.

[0305] In some embodiments, at least a portion of the transmembrane and cytoplasmic domain of the MHC molecule comprises the amino acid sequence IVGIVAGLAVLAVVVIGAVVATVMCRRKSSGGKGGSYSQAASSDSAQGSDVSLTA. In some embodiments, the polynucleotide sequence encoding at least a portion of the transmembrane and cytoplasmic domain of the MHC molecule comprises the sequence AUCGUGGGAAUUGUGGCAGGACUGGCAGUGCUGGCCGUGGUGGUGAUCGGAGCCGUGGUGGCUACCGUGAUGUGCAGACGGAAGUCCAGCGGAGGCAAGGGCGGCAGCUACAGCCAGGCCGCCAGCUCUGAUAGCGCCCAGGGCAGCGAC GUGUCACUGACAGCC.

[0306] In some embodiments, the 3' untranslated region of the AES mRNA comprises the sequence CUGGUACUGCAUGCACGCAAUGCUAGCUGCCCCUUUCCCGUCCUGGGUACCCCGAGUCUCCCCCGACCUCGGGUCCCAGGUAUGCUCCCACCUCCACCUGCCCCACUCACCACCUCUGCUAGUUCCAGACACCUCC. In some embodiments, the non-coding RNA of the mitochondrial-encoded 12S RNA comprises the sequence CAAGCACGCAGCAAUGCAGCUCAAAACGCUUAGCCUAGCCACACCCCCACGGGAAACAGCAGUGAUUAACCUUUAGCAAUAAACGAAAGUUUAACUAAGCUAUACUAACCCCAGGGUUGGUCAAUUUCGUGCCAGCCACACCG. In some embodiments, the 3' UTR comprises the sequence CUCGAGCUGGUACUGCAUGCACGCAAUGCUAGCUGCCCCUUUCCCGUCCUGGGUACCCCGAGUCUCCCCCGACCUCGGGUCCCAGGUAUGCUCCCACCUCCACCUGCCCCACUCACCACCUCUGCUAGUUCCAGACACCUCCCAAGCACGCAGCAAUGCAGCUCAAAACGCUUAGCCUAGCCACACCCCCACGGGAAACAGCAGUGAUUAACCUUUAGCAAUAAACGAAAGUUUAACUAAGCUAUACUAACCCCAGGGUUGGUCAAUUUCGUGCCAGCCACACCGAGACCUGGUCCAGAGUCGCUAGCCGCGUCGCU.

[0307] In some embodiments, the poly(A) sequence comprises 120 adenine nucleotides.

[0308] In some embodiments, the RNA-based vaccine comprises an RNA molecule comprising, in the 5'→3' direction: a polynucleotide sequence GGCGAACUAGUAUUCUUCUGGUCCCCACAGACUCAGAGAGAACCCGCCACCAUGAGAGUGAUGGCCCCCAGAACCCUGAUCCUGCUGCUGUCUGGCGCCCUGGCCCUGACAGAGACAUGGGCCGGAAGC; a polynucleotide sequence encoding one or more mutant peptides derived from cancer-specific somatic mutations present in a tumor sample; and a polynucleotide sequenceAUCGUGGGAAUUGUGGCAGGACUGGCAGUGCUGGCCGUGGUGGUGAUCGGAGCCGUGGUGGCUACCGUGAUGUGCAGACGGAAGUCCAGCGGAGGCAAGGGCGGCAGC UACAGCCAGGCCGCCAGCUCUGAUAGCGCCCAGGGCAGCGACGUGUCACUGACAGCCUAGUAACUCGAGCUGGUACUGCAUGCACGCAAUGCUAGCUGCCCCUUUCCCGUCCUGGGUACCCCGAGUC UCCCCCGACCUCGGGUCCCAGGUAUGCUCCCACCUCCACCUGCCCCACUCACCACCUCUGCUAGUUCCAGACACCUCCCAAGCACGCAGCAAUGCAGCUCAAAACGCUUAGCCUAGCCACACCCCCA CGGGAAACAGCAGUGAUUAACCUUUAGCAAUAAACGAAAGUUUAACUAAGCUAUACUAACCCCAGGGUUGGUCAAUUUCGUGCCAGCCACACCGAGACCUGGUCCAGAGUCGCUAGCCGCGUCGCU.

[0309] In some embodiments, the mutant peptides described herein (e.g., comprising or consisting of an ordered set of amino acids as identified by variant-encoding sequences selected based on the results of the machine learning technique described herein) can be used to generate mutant peptide-specific therapeutic agents, such as antibody-based therapeutic agents. For example, mutant peptides can be used to elicit and / or identify antibodies that specifically recognize mutant peptides. These antibodies can be used as therapeutic agents. Synthetic short peptides have been used to generate protein-reactive antibodies. An advantage of immunization with synthetic peptides is that an unlimited amount of pure, stable antigen can be used.This approach involves synthesizing short peptide sequences, combining them with a large carrier molecule, and immunizing the subject with the peptide carrier molecule. Antibody properties depend on the primary sequence information. A good response to a desired peptide can usually be generated by carefully selecting the sequence and combination method. Most peptides can elicit a good response. The advantage of peptide antibodies is that they can be generated immediately after determining the amino acid sequence of the mutant peptide, and specific regions of the protein can be specifically targeted for antibody production. Selecting mutant peptides for which a machine learning model has predicted immunogenicity and / or selecting for it can result in a high chance that the resulting antibody will recognize the naive protein in the tumor setting.The mutant peptide may consist of, for example, 15 or less, 18 or less, or 20 or less, 25 or less, 30 or less, 35 or less, 40 or less, 50 or less, 60 or less, 70 or less, 85 or less, 100 or less, 110 or less residues. The mutant peptide may consist of, for example, 9 or more, 10 or more, 15 or more, 20 or more, 25 or more, 30 or more, 50 or more, or 70 or more residues. Shorter peptides may improve antibody production.

[0310] A peptide-carrier protein combination can be used to facilitate the production of high antibody titers. The coupling method can include, for example, site-directed coupling and / or a technique that is based on reactive functional groups in amino acids, such as -NH2, -COOH, -SH, and phenolic -OH. Any suitable method used in the production of antibodies to a peptide can be used with the mutant peptides identified by the methods of the present invention. Two such known methods are the multiple antigenic peptide (MAP) system and lipid core peptides (LCP method). The advantage of MAP is that a conjugation method is not required. No carrier protein or linker bond is introduced into the immunized host. One disadvantage is that the peptide purity is more difficult to control. In addition, MAPs can evade the immune response system in some hosts.The PLN method is known to provide higher titers than other peptide antibody-based vaccine systems and may thus be preferred.

[0311] The present document also provides isolated MHC / peptide complexes comprising one or more mutant peptides identified using the technique disclosed herein. Such MHC / peptide complexes can be used, for example, to identify antibodies, soluble TCRs, or TCR analogs. One type of these antibodies has been termed TCR mimics because they are antibodies that bind peptides from tumor-specific antigens in the context of specific HLA environments. This type of antibody has been shown to mediate lysis of cells expressing the complex on their surface and also protect mice from implanted cancer cell lines that express the complex (see, for example, Wittman et al., J. of Immunol. 177:4187–4195 (2006)). One advantage of TCR mimics as IgG mAbs is that affinity maturation can be performed and the molecules bind to immune effector functions via the exposed Fc domain.These antibodies can also be used to target therapeutic molecules to tumors, such as toxins, cytokines, or drugs.

[0312] Other types of molecules that have been developed using mutant peptides, such as those selected using the methods of the present invention, using non-hybridoma-based antibody production or the production of binding-competent antibody fragments, such as Fab molecules of an antibody to a peptide on a bacteriophage. These fragments can also be conjugated with other therapeutic molecules for tumor delivery, such as anti-MHC peptide antibody Fab-immunotoxin conjugates, anti-MHC peptide antibody Fab-cytokine conjugates, and anti-MHC peptide antibody Fab-drug conjugates. IV. Methods of treatment involving immunogenic vaccines or T cells

[0313] Some embodiments provide methods of treatment, including a vaccine, which can be an immunogenic vaccine. In some embodiments, a method of treating a disease (such as cancer) is provided, which can include administering to an individual an effective amount of the composition described herein, a mutant peptide identified using the techniques disclosed herein, a precursor thereof, or nucleic acids encoding the mutant peptide (or precursor) identified using the techniques described herein.

[0314] In some embodiments, a method of treating a disease (e.g., cancer) is provided. The method may include collecting a sample (e.g., a blood sample) from a subject. T cells can be isolated and stimulated. Isolation can be performed using, for example, density gradient sedimentation (e.g., and centrifugation), immunomagnetic selection, and / or antibody-complex filtration. Stimulation can include, for example, antigen-independent stimulation, which can use a mitogen (e.g., PHA or Con A) or anti-CD3 antibodies (e.g., to bind to CD3 and activate the T cell receptor complex) and anti-CD28 antibody (e.g., to bind to CD28 and stimulate T cells).One or more mutant peptides may be (or could be) selected for use in treating a subject (e.g., based on results produced by a machine learning model corresponding to predictions regarding whether and / or to what extent each set of mutant peptides will bind to an MHC molecule of the individual, will be presented by the MHC molecule of the individual, and / or whether an immune response is elicited in the individual, according to one or more techniques disclosed herein). The one or more mutant peptides could be selected based on the technique disclosed herein, which includes identifying and processing one or more presentation sequences associated with the subject (e.g., a presentation: an MHC sequence, a set of variant-encoding sequences, and / or a T-cell receptor sequence).One or more sequences can be detected using the sample from which the T cells were isolated or another sample.

[0315] In some cases, one or more mutant peptides (or their precursors) can be used to generate T cells specific for a mutant peptide (e.g., a neoantigen). For example, peripheral blood T cells can be isolated from a subject and contacted with one or more mutant peptides to generate mutant peptide-specific T cell populations that can be administered to the subject. In some examples, the T cell receptor sequence of the T cells reactive with the mutant peptide can be sequenced. If the sequencing identifies an ordered set of nucleic acids, each codon of the nucleic acids can be translated into an amino acid (e.g., using a search technique). Once the T cell receptor sequence (e.g., the T cell receptor amino acid sequence) is obtained, T cells can be engineered to contain a T cell receptor that specifically recognizes the mutant peptide.These engineered T cells can then be administered to a subject. See, for example, Matsuda et al. "Induction of Neoantigen-Specific Cytotoxic T Cells and Construction of T-cell Receptor Engineered T Cells for Ovarian Cancer," Clin. Cancer Res. 1-11 (2018), which is hereby incorporated by reference in its entirety for all purposes. In any of the methods provided herein, T cells can be expanded in vitro and / or ex vivo prior to administration to a subject. The subject can then be administered (e.g., by infusion) a composition that contains the expanded T cell population.

[0316] In some cases, a method of treating a disease (such as cancer) is provided, which may include administering to an individual a composition that comprises one or more mutant peptides (or one or more precursors thereof) in an amount effective to, for example, prime, activate, and grow T cells in vivo.

[0317] In some embodiments, a method of treating a disease (e.g., cancer) is provided, which may comprise administering to an individual an effective amount of a composition comprising a precursor of a mutant peptide selected using the techniques described herein. In some embodiments, an immunogenic vaccine may comprise a pharmaceutically acceptable mutant peptide selected using the techniques described herein. In some embodiments, an immunogenic vaccine may comprise a pharmaceutically acceptable precursor for a mutant peptide selected using the techniques described herein (such as a protein, peptide, DNA and / or RNA). In some embodiments, a method of treating a disease (e.g., cancer) is provided, which may comprise administering to an individual an effective amount of an antibody that specifically recognizes a mutant peptide selected using the techniques described herein.In some embodiments, a method of treating a disease (e.g., cancer) is provided that may comprise administering to an individual an effective amount of a soluble TCR or TCR analog that specifically recognizes a mutant peptide selected using the techniques described herein.

[0318] In some embodiments, the cancer is one of: carcinoma, lymphoma, blastema, sarcoma, leukemia, squamous cell carcinoma, lung cancer (including small cell lung cancer, non-small cell lung cancer, lung adenocarcinoma, and squamous cell lung cancer), abdominal cancer, hepatocellular cancer, gastric cancer (including gastrointestinal cancer), pancreatic cancer, glioblastoma, cervical cancer, ovarian cancer, liver cancer, bladder cancer, hepatoma, breast cancer, colon cancer, melanoma, endometrial or uterine cancer, salivary gland carcinoma, kidney cancer, liver cancer, prostate cancer, vulvar cancer, thyroid cancer, liver carcinoma, head and neck cancer, colorectal cancer, rectal cancer, soft tissue sarcoma, Kaposi's sarcoma, B-cell lymphoma (including low-grade / follicular non-Hodgkin's lymphoma (NHL), small lymphocyte (SL) NHL, intermediate-grade / follicular NHL, intermediate-grade diffuse NHL,high-grade immunoblastic NHL, high-grade lymphoblastic NHL, high-grade small noncleaved cell NHL, bulky NHL, mantle cell lymphoma, AIDS-related lymphoma, and Waldenstrom's macroglobulinemia), chronic lymphocytic leukemia (CLL), acute lymphoblastic leukemia (ALL), myeloma, hairy cell leukemia, chronic myeloid leukemia, and post-transplant lymphoproliferative disorder (PTLD), as well as atypical vascular proliferation associated with phakomatoses, edemas (e.g., associated with brain tumors), and Meigs' syndrome.

[0319] Embodiments disclosed herein may include identifying part or all of and / or implementing part or all of a personalized medical strategy. For example, one or more mutant peptides can be selected for use in a vaccine by: determining an MHC sequence and / or a set of variant-encoding sequences using a sample from an individual; and processing the representations of the MHC sequence and variant-encoding sequences using a machine learning model disclosed herein (e.g., an attention-based machine learning model). One or more mutant peptides (and / or their precursors) can then be administered to the same individual.

[0320] In some embodiments, a method of treating a disease (such as cancer) in an individual is provided, which comprises: a) identifying one or more mutant peptides in the individual (e.g., based on results generated by a machine learning model corresponding to predictions regarding whether and / or to what extent each of a set of mutant peptides will bind to an MHC molecule of the individual, will be presented by an MHC molecule of the individual, and / or elicit an immune response in the individual, according to one or more techniques disclosed herein); b) synthesizing the identified mutant peptide(s) or one or more mutant peptide precursors(s) or nucleic acid(s) (e.g., a polynucleotide such as DNA or RNA) encoding the identified peptide(s) or peptide precursor(s); and c) administering the mutant peptide(s), mutant peptide precursor(s), or nucleic acid(s) to the individual.

[0321] In some embodiments, a method of treating a disease (such as cancer) in an individual is provided, which comprises: a) identifying one or more mutant peptides in the individual (e.g., based on results obtained by a machine learning model corresponding to predictions regarding whether and / or to what extent each of a set of mutant peptides will bind to an MHC molecule of the individual, will be presented by an MHC molecule of the individual, and / or elicit an immune response in the individual, according to one or more techniques disclosed herein); b) identifying a set of nucleic acids (e.g., polynucleotides such as DNA or RNA) that encode the identified mutant peptide(s) or one or more precursors of the mutant peptide(s); c) synthesizing the set of nucleic acids; and d) administering the set of nucleic acids to the individual.

[0322] In some embodiments, a method of treating a disease (such as cancer) in an individual is provided, which comprises: a) identifying one or more mutant peptides in the individual (e.g., based on results obtained by a machine learning model corresponding to predictions regarding whether and / or to what extent each of a set of mutant peptides will bind to an MHC molecule of the individual, will be presented by an MHC molecule of the individual, and / or elicit an immune response in the individual, according to one or more techniques disclosed herein); b) producing an antibody that specifically recognizes the mutant peptide; and c) administering the peptide to the individual.

[0323] The methods provided herein can be used to treat an individual (e.g., a human) who has been diagnosed with or is suspected of having cancer. In some embodiments, the individual may be a human. In some embodiments, the individual may be at least about any of 18, 20, 25, 30, 35, 40, 45, 50, 55, 60, 65, 70, 75, 80, or 85 years of age. In some embodiments, the individual may be male. In some embodiments, the individual may be female. In some embodiments, the individual may have declined surgery. In some embodiments, the individual may be medically inoperable. In some embodiments, the individual may be at clinical stage Ta, Tis, T1, T2, T3a, T3b, or T4. In some embodiments, the cancer may be recurrent.In some embodiments, the individual may be a person who exhibits one or more symptoms associated with cancer. In some embodiments, the individual may be genetically or otherwise predisposed (e.g., have a risk factor) to developing cancer.

[0324] The methods provided herein can be performed in an adjuvant setting. In some embodiments, the method is practiced in a neoadjuvant setting, i.e., the method can be performed before primary / definitive therapy. In some embodiments, the method is used to treat a subject who has previously been treated. Any of the treatment methods provided herein can be used to treat a subject who has not previously been treated. In some embodiments, the method is used as first-line therapy. In some embodiments, the method is used as second-line therapy.

[0325] In some embodiments, a method is provided for reducing the incidence or burden of pre-existing cancer with metastatic tumors (e.g., pulmonary metastasis or lymph node metastasis) in an individual, comprising administering to the individual an effective amount of the composition disclosed herein. In some embodiments, a method is provided for prolonging the time of disease progression of cancer in an individual, comprising administering to the individual an effective amount of the composition disclosed herein. In some embodiments, a method is provided for prolonging the survival of an individual with cancer, comprising administering to the individual an effective amount of the composition disclosed herein.

[0326] In some embodiments, at least one or more chemotherapeutic agents can be administered in addition to the composition disclosed herein. In some embodiments, the one or more chemotherapeutic agents can (but need not) belong to different classes of chemotherapeutic agents.

[0327] In some embodiments, there is provided a method of treating a disease (such as cancer) in an individual, comprising administering: a) a vaccine disclosed herein (e.g., which comprises a mutant peptide selected based on the machine learning technique disclosed herein, or a precursor thereof), and b) an immunomodulator. In some embodiments, there is provided a method of treating a disease (such as cancer) in an individual, comprising administering: a) a vaccine disclosed herein (e.g., which comprises a mutant peptide selected based on the machine learning technique disclosed herein, or a precursor thereof), and b) a checkpoint protein antagonist.In some embodiments, there is provided a method of treating a disease (such as cancer) in an individual, comprising administering: a) a vaccine disclosed herein (e.g., comprising a mutant peptide selected based on the machine learning technique disclosed herein, or a precursor thereof), and b) a programmed cell death 1 (PD-1) antagonist, such as an anti-PD-1 antibody. In some embodiments, there is provided a method of treating a disease (such as cancer) in an individual, comprising administering: a) a vaccine disclosed herein (e.g., comprising a mutant peptide selected based on the machine learning technique disclosed herein, or a precursor thereof), and b) a programmed cell death ligand 1 (PD-L1) antagonist, such as an anti-PD-L1 antibody.In some embodiments, a method of treating a disease (such as cancer) in an individual is provided, comprising administering: a) a vaccine disclosed herein (e.g., that comprises a mutant peptide selected based on the machine learning technique disclosed herein, or a precursor thereof), and b) a cytotoxic T-lymphocyte-associated protein 4 (CTLA-4) antagonist, such as an anti-CTLA-4 antibody.

[0328] It will be appreciated that various disclosures relate to the use of amino acid sequences. Nucleic acid sequences may be used additionally or alternatively. For example, a disease-specific sample may be sequenced to identify a set of nucleic acid sequences that are not present in a corresponding disease-specific non-disease-related sample (e.g., from the same subject or different subjects). Similarly, the nucleic acid sequence of an MHC molecule and / or T-cell receptor may also be identified. Representations of each of the disease-specific nucleic acid sequence and the MHC molecule (or T-cell receptor) may be processed by an attention-based model as described herein (e.g., and potentially were trained using the nucleic acid sequence representations). V. Examples V.A Overview

[0329] A typical attention-based machine learning model for peptide-MHC (MHC class I) (hereinafter "P-MHC-I model") and a typical attention-based machine learning model for peptide-MHC (MHC class II) (hereinafter "P-MHC-II model") (jointly and individually referred to herein as the P-MHC model) were developed. These models are implementation examples for the machine learning model 132 in Fig. 1. Both the architecture of the P-MHC-I model and the architecture of the P-MHC-II model were implemented according to the architecture shown in Fig. 3 and Fig. 4A.

[0330] The P-MHC model is a typical attention-based deep learning model for neoantigen presentation prediction in personalized cancer vaccine development. The P-MHC model takes the N-flanking sequence, peptide sequence, and MHC sequence (MHC pseudo-sequence) as inputs and outputs the presentation or eluted ligand (EL) score. A dictionary covering the space of naturally occurring amino acids was constructed and tokenized to represent amino acid sequences. The input amino acid sequences were tokenized as characters, each represented by a unique character. The model pairs the input N-flanking sequence and peptide sequence with one of 6 MHC alleles and 6 pairwise interactions were forwarded to the P-MHC-I model, and one of 12 MHC allotypes and 12 pairwise interactions were forwarded to the P-MHC-II model to select a specific binding MHC allele.

[0331] Thus, the R-MHC model internally performs a data deconvolution for multiple alleles. The most probable peptide-MHC interaction result is normalized as a value between 0 and 1 and compared to the true representation value using a binary cross-entropy loss function to generate an error term for tuning the model parameters. To prevent overfitting and improve model robustness, the R-MHC model uses systematic methods in model training.

[0332] Representative results and statistics corresponding to the training and performance of the P-MHC-I model and the P-MHC-II model compared to other previously available models (e.g., NetMHCpan-4.0 (herein "Model A"), Immune Epitope Database and Analysis Source (IEDB) v2.13 (herein "Model B") for the P-MHC-I model, and NetMHCIIpan-4.1 (herein "Model C") for the P-MHC-II model. The P-MHC-I and P-MHC-II models consistently performed better than other models for peptide representation, and the P-MHC-I model performed better than other models for CD8 T cell response prediction. The P-MHC model performs better at least because it inversely convolves peptide-MHC pairs from multi-allelic data, can be easily trained on extended training data in both monoallelic and multi-allelic formats.V.B. Materials and methods.V.B.1. Training the R-MHC model - Immunopeptidomic data

[0333] Peptide elution data from mass spectrometry experiments was used to construct an immunopeptidomic dataset for training the R-MHC model. This data includes a mixture of private and public datasets that include multiallelic and monoallelic peptide elution data from cell lines, tissue samples, and PBMC donors. V.B.1.a. Representation labeling data

[0334] Positive set (EL=1). For each batch, positive peptide-MHC pairs (e.g., peptide-HLA) were processed as follows: 1) Peptides were aligned with the human proteome. 2) For each peptide, flanking sequences of up to 10 amino acids were retained at the N-terminus and C-terminus positions. 3) Peptides that mapped to multiple genes were removed from the following analysis. Such peptides are not shown in the EL=1 sets. (No such restriction was imposed on EL=0, since EL=0 peptides were generated only from proteins that had evidence of EL=1 peptides.) 48,329 class I peptides were filtered by this criterion. Although this is a large number, it increases the confidence in the negative set. 4) Peptides that mapped to the same gene but with different flanking sequences were also removed from the following analysis. This further removes 11,443 class I peptides.5) Peptides containing post-translational modifications (PTMs) were also removed from the following analysis. 7080 such class I peptides were removed.

[0335] Negative set (EL=0). Negative peptide-MHC pairs (e.g., peptide-HLA) were computationally generated. For each allele for each protein of origin in the positive set (EL=1), all possible peptide fragments of length 8-11 were generated for MHC class I and 8-30 for MHC class II, with uniform probability for each length. N-terminal and C-terminal flanking sequences were also retained, with a maximum length of 10 amino acids. All peptide-genotype pairs shown in the EL=1 data were removed from the EL=0 data. Additionally, for databases constructed for MHC class II, peptide-genotype pairs with any length 9 subsequence that can be found in the EL=1 peptide (paired with the same genotype) were removed. V.B.1.b. Control data set

[0336] The validation dataset was created by splitting the above EL data discussed in Section VI.E.1.a into training, validation, and test sets. The training and validation sets were used to train the R-MHC model, while the test set was not explicitly used for training and was used only for model performance. For class I MHC data, monoallelic data was used to generate the test dataset by retaining 10% of the peptides from the monoallelic data for each allele. For class II MHC data, all data, multiallelic and monoallelic, were used to generate the test / validation datasets.

[0337] Dataset features include: All peptide lengths were restricted to the range of [8,14] amino acids for class I and [8,30] amino acids for class II. All peptides were restricted to contain canonical amino acids in the core sequence (i.e., epitope) and flanking sequences. All model names were replaced by 34 amino acid subsequences defined by the following amino acid positions in the MHC1 protein: (7, 9, 24, 45, 59, 62, 63, 66, 67, 69, 70, 73, 74, 76, 77, 80, 81, 84, 95, 97, 99, 114, 116, 118, 143, 147, 150, 152, 156, 158, 159, 163, 167, 171), or positions in the MHCII alpha and beta proteins:alpha: 9,11,22,24,31,52,53,58,59,61,65,66,68,72,73; Ibeta: 9,11,13,26,28,30,47,57,67,70,71,74,77,78,81,85,86,89,90. These positions were previously described as positions in the binding pocket where the MHC-I / II protein contacts the peptide. The set of unique subsequences for a data point can henceforth be referred to as a "pseudogenotype."In some cases, the names of multiple alleles may indicate the same 34 amino acid subsequences. These alleles were considered identical for training the attention-based R-MHC model. All empty flanking sequences (peptide maps for the protein end) were assigned a special amino acid symbol, "$." Six data points where the flanking sequences read "NA" in the amino acid alphabet were removed due to the fact that in some programming languages, the interpretation of NA is "not applicable."

[0338] The training / validation / test splits were performed as follows:

[0339] For EL=1: For each processing batch (each batch was based on the original source from the dataset), the monoallelic data were randomly split into training / validation / testing groups with a 70 / 20 / 10 ratio. For the MHCII class, no length 9 subsequence from the peptide sequence overlapped between the training / validation / testing datasets for peptides with exact genotype matches. The monoallelic data consist of 105 (41) unique subsequences, representing 111 (39) unique MHCI class (MHCII class), respectively, alleles in the entire dataset. All multiallelic data were used entirely for training the class I datasets. The multiallelic data consist of 126 (76) unique MHCI class (MHCII class) genotypes in the entire dataset.Data in processing batches were pooled and duplicated {peptide, nFlank, cFlank, mhc0, mhc1, mhc2, mhc3, mhc4, mhc5} (MHC class I) and {peptide, nFlank, cFlank, mhc_dq1_1, mhc_dq1_2, mhc_dq1_3, mhc_dq1_4, mhc_dp1_1, mhc_dp1_2, mhc_dp1_3, mhc_dp1_4, mhc_dr1_1, mhc_dr1_2, mhc_dr3_1, mhc_dr3_2, mhc_dr4_1, mhc_dr4_2, mhc_dr5_1, mhc_dr5_2} (MHC class II) entry For EL=0: For each treatment batch, for each {peptide, pseudogenotype} pair, negative peptide data were selected at a 1:1 ratio with EL=1 data in the training and validation groups. In the test group, they were selected at a 1:99 ratio for MHC class I and a 1:9 ratio for MHC class II. Data across treatment batches were pooled and duplicate observations were removed. This finally resulted in 1.71% of test data being positive (instead of 1%) for MHCI class and 11.15% for MHC class II (instead of 10%).

[0341] For observations with multiple subsequences in the "pseudogenotype", i.e., multi-allelic data, negative peptides were generated by excluding positive peptides for each allele, and then random peptides were selected from protein sources.V.B.1.c. Baseline quality control

[0342] The following downstream quality control procedure was followed to ensure that there was no data redundancy: 1) Only canonical amino acids are allowed in the peptide, N-flank, and C-flank sequences; 2) Each set of {N-flank, peptide, C-flank, pseudogenotype} entries is unique; and 3) there is no overlap of {N-flank, peptide, C-flank, pseudogenotype} entries in the set EL=1 and EL=0. For MHC class II, there is also no overlap between the length of 9 subsequences in the peptide sequence between EL=1 and EL=0, for peptides with identical pseudogenotypes.

[0343] The number of MHC (HLA) pseudogenotypes may differ from the number of alleles, since some alleles with different allele names (at 2-field resolution, i.e., 4-bit resolution) may have the same pseudogenotype.V.B.2. Immunogenicity datasets for evaluating the performance of the R-MHC-I model

[0344] Two different datasets were used to evaluate the performance of the P-MHC-I model. For the first dataset, subjects with cancer had their DNA sequences, from which standard P-MHC binding was performed, predictions were made using IEDPv.2.13BA to predict neoantigens that were expressed and / or bound in MHC. The neoantigens predicted in this way were further prioritized using their expression, allelic frequency of variants, and clonal properties in tumor tissue. Subjects were then dosed with the RNA vaccine as previously administered. T-cell responses to neoantigens delivered in the RNA vaccine were monitored in dosed subjects using multimer and ELISPOT assays. T-cell responses, considered technical artifacts, were removed using multiple controls in these assays.In the second immunogenicity test dataset, sequence determination data were obtained from cancer subjects receiving checkpoint blockade therapy (but not RNA vaccine therapy) identified by the Tumor Selection Alliance Agreement (TESLA). P-MHC binding prediction was performed using NetMHCcons 1.0 to predict neoantigens that were expressed and / or bound by the MHC. Immunogenicity assays were performed on neoantigens predicted by the P-MHC-I model and used to evaluate the performance of the P-MHC-IVB.2.a. Multivariate analysis of the dosed subject.

[0345] For the first immunogenicity test data set, multimer assay data were scored as positive or negative for CD8 T cell detection by peptide-MHC multimers. Conservative criteria were used to define a positive result: specifically, whether the proportion of double tetramer-positive CD8 T cells was greater than 0.05%. Some neoepitopes were declared positive despite the presence of less than 0.05% neoepitope-specific CD8 T cells if closer examination of the T cell phenotype strongly suggested the presence of a T cell response. Of the multimer assay data, 1,318 neoepitopes were scored as negative, and based on conservative criteria, a small proportion of these are expected to be false negatives. 27 neoepitope-HLA pairs were reported as positive only after vaccination (called de novo responses) and 20 pairs were reported as pre-existing CD8 T cell responses.V.B.2.b.ELISpot assay of the dosed subject.

[0346] In addition to the first set of immunogenicity test data, ELISpot data were collected. Spot counts of negative controls without peptide re-stimulation and test cases with peptide re-stimulation were statistically evaluated to indicate positive responses (using a permutation approach) and further confirmed manually to assign a positive or negative result for neoantigen immunogenicity for a given subject visit. A neoantigen was reported as positive in the ELISpot assay if it showed a positive result at any of the subject visits, either before treatment or after treatment. Neoantigens were further filtered based on the following criteria: (a) the assessed assay result value was not "NA"; (b) none of the assessed P-MHC-I assay methods (P-MHC-I, model A, model B) reported an "NA" value for the neoantigen; and (c) the collected neoantigens were used for restimulation, which was excluded from consideration.

[0347] After all filtering steps, the distribution of positive (immunogenic) and negative (non-immunogenic) neoantigens for each cell type assessed in the ELISpot assays is shown below. Assay.value binary=TRUE indicates an immunogenic neoantigen, while non-immunogenic results are marked as Assay.value_binary=FALSE. Positive assays were further classified into two sets based on spot counts from the ELISpot assay. Each ELISpot assay had duplicate experiments, and the average spot count was determined in duplicates. For a positive neoantigen, the highest average spot count across all visits was considered, and positive neoantigens were divided into two sets: one with this spot count value <50 and the other with this spot count value >50. The latter set represented neoantigens that induced more extensive T-cell responses and were less likely to contain false-positive ELISpot result interpretations compared to the set with a lower spot count number. The choice of 50 spots was arbitrary, as it was correspondingly higher than the original cutoff value used for positive ELISpot values ​​(spot count >15). V.B.2.c. Multimeric TESLA analysis

[0348] For the second immunogenicity test data set, the TESLA agreement validated neoantigen predictions. Assay data were available for subjects 1, 2, 3, 4, 10, 12, and 16 from the TESLA subject identifiers. Assay results were provided by TESLA based on four different assays: TCR_FLOW_I, TCR_FLOW_II, the nanoparticle assay, and the TCR reactivity assay. The TCR_FLOW_I assay results were used in this example. The other assays were omitted due to the following reasons: (a) the nanoparticle assay was expected to have a higher false-positive rate, as it is a single-cell assay designed to be very sensitive; (b) TCR_FLOW_II is highly redundant with TCR_FLOW_I, with both being performed in different laboratories and TCR_FLOW_II having fewer data points.The TCR reactivity assay is an intracellular assay of IFNγ / TNFα staining after prestimulation of T cells with IL-2 and short peptides for 7 days, followed by re-stimulation with the short peptide. The TESLA team has not validated the use of this assay to predict peptide-MHC presentation. The selected assay had 16 positive results and 196 negative results. V.B.3. Comparison models - NetMHCpan and IEDB metrics.

[0349] To perform the comparison against the P-MHC-I model, Model A and Model B were used to assign BA and EL values ​​to peptide-HLA pairs. To perform the comparison against the P-MHC-II model, Model C was used to assign EL values ​​to peptide-MHC (HLA) pairs. The BA and EL values ​​obtained as percentiles by these methods are referred to (in this example) as BA or EL. These percentile values ​​behave such that a lower value implies higher affinity or probability of presentation. A transformed score scheme was used by inverting these values ​​to obtain scores (e.g., for MHC-I, the binding affinity score for Model A, the elution score for Model A, and the binding affinity score for Model B; for MHC-II, the binding affinity score for Model C) that behave such that a higher value indicates stronger affinity or probability of presentation. One such score was obtained for neoepitope-HLA pairs.For neoantigens, all neoepitope-HLA pairs were considered for 8-14 measures of candidate neoepitopes containing the mutation, and the pair with the highest score was selected to represent the neoantigen score.VB ResultsV.B.1. Performance of the P-MHC-I model on the representation data.

[0350] Fig. 14A-B are graphs with exemplary precision response (PR) curves according to one or more embodiments. Fig. 14A-B show the performance of the P-MHC-I model compared to previously used approaches. The eluted ligand (EL) test data set was used to evaluate the performance of predicting representations between the EL result of the P-MHC-I model, the EL result of model A, and the binding affinity (BA) result of model C.

[0351] Fig. 14A contains a graph 1400 showing the performance of the P-MHC-I model. Fig. 14B contains a graph 1402 showing the performance of model A relative to its elution result. Fig. 14C contains a graph 1404 showing the performance of model B relative to its binding affinity result. The point on the curve of each of graphs 1400, 1402, and 1404 corresponds to the performance cutoff value for the upper 1.71% quantile of the performance (chosen due to 1.71% of the gold standard test data being positive). The average precision (AP) is indicative of the threshold-independent performance. The F1 score, accuracy, and recall values ​​are based on the 1.71% cutoff value.

[0352] The values ​​of Model A and Model B were the percentile rank results of these methods. The values ​​of the P-MHC-I model were selected from the result (terminal node) of the P-MHC-I model. Based on these PR curves, the results in Fig. 14A-B indicate that the P-MHC-I model showed improved performance relative to both Model A and Model C (CT value of 0.85 versus 0.78 for Model A and 0.57 for Model B). The CT values ​​of the methods were compared on an allele-by-allele basis.

[0353] Fig. 15 is a graph 1500 comparing exemplary average accuracy values ​​of the ligand elution results of model A and the P-MHC-I model for each allele in a test data set according to one or more embodiments. The test data set, which was monoallelic, included at least 1000 data points, with 67 alleles meeting the criteria. As shown in graph 1500, the P-MHC-I model showed better performance relative to model A. The marker patterns in graph 1500 indicate whether the alleles were from the HLA-A, B, or C gene. The marker sizes represent the amount of monoallelic data used in training the P-MHC-I model for this allele, which also correspond to the amount of test data for each allele.

[0354] Fig. 16A and 16B are graphs 1600 and graph 1602, respectively, that compare the performance of the P-MHC-I model on a human data set with the performance of the P-MHC-I model on a mouse data set according to one or more embodiments. As shown by these graphs, the P-MHC-I model performed well for both data sets, with the average accuracy of the P-MHC-I model being similar in both the human and mouse data sets. These results indicate that the P-MHC-I model can be a pan-species model that can be used with desirable performance characteristics for a variety of particles. VB2. Performance of the P-MHC-III model on representation data

[0355] Fig. 17A and 17B are a graph 1700 and a graph 1702, respectively, that compare the performance of the R-MHC-II model with the C model on the representation data according to one or more embodiments. The values ​​of the C model were the percentile rank results. The values ​​of the R-MHC-II model were taken from the result (terminal node) of the R-MHC-II model. Using the average accuracy from the PR curves, the results in Fig. 17A and 17B show that the R-MHC-II model with a CT of 0.69 showed improved performance relative to the C model with a CT of 0.31. The CT values ​​of these two methods were compared on a relative allele basis.

[0356] Fig. 18A and 18B are a graph 1800 and a graph 1802, respectively, which compare the performance of the R-MHC-II model with the C model, respectively, on an evasion data set according to one or more embodiments. Again, the R-MHC-II model with an AP of 0.84 shows improved performance relative to the C model with a CT of 0.46.

[0357] Fig. 19 is a graph 1900 showing a genotype comparison of the average accuracy for the P-MHC-II model with the C model on a test data set according to one or more embodiments. Based on the genotype, the P-MHC-II model had improved performance relative to the C.V.B.3 model. Performance relative to the first and second test data sets regarding immunogenicity

[0358] The first and second immunogenicity test data sets were used to evaluate the performance of R-MHC presentation predictions on T cell response data. In these evaluations, no training was performed on the immunogenicity data, and only the amino acid sequence of neoantigens and MHC proteins were used to calculate R-MHC presentation scores. Other features, such as gene or mutant allele expression, were not used to provide a simplified estimate of the contribution of R-MHC presentation prediction to CD8 T cell response prediction. V.B.3.a. Multivariate analysis of the dosed subject

[0359] Fig. 20 is a graph of 2000 receiver operating characteristic (ROC) curves that shows the performance of the P-MHC-I model (EL result), model A (EL result), and model B (BA result) relative to the CD8 multimer assay data (the first set of test immunogenicity data) according to one or more embodiments. The performance was assessed relative to the ability to predict positive neoepitopes from the multimer assay. For model A and model B, the values ​​were inversely transformed to obtain EL and BA scores, respectively, so that a higher value indicated a stronger binding affinity or presentation probability. The area under the curve (AUC) was calculated based on a step function. The step function for plotting the ROC curve connected the points representing the true positive rates (tpr) and false positive rates (fpr) in a horizontal and then vertical direction.True positive rate (tpr) and false positive rate (fpr) values ​​were calculated using the R package ROCR.V.B.3.b. Dosed subject ELISpot assay.

[0360] Fig. 21A-D are graphs 2102, 2104, 2106 and 2108, respectively, which show the performance of the P-MHC-I model (E1 result), model A (EL result) and model B (BA result) relative to ELISpot assays (first set of test immunogenicity data) according to one or more embodiments. As shown, the P-MHC-I model performed well with strong predictive power. The graphs show typical ROC curves with separate subgraphs shown for the PBMC ELISpots (Fig. 21A, PBMC panel) and the CD8 ELISpot (Fig. 21B, CD8 panel). Positive CD8 ELISpot data were further divided into two sets, and ROC curves were generated for stronger T-cell responses (Fig. 21B, CD8 spots >= 50) and relatively weak T-cell responses (Fig. 21D, CD8 spots <50). The same negative neoantigen set was used to generate ROC curves for these two sets. V.B.3.c. Multivariate TESLA analysis

[0361] Fig. 22A-D are graphs 2202, 2204, 2206, and 2208, respectively, that show the performance of model A (BA result), model A (EL result), model C (BA result), and model P-MHC-I (EL result), respectively, according to one or more embodiments. The performance was assessed on TESLA immunogenicity data (the second set of immunogenicity test data), with the results of the multivariate analyses being used. These graphs are a scatter plot corresponding to typical neoepitope-HLA pairs assessed by the multivariate analyses from the TESLA study. The answer is TRUE for positive assay responses, as determined by TESLA, and FALSE for non-immunogenic neoepitopes. The Wilcoxon rank sum test was used to calculate p-values ​​for the two-sided alternative hypothesis. The y-axes show the converted values ​​such that a higher value corresponds to stronger peptide-MHC binding or presentation.

[0362] Fig. 23 is an illustration of a graph 2300 comparing ROC curves for model A (EL result), model B (BA result), and model P-MHC-I (EL result) using TESLA multimer assay data according to one or more embodiments. The multimer assay was the TCR FLOW I assay. The area under the curve was largest for model P-MHC-IVG. Conclusion

[0363] Thus, the methods for predicting R-MHC presentation were evaluated on two types of evaluation datasets: R-MHC presentation data from immunopeptidomic experiments and T-cell response data from various immunogenicity assays. Presentation predictors trained on immunopeptidomic data performed better than the current production method (IEDBv2.13BA result) on a variety of these datasets. The R-MHC model showed improved performance values ​​across many datasets. Therefore, the use of attention-based techniques trained on immunopeptidomic data may be superior to models based on in vitro binding affinity data. VI. Computer-implemented system

[0364] Fig. 24 is a block diagram of a computing system according to various embodiments. Computing system 2400 may be an example of one implementation for the computing platform 102 described above in Fig. 1.

[0365] In one or more examples, the computing system 2400 may comprise a bus 2402 or other communication mechanism for transmitting information, and a processor 2404 connected to the bus 2402 for processing the information. In various embodiments, the computing system 2400 may also include memory, which may be random access memory (RAM) 2406 or another dynamic memory device connected to the bus 2402 for determining instructions to be executed by the processor 2404. The memory may also be used to store temporary variables or other intermediate information when executing instructions to be executed by the processor 2404. In various embodiments, the computing system 2400 may also include read-only memory (ROM) 2408 or another static memory device connected to the bus 2402 for storing static information and instructions for the processor 2404.A storage device 2410, such as a magnetic disk or an optical disk, may be provided and connected to the bus 2402 for storing information and instructions.

[0366] In various embodiments, the computing system 2400 may be connected via a bus 2402 to a display 2412, such as a cathode ray tube (CRT) or a liquid crystal display (LCD), for displaying information on a user's computer. An input device 2414, including alphanumeric and other keys, can be connected to the bus 2402 for transmitting information and selecting commands for the processor 2404. Another type of user input device is a cursor control 2416, such as a mouse, a joystick, a trackball, a gesture input device, a gaze-based input device, or cursor direction keys for transmitting control information and selecting commands for the processor 2404 and for controlling the movement of the cursor on the display 2412. This input device 2414 typically has two degrees of freedom on two axes, a first axis (e.g., x) and a second axis (e.g., y), which provide for determining the positions of the device on a plane.However, it should be understood that input devices 2414 that provide three-dimensional (e.g., x, y, and z) cursor movement are also contemplated herein.

[0367] According to some embodiments of the teachings of the present invention, results can be provided by the computing system 2400 in response to the processor 2404 executing one or more sequences of one or more instructions contained in the RAM 2406. Such instructions can be read into the RAM 2406 from another computer-readable medium or a computer-readable storage medium, such as the storage device 2410. Execution of the sequences of instructions contained in the RAM 2406 can cause the processor 2404 to perform the processes described herein. Alternatively, a hardware circuit can be used instead of or in combination with software instructions to implement the teachings of the present invention. Thus, embodiments of the teachings of the present invention are not limited to any particular combination of a hardware circuit and software.

[0368] The term "machine-readable medium" (e.g., data array, data storage, storage device, data storage device, etc.) or "machine-readable storage medium" when used herein refers to any medium that is involved in providing instructions to the processor 2404 for execution. Such a medium may have many forms, including, but not limited to, non-volatile media, volatile media, and communication media. Examples of non-volatile media may include, but are not limited to, optical, solid-state, magnetic disks, such as storage device 2410. Examples of volatile media may include, but are not limited to, dynamic memory, such as RAM 2406. Examples of communication media may include, but are not limited to, coaxial cables, copper wire, and fiber optic cables, including cables that contain bus 2402.

[0369] General forms of computer-readable storage medium include, for example, a floppy disk, a floppy disk, a hard disk, magnetic tape or any other magnetic storage medium, a CD-ROM, any other optical storage medium, punched cards, paper tape and other physical storage medium with a pattern of holes, RAM, PROM and EPROM, FLASH-EPROM, any other memory chip or cartridge, or any other tangible storage medium that can be read by a computer.

[0370] In addition to a computer-readable medium, instructions or data can be provided in the form of signals on communication means included in a communication device or system to provide sequences of one or more instructions to the processor 2404 of the computing system 2400 for execution. For example, a communication device may include a transceiver with signals indicative of instructions and data. The instructions and data are configured to cause one or more processors to implement the functions specified in the disclosure of the present invention. Typical examples of data transmission connections may include, but are not limited to, communication via a telephone modem, wide area networks (WAN), local area networks (LAN), communication via infrared sensors, NFC communication, optical communication, etc.

[0371] It should be appreciated that the techniques described herein, the flow charts, diagrams, and accompanying disclosure may be implemented using the computing system 2400 as a single device or a distributed network of shared computer processing resources, such as a cloud computing network.

[0372] The techniques described in this document can be implemented by various means depending on the application. For example, these techniques can be implemented in hardware, firmware, software, or any combination thereof. For a hardware implementation, the processing unit can be implemented as one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), processors, controllers, microcontrollers, microprocessors, electronic devices, other electronic units designed to perform the functions described in this document, or a combination thereof.

[0373] In various embodiments, the methods of the teachings of the present invention may be implemented as firmware and / or software and applications written in conventional programming languages ​​such as C, C++, Python, etc. If implemented as firmware and / or software, the embodiments described herein may be implemented on a transient computer-readable medium in which a program is stored to cause a computer to execute the methods described above.It should be understood that the various mechanisms described herein may be provided in a computing system, such as computing system 2400, with processor 2404 performing the analyses and determinations represented by these mechanisms, being subject to instructions represented by any or a combination of memory components RAM 2406, ROM, 2408, or storage device 2410, and user input represented by input device 2414.VII. Illustrative Descriptions of Terms.

[0374] As used herein, the terms "peptide," "polypeptide," and "protein" are used interchangeably to refer to a polymer of amino acid residues. The term encompasses amino acid chains of any length, including full-length proteins with amino acid residues linked by covalent peptide bonds.

[0375] As used herein, a "mutant peptide" may refer to a peptide that is not present in normal tissue (e.g., in the wild-type amino acid sequence of normal tissue) of an individual subject. A mutant peptide contains at least one mutant amino acid and may be present in disease-related tissue (e.g., collected from a specific subject) and not in normal tissue (e.g., collected from a specific subject, collected from another subject, and / or as determined in a database corresponding to normal tissue). The mutant peptide may contain an epitope. An epitope is a portion of a mutant peptide to which an MHC molecule or T cell receptor (TCR) binds. Thus, this binding between the epitope of the mutant peptide and the MHC molecule or TCR can elicit an immune response (as a result of the mutant peptide not binding to the subject "itself"). The mutant peptide may contain or be a neoantigen.A mutant peptide may arise from, for example: a nonsynonymous mutation resulting in different amino acids in the protein (e.g., a point mutation); a read-through mutation in which the stop codon is modified or deleted, resulting in the translation of a longer protein with a new tumor-specific sequence at the C-terminus; a splice site mutation that results in a unique tumor-specific protein sequence; a chromosomal rearrangement that results in a chimeric protein with a tumor-specific sequence at the junction of two proteins (i.e., gene fusion); and / or a frameshift insertion or deletion that results in a new open reading frame with a tumor-specific protein sequence. A mutant peptide may comprise a polypeptide (as characterized by a polypeptide sequence) and / or may be encoded by a nucleotide sequence.

[0376] As used herein, the "C-flank" of a peptide refers to one or more amino acids upstream of the C-terminus of the peptide from the parent protein. Optionally, the C-flank of a peptide includes one, two, three, four, five, or more amino acid residues upstream of the C-terminus of the peptide.

[0377] As used herein, the "N-flank" of a peptide refers to one or more amino acids downstream of the N-terminus of the peptide, from the parent peptide. Optionally, the N-flank of a peptide includes one, two, three, four, five, or more amino acid residues downstream of the N-terminus of the peptide.

[0378] As used herein, the "epitope" of a peptide may refer to the region of the peptide between the C-flank and N-flank that can be recognized by the TCR. A peptide epitope is the portion of the peptide that is recognized by the TCR on a T cell and the MHC class I on an antigen-presenting cell. For example, an epitope may be a peptide that the TCR binds to, such as the peptide that the TCR binds to when the peptide binds to the MHC class I on an antigen-presenting cell.

[0379] As used herein, "ligand" is a peptide that is found to be presented by an MHC molecule on the cell surface from elution experiments or that is found to bind to MHC in an in vitro assay.

[0380] As used herein, "sequence" refers to an amino acid sequence that contains an ordered set of amino acid identifiers.

[0381] As used herein, "peptide sequence" refers to a sequence that identifies the amino acids of at least a portion of a peptide. In some cases, the peptide sequence contains a variant-encoding sequence that contains a variant that is not observed in the corresponding reference sequence.

[0382] When a peptide includes a mutant peptide, the variant-encoding sequence identifies the amino acids of the mutation or variant. However, when the peptide does not contain a mutation or variant, the variant-encoding sequence does not identify the amino acids of the mutation or variant (and in this case is the same as the reference sequence). The variant-encoding sequence can be determined by collecting a disease sample and / or a tumor sample (e.g., containing tumor cells) and performing a sequence determination analysis to identify one or more sequences corresponding to the disease and / or tumor cells in the sample. In some cases, the results of the sequence determination analysis of the amino acid sequence.In some cases, the results of a nucleic acid sequence analysis can be further processed to convert codons into amino acid identifiers, thereby obtaining an amino acid sequence. The variant coding sequence may comprise a neoantigen sequence. The variant coding sequence may, but need not, comprise one or more termini (e.g., the C-terminus and / or N-terminus) of a peptide. The variant coding sequence may comprise a peptide epitope. The variant coding sequence may identify amino acids in a peptide that have one or more variants (e.g., one or more amino acid differences) relative to a corresponding reference sequence. In some cases, the variant coding sequence comprises an ordered set of amino acids.In some cases, the variant coding sequence identifies a reference peptide (e.g., by identifying a genetic reference sequence, such as by gene, start position, and / or end position; or by gene, start position, and / or length) and one or more point mutations relative to the reference peptide.

[0383] As used herein, a "reference sequence" may refer to a sequence that identifies amino acids in at least a portion of a non-mutated or wild-type peptide (e.g., a wild-type parent sequence). The non-mutated or wild-type peptide may contain no variants or fewer variants than those included in the mutant peptide. The reference sequence may comprise an amino acid sequence encoded by a genetic sequence in the same gene relative to the gene that contains the corresponding variant-coding sequence. The reference sequence may comprise an amino acid sequence encoded by a genetic sequence encompassing the same start and stop in the gene relative to the intragenic positions associated with the genetic sequence associated with the corresponding variant-coding sequence.A reference sequence may be identified by collecting a non-disease and / or non-neoplastic sample from one or more subjects (which may, but need not, include the subject from whom the disease sample was collected to determine the variant-encoding sequence) and performing sequence determination analysis using the sample.

[0384] As used herein, a "pseudosequence" of an MHC molecule may refer to an ordered set of amino acids of an MHC molecule that contacts a peptide.

[0385] As used herein, a "representation" of a sequence may comprise a set of values ​​that represent or identify amino acids in the sequence and / or a set of values ​​that represent or identify nucleic acids that encode the sequence. For example, each amino acid may be represented by a binary string and / or a vector of values ​​that is distinct from another binary string and / or vector representing another amino acid. The representation may be generated using, for example, one-hot encoding or the BLOcks Substitution Matrix (BLOSUM). For example, a multidimensional (e.g., 20- or 21-dimensional) matrix is ​​initialized (e.g., randomly or pseudo-randomly initialized). The initialized matrix may contain, for each amino acid, a unique vector corresponding to that amino acid.Values ​​can be fixed such that the use of such a unique vector can be assumed to represent the corresponding amino acid. There can be many possible representations of nucleic acids of a given sequence, provided that any of the many codons can encode a single amino acid.

[0386] As used herein, "presentation" of a peptide refers to at least a portion of the peptide being displayed on the cell surface by binding to an MHC molecule in a specific manner. The displayed peptide can then be accessible to other cells, such as neighboring T cells.

[0387] As used herein, "specimen" may include tissue (e.g., a biopsy), a single cell, a plurality of cells, cell fragments, or an aliquot of bodily fluid. A sample may be obtained from a subject by, for example, but not limited to, venipuncture, excretion, ejaculation, massage, biopsy, needle aspiration, sample irrigation, scraping, surgical incision, intervention, other types of sample collection means, or a combination thereof.

[0388] As used herein, "subject" refers to one or more cells, tissues, or organisms. A subject may be human or non-human, in vivo, ex vivo, or in vitro, male or female. A subject may be a mammal, such as a human.

[0389] As used herein, "binding affinity" refers to the binding affinity between a peptide (e.g., a specific antigen) and an MHC (e.g., an MHC molecule and / or an MHC allele). Binding affinity can characterize the stability, tendency, and / or strength of binding between a peptide and an MHC molecule.

[0390] As used herein, "immunogenicity" may refer to the elicitation of an immune response (e.g., via T cells and / or B cells). A peptide that is "immunogenic" may be one that is capable of eliciting an immune response.

[0391] As used herein, "MHC" refers to the major histocompatibility complex. Human MHC is also called human leukocyte antigen (HLA) complex. VIII. Exemplary embodiments

[0392] Embodiment 1. A method is provided. The method comprises accessing a set of peptide sequences characterizing a set of peptides, wherein each peptide sequence from the set of peptide sequences was determined by processing a disease sample from a subject. The method comprises accessing an immunoglobulin complex (IPC) sequence identified for the immunoglobulin complex (IPC) of the subject.The method comprises processing a set of peptide representations, which represents a set of peptide sequences, using a first attention unit in an initial attention subsystem of an attention-based machine learning model, and a representation of an immunoglobulin complex (IPC), which represents an IPC sequence, using a second attention unit in the initial attention subsystem, to generate a result, wherein the result includes at least one of an interaction prediction, an interaction affinity prediction, or an immunogenicity prediction for the corresponding peptide-IPC combination. The method comprises generating a report based on the result.

[0393] Embodiment 2. The method of embodiment 1, wherein at least one peptide sequence of the set of peptide sequences comprises a variant-encoding sequence that comprises a variant relative to a corresponding reference sequence.

[0394]

[0395] Embodiment 3. The method of embodiment 1 or embodiment 2, wherein the processing comprises: obtaining a peptide representation from a set of peptide representations for a corresponding peptide sequence from a set of peptide sequences; and transforming the peptide representation by a first attention unit into a transformed peptide representation, wherein the first attention unit comprises a set of attention sub-units, wherein each attention sub-unit of the set of attention sub-units comprises an internal attention layer.

[0396] Embodiment 4. The method according to any one of embodiments 1-3, wherein the processing includes: obtaining a representation of the IPC; and transforming the representation of the IPC by a second attention unit into a transformed representation of the IPC, wherein the second attention unit comprises a set of attention sub-units, in which each attention sub-unit of the set of attention sub-units comprises an inner-attention layer.

[0397] Embodiment 5. The method according to any one of embodiments 1-4, wherein at least a portion of the peptide representation corresponds to a monomer in the peptide sequence and at least a portion of the IPC representation corresponds to a monomer in the IPC sequence; and wherein the processing comprises: generating a transformed peptide representation based on the peptide representation using a first attention unit and a first set of weights; generating a transformed IPC representation based on the IPC representation using a second attention unit and a second set of weights; and generating a composite representation using the transformed peptide representation and the transformed IPC representation.

[0398] Embodiment 6. The method according to any one of embodiments 1-5, which further comprises inserting a peptide sequence from a set of peptide sequences to generate an inserted peptide representation for the peptide sequence; and encoding, positionally, the inserted peptide representation for the peptide sequence to generate a peptide representation from the set of peptide representations that represents the peptide sequence.

[0399] Embodiment 7. The method according to any one of embodiments 1-6, wherein the first attention unit comprises a set of attention sub-units; and each attention sub-unit of the set of attention sub-units comprises a neural network that comprises at least one self-attention layer.

[0400] Embodiment 8. The method according to any one of embodiments 1-7, wherein the second attention unit comprises a set of attention sub-units; and each attention sub-unit of the set of attention sub-units comprises a neural network that comprises at least one self-attention layer.

[0401] Embodiment 9. The method according to any one of embodiments 1-8, in which the first attention unit comprises a first plurality of attention sub-units; the second attention unit comprises a first plurality of attention sub-units; and each attention sub-unit of the first set of attention sub-units and the second set of attention sub-units comprises a neural network that comprises at least one self-attention layer.

[0402] Embodiment 10. The method according to any one of embodiments 1-9, wherein the peptide representation from the set of peptide representations forms a first portion of an aggregate representation processed by the first attention unit; and the second portion of the aggregate representation represents at least one of an N-flanking sequence or a C-flanking sequence.

[0403] Embodiment 11. The method according to any one of embodiments 1-10, wherein a peptide sequence from a set of peptide sequences forms a first portion of an aggregate sequence; and a second portion of the aggregate sequence comprises at least one of an N-flanking sequence or a C-flanking sequence; and the attention-based machine learning model comprises a representation unit that receives and processes the sequence of aggregates to form a representation of the aggregate that comprises a representation of a peptide from a set of peptide representations corresponding to the peptide sequence, wherein the aggregate representation is processed using the first attention unit.

[0404] Embodiment 12. The method according to any one of embodiments 1-11, further comprising inserting an IPC sequence to generate an inserted IPC representation of the IPC sequence; and encoding, positionally, the inserted IPC representation from the IPC sequence to generate the IPC representation.

[0405] Embodiment 13. The method according to any one of embodiments 1-12, wherein the attention-based machine learning model comprises a plurality of self-attention layers and, for each of the plurality of self-attention layers, a corresponding top-down feedforward neural network.

[0406] Embodiment 14. The method according to any one of embodiments 1-13, wherein the first attention unit comprises a first neural network configured to receive and process a peptide representation from a set of peptide representations to generate a transformed peptide representation; and the second attention unit comprises a second neural network configured to receive and process an IPC representation to generate the transformed IPC representation; and wherein each of the first neural network and the second neural network comprises at least one self-attention layer; and wherein the attention-based machine learning model is configured to generate a composite representation using the transformed peptide representation and the transformed IPC representation.

[0407] Embodiment 15. The method according to any one of embodiments 1-14, wherein the attention-based machine learning model further comprises: a composite attention unit that comprises a neural network configured to receive and process a representation of the composite, wherein the neural network comprises an internal attention layer.

[0408] Embodiment 16. The method of any one of embodiments 1-15, wherein the attention-based machine learning model further comprises: a composite attention unit that comprises a set of attention sub-units, wherein each attention sub-unit of the set of attention sub-units comprises a neural network that comprises at least one self-attention layer. Embodiment 17. The method of any one of embodiments 1-16, wherein the IPC comprises a major histocompatibility complex (MHC), and the corresponding peptide-IPC combination comprises a peptide from the set of peptides and an MHC, and wherein predicting the interaction affinity for the corresponding peptide-IPC combination predicts the binding affinity between the peptide and the MHC; the interaction prediction for the corresponding peptide-IPC combination predicts whether the MHC will present the peptide on the cell surface.

[0409] Embodiment 18. The method according to any one of embodiments 1-17, wherein the attention-based machine learning model is trained using a training data set that comprises at least one of experimental interaction affinity data or experimental interaction data for a plurality of training peptide sequences and a set of training MHC sequences.

[0410] Embodiment 19. The method according to any one of embodiments 1-18, wherein the IPC is a T cell receptor (TCR), and the corresponding peptide-IPC pair comprises a peptide from a set of peptides and either a TCR or a TCR and a major histocompatibility complex (MHC), and wherein the immunogenicity prediction for the corresponding peptide-IPC combination predicts the immunogenicity of the peptide relative to the TCR; and the attention-based machine learning model is trained using a training data set that comprises experimental data for immunogenicity for a plurality of training peptide sequences and a set of training TCR sequences.

[0411] Embodiment 20. The method of any one of embodiments 1-19, wherein the training data set comprises a plurality of training data items, at least one training data item of the plurality of training data items comprises at least one of: a training peptide sequence characterizing a training peptide that is not included in the peptide set; a training IPC sequence characterizing a training IPC that is different from the IPC; and an experimental result identifying an indication of an affinity of interaction between the training peptide and the training IPC, wherein the indication of affinity of interaction was detected using a biosensor-based assay or methodology.

[0412] Embodiment 21. The method of any one of embodiments 1-20, wherein the training data set comprises a plurality of training data items, at least one training data item of the plurality of training data items comprises at least one of: a training peptide sequence characterizing a training peptide that is not included in the peptide set; a training MHC sequence characterizing a training MHC that is different from IPC; and a result based on an experiment including an interaction readout that identifies whether the training peptide was presented by the training MHC on the cell surface, wherein at least one of immunoprecipitation or mass spectrometry was used to determine the interaction readout.

[0413] Embodiment 22. The method of any one of embodiments 1-21, further comprising training an attention-based machine learning model before the processing step using a training data set that comprises at least one of binding affinities, interaction readings, or immunogenicity readings for a plurality of peptide-IPC combinations, wherein the training data set comprises a plurality of training peptide sequences and at least one of a plurality of major histocompatibility complex (MHC) training sequences or a plurality of T cell receptor (TCR) training sequences.

[0414] Embodiment 23. The method of any one of embodiments 1-22, wherein the processing comprises processing a set of peptide representations using a first attention unit and an IPC representation using a second attention unit to generate a set of composite representations for the set of peptide-IPC combinations; processing the set of composite representations to generate a set of results; selecting a subset of the set of peptide-IPC combinations, wherein the set of selected interactions is more likely to occur with each peptide-IPC combination of the subset compared to the remaining subset of the set of peptide-IPC combinations, wherein the report identifies each peptide in the subset.

[0415] Embodiment 24. The method according to any one of embodiments 1-23, wherein each peptide from the set of peptides is used to form a set of peptide-IPC combinations; and the attention-based machine learning model is configured to generate an immunogenicity prediction for each peptide-IPC combination from the set of peptide-IPC combinations, wherein the immunogenicity prediction for the peptide-IPC combination from the set of peptide-IPC combinations is a prediction of tumor-specific immunogenicity of the peptide in the peptide-IPC combination.

[0416] Embodiment 25. The method of any one of embodiments 1-24, wherein the report identifies a subset of peptides from the set of peptides with increased tumor-specific immunogenicity relative to the remainder of the set of peptides.

[0417] Embodiment 26. The method of any one of embodiments 1-25, wherein the IPC is a major histocompatibility complex (MHC); each peptide of the set of peptides is used to form a set of peptide-MHC combinations; and the attention-based machine learning model is configured to generate an interaction prediction for each peptide-MHC combination of the set of peptide-MHC combinations, wherein the interaction prediction for the peptide-MHC combination of the set of peptide-MHC combinations is a prediction of whether the peptide in the peptide-MHC combination is presented by the MHC on the cell surface.

[0418] Embodiment 27. The method of embodiment 26, wherein the report identifies a subset of peptides from the set of peptides with an increased probability of being represented by the MHC relative to the rest of the set of peptides.

[0419] Embodiment 28. The method according to any one of embodiments 1-27, wherein the peptide sequence of the set of peptide sequences is a variant-coding sequence characterizing a mutant peptide, wherein the variant-coding sequence comprises: a first portion identifying a sequence at the N-terminus of the mutant peptide; and a second portion identifying an epitope sequence of the mutant peptide; and the processing comprises: processing a first representation of the first portion of the variant-coding sequence using a first self-attention layer of the source attention subsystem; and processing a second representation of the second portion of the variant-coding sequence using a second self-attention layer of the source attention subsystem.

[0420] Embodiment 29. The method of embodiment 28, wherein the first representation and the second representation are processed in a first attention block.

[0421] Embodiment 30. The method according to any one of embodiments 1-29, wherein the attention-based machine learning model comprises one or more transformer encoders, wherein each of the one or more transformer encoders comprises an internal attention layer.

[0422] Embodiment 31. The method according to any one of embodiments 1-30, wherein the IPC sequence and each of the set of peptide sequences comprises an ordered set of amino acid identifiers.

[0423] Embodiment 32. The method according to any one of embodiments 1-31, wherein the IPC sequence is identified using a disease sample.

[0424] Embodiment 33. The method according to any one of embodiments 1-32, wherein the IPC sequence is identified using a biological sample from the subject.

[0425] Embodiment 34. The method according to any one of embodiments 1-33, wherein the disease sample comprises cancer cells.

[0426] Embodiment 35. The method according to any one of embodiments 1-34, wherein the IPC of the subject comprises a major histocompatibility complex (MHC); the sequence of the IPC comprises a sequence of the MHC; and the presentation of the IPC comprises a presentation of the MHC.

[0427] Embodiment 36. The method of embodiment 35, wherein the MHC comprises a class I MHC molecule.

[0428] Embodiment 37. The method of embodiment 35, wherein the MHC comprises a class II MHC molecule.

[0429] Embodiment 38. The method according to any one of embodiments 1-35, wherein the IPC of the subject comprises a T cell receptor (TCR); the sequence of the IPC comprises a sequence of the TCR; and the presentation of the IPC comprises the presentation of the TCR.

[0430] Embodiment 39. The method according to any one of embodiments 1-38, wherein the disease sample comprises tissue.

[0431] Embodiment 40. The method according to any one of embodiments 1-39, wherein at least one peptide from the set of peptides is a neoantigen.

[0432] Embodiment 41. The method according to any one of embodiments 1-40, wherein at least one peptide sequence of the set of peptide sequences is a genomic sequence obtained from a disease sample.

[0433] Embodiment 42. The method according to any one of embodiments 1-41, wherein each of the at least one set of variant encoding sequences is based on RNA sequences of a disease sample.

[0434] Embodiment 43. The method according to any one of embodiments 1-42, wherein the corresponding peptide-IPC combination comprises a peptide from the set of peptides and an IPC; the IPC is a major histocompatibility complex (MHC); the prediction of the interaction affinity is the prediction of the binding affinity for binding between the peptide and the MHC; and the interaction prediction is the prediction of the presentation of the peptide by the MHC on the cell surface.

[0435] Embodiment 44. The method of any one of embodiments 1-43, further comprising receiving input data entered by a user, wherein the input data corresponds to a subject; wherein accessing a set of peptide sequences and an IPC sequence in response to receiving the input data by retrieving from an array of data; and wherein the report identifies a subset of peptides from the set of peptides for inclusion in a personalized vaccine for treating a medical condition of the subject.

[0436] Embodiment 45. The method of embodiment 44 further includes generating a treatment recommendation for the subject that includes a personalized vaccine.

[0437] Embodiment 46. The method of any one of embodiments 1-45, further comprising receiving input data entered by a user, wherein the input data corresponds to a subject; wherein accessing a set of peptide sequences and an IPC sequence in response to receiving the input data by retrieving from an array of data; and determining a set of peptides for treatment for inclusion in a personalized vaccine based on the report; and initiating an action that facilitates the production of a personalized vaccine that includes the set of peptides for treatment.

[0438] Embodiment 47. The method of embodiment 46, wherein initiating the action comprises: generating an alert that activates a computerized process involved in producing a personalized vaccine.

[0439] Embodiment 48. The method according to any one of embodiments 1-47, wherein the processing comprises: receiving, from an insertion unit in an attention-based machine learning model, a representation that comprises a plurality of elements, wherein the representation is either a representation of a peptide from a set of peptide representations that represents a peptide sequence in a set of peptide sequences, or an IPC representation that represents an IPC sequence; and wherein each element in the data set of the plurality of elements corresponds to a monomer in either the peptide sequence or the IPC sequence; determining, for each element of the plurality of elements, a key vector, a value vector, and a query vector based on a set of key weights, a set of value weights, and a set of query weights, respectively, associated with an internal attention layer of the attention-based machine learning model;performing a transformation on a plurality of elements to form a plurality of modified elements, wherein the transformation is performed using attention indicators for the plurality of elements and a vector of values ​​determined for each of the plurality of elements; and generating a result based on the plurality of modified elements.

[0440] Embodiment 49. The method of embodiment 48, wherein performing a transformation on a selected element from a plurality of elements includes determining an attention score of the selected element using a key vector and a query vector of the element, wherein the remainder of the plurality of elements other than the selected element forms a set of remaining elements; determining an additional attention score for each remaining element from the set of remaining elements using the key vector of the remaining element and the query vector of the selected element to form a set of additional attention scores; and generating a modified element using the attention score, the set of additional attention scores, and the vector of values ​​of each element from the plurality of elements.

[0441] Embodiment 50. The method of any one of embodiments 1-49, which further comprises displaying the report on a graphical user interface on the display system.

[0442] Embodiment 51. The method of any one of embodiments 1-50, wherein the processing is performed on the first computing platform and further includes sending a report to the second computing platform regarding the set of communication links that comprise at least one of wired communication links or wireless communication links.

[0443] Embodiment 52. The method according to any one of embodiments 1-51, further comprising determining to include at least one peptide from the set of peptides as a target for immunotherapy based on the report.

[0444] Embodiment 53. The method of embodiment 52, wherein the immunotherapy is selected from the group consisting of T cell therapy, personalized cancer therapy, antigen-specific immunotherapy, antigen-dependent immunotherapy, a vaccine, and natural killer (NK) cell therapy.

[0445] Embodiment 54. The method according to any one of embodiments 1-53 further comprises determining to exclude at least one peptide from the set of peptides as a target for immunotherapy based on the report.

[0446] Embodiment 55. The method of embodiment 54, wherein the immunotherapy is selected from the group consisting of T cell therapy, personalized cancer therapy, antigen-specific immunotherapy, antigen-dependent immunotherapy, a vaccine, and natural killer (NK) cell therapy.

[0447] Embodiment 56. The method according to any one of embodiments 1-55, wherein the IPC is a human leukocyte antigen (HLA) molecule.

[0448] Embodiment 57. The method of any one of embodiments 1-56, further comprising determining a sequence of a disease sample from the subject; determining a set of peptide sequences based on the determination of a sequence of a disease sample from the subject; identifying, based on the report, a subset of the set of peptide sequences; synthesizing mRNA that encodes at least one peptide included in the subset of the set of peptides; complexing the mRNA with lipids to produce an mRNA-lipoplex therapeutic agent; and administering the mRNA-lipoplex therapeutic agent to the subject.

[0449] Embodiment 58. The vaccine comprises one or more peptides; a plurality of nucleic acids that encode the one or more peptides; or a plurality of cells expressing one or more peptides, wherein the one or more peptides are selected from a set of peptides based on a report generated by the method of any one of embodiments 1-49, wherein the one or more peptides represent a partial subset of the set of peptides.

[0450] Embodiment 59. The vaccine of embodiment 58, wherein the vaccine comprises either DNA that contains a plurality of nucleic acids or RNA that contains a plurality of nucleic acids.

[0451] Embodiment 60. The vaccine of embodiment 58 or embodiment 59, wherein the vaccine comprises mRNA that comprises a plurality of nucleic acids.

[0452] Embodiment 61. The vaccine of any one of embodiments 58-60, wherein the vaccine is an antitumor vaccine.

[0453] Embodiment 62. A method for producing a vaccine comprises obtaining a vaccine comprising: one or more peptides; a plurality of nucleic acids that encode the one or more peptides; or a plurality of cells expressing one or more peptides, wherein the one or more peptides are selected from a set of peptides based on a report generated by the method of any one of embodiments 1-49, wherein the one or more peptides represent a partial subset of the set of peptides.

[0454] Embodiment 63. The method of embodiment 62, wherein the vaccine comprises DNA that contains a plurality of nucleic acids, RNA that contains a plurality of nucleic acids, or mRNA that contains a plurality of nucleic acids.

[0455] Embodiment 64. The method of embodiment 62 or embodiment 63 further comprises identifying, based on amino acids in one or more peptides, a plurality of nucleic acids that encode the one or more peptides, wherein the vaccine comprises the plurality of nucleic acids.

[0456] Embodiment 65. The method of any one of embodiments 62-64, wherein the vaccine is a tumor vaccine.

[0457] Embodiment 66. The method of embodiment 65, wherein for each peptide of the one or more peptides, the tumor vaccine comprises at least one of: a nucleotide sequence encoding each peptide, an amino acid sequence corresponding to each peptide, RNA corresponding to each peptide, DNA corresponding to each peptide, a cell corresponding to each peptide, a plasmid corresponding to each peptide, or a vector corresponding to each peptide.

[0458] Embodiment 67. The method of any one of embodiments 62-66, wherein the vaccine further comprises at least one of an excipient or an adjuvant.

[0459] Embodiment 68. The method of any one of embodiments 62-67, wherein the vaccine comprises an RNA molecule comprising, in 5'→3' direction: a 5' cap; a 5' untranslated region (UTR); a polynucleotide sequence encoding a secretory signal peptide; a polynucleotide sequence encoding one or more peptides; a polynucleotide sequence encoding at least a portion of a transmembrane and cytoplasmic domain of a major histocompatibility complex (MHC) molecule; a 3' UTR including: a 3' untranslated region of an amino-terminal cleavage enhancer (ACTE) mRNA or a fragment thereof; and a non-coding RNA of mitochondrially encoded 12S RNA or a fragment thereof; and a poly(A) sequence.

[0460] Embodiment 69. The pharmaceutical composition comprises one or more peptides selected from a set of peptides based on a report generated by the method of any one of embodiments 1-49, wherein the one or more peptides are a partial subset of the set of peptides.

[0461] Embodiment 70. The pharmaceutical composition comprises a nucleic acid sequence that encodes one or more peptides that have been selected from a set of peptides based on a report generated by the method of any one of embodiments 1-49, wherein the one or more peptides are a partial subset of the set of peptides.

[0462] Embodiment 71. An immunogenic peptide is identified based on a report generated by the method of any one of embodiments 1-49.

[0463] Embodiment 72. The nucleic acid sequence is identified based on a report generated by the method of any one of embodiments 1-49.

[0464] Embodiment 73. The nucleic acid sequence of embodiment 72, wherein the nucleic acid sequence comprises a DNA sequence.

[0465] Embodiment 74. The nucleic acid sequence of embodiment 72 or embodiment 73, wherein the nucleic acid sequence comprises an RNA sequence.

[0466] Embodiment 75. The nucleic acid sequence of any one of embodiments 72-74, wherein the nucleic acid sequence comprises an mRNA sequence.

[0467] Embodiment 76. A method of treating a subject comprises administering at least one of one or more peptides, one or more pharmaceutical compositions, or one or more nucleic acid sequences identified based on a report generated by the method of any one of embodiments 1-49.

[0468] Embodiment 77. The method includes processing a set of biological samples obtained from a subject to generate a set of peptide sequences characterized by the set of peptides; processing the set of biological samples obtained from the subject to generate an immunoglobulin complex (IPC) sequence identified for an immunoglobulin complex (IPC) of the subject; generating a set of peptide representations that represents the set of peptide sequences using a first attention unit in an initial attention subsystem of an attention-based machine learning model; generating a representation of the immunoglobulin complex (IPC) that represents the IPC sequence using a second attention unit in the initial attention subsystem;processing a set of peptide representations and an IPC representation to generate a result, wherein the result comprises at least one of an interaction prediction, an interaction affinity prediction, or an immunogenicity prediction for a corresponding peptide-IPC combination, wherein the corresponding peptide-IPC combination comprises a peptide from the set of peptides;

[0469] Embodiment 78. The method of embodiment 77, wherein processing a set of biological samples obtained from a subject to generate peptide sequences comprises processing a disease sample in the set of biological samples obtained from the subject to generate a set of peptide sequences.

[0470] Embodiment 79. The method of embodiment 77 or embodiment 78 further comprises obtaining a set of biological samples from the subject, wherein the set of biological samples comprises a disease sample.

[0471] Embodiment 80. The method according to any one of embodiments 77-79 further includes generating a report based on the result.

[0472] Embodiment 81. The method includes receiving at a user device a request to develop a personalized vaccine for a subject; transmitting from the user device a communication to a remote system, wherein the communication includes an identifier of the subject, wherein the remote system is configured to: access a set of peptide sequences characterizing a set of peptides, wherein each peptide sequence of the set of peptide sequences has been identified by processing a disease sample from the subject, and access a sequence of an immunoglobulin complex (IPC) identified for an immunoglobulin complex (IPC) of the subject;processing a set of peptide representations that represents a set of peptide sequences using a first attention unit in an initial attention subsystem of an attention-based machine learning model and an immunoglobulin complex (IPC) representation that represents an IPC sequence using a second attention unit in the initial attention subsystem to generate a result, wherein the result includes at least one of an interaction prediction, an interaction affinity prediction, or an immunogenicity prediction for the corresponding peptide-IPC combination; and generating a report based on the result; and transmitting the report to a user device; and receiving the report at the user device.

[0473] Embodiment 82. The method of embodiment 81 further comprises collecting a disease sample from a subject; eluting a plurality of peptides that comprise a set of peptides from MHC molecules into the disease sample using at least one of chromatography or mass spectrometry; determining the sequence of the set of peptides to generate a set of initial sequences; comparing each initial sequence of the set of initial sequences with a reference sequence; and determining a set of peptide sequences based on the comparisons, wherein each peptide sequence in the set of peptide sequences is a sequence encoding a variant that comprises a variant relative to the reference sequence.

[0474] Embodiment 83. A method for producing a treatment for a subject is provided. The method includes receiving a report from a computing device that is configured to: access a set of peptide sequences that characterize a set of peptides, wherein each peptide sequence from the set of peptide sequences was identified by processing a disease sample from the subject, and access an immunoglobulin complex (IPC) sequence identified for the immunoglobulin complex (IPC) of the subject;processing a set of peptide representations that represents a set of peptide sequences using a first attention unit in an initial attention subsystem of an attention-based machine learning model and an immunoglobulin complex (IPC) representation that represents an IPC sequence using a second attention unit in the initial attention subsystem to generate a result, wherein the result includes at least one of an interaction prediction, an interaction affinity prediction, or an immunogenicity prediction for a corresponding peptide-IPC combination; and generating a report based on the result; and generating a treatment agent manufacturing plan for manufacturing the treatment agent based on the report.

[0475] Embodiment 84. The method of embodiment 83 further includes producing the treatment agent based on the treatment agent production plan.

[0476] Embodiment 85. The method includes inputting a plurality of variant coding sequences characterizing a plurality of mutant peptides into an attention-based machine learning model, wherein each variant coding sequence of the plurality of variant coding sequences has been identified by processing a disease sample from a subject;inputting an immunoglobulin complex (IPC) sequence identified for an immunoglobulin complex (IPC) of a subject into an attention-based machine learning model, wherein the attention-based machine learning model is configured to process a plurality of variant representations that represents a plurality of variant-encoding sequences using a first attention unit in an initial attention subsystem of the attention-based machine learning model and a representation of the immunoglobulin complex (IPC) that represents the IPC sequence using a second attention unit in the initial attention subsystem to generate a result, wherein the result comprises at least one of an interaction prediction, an interaction affinity prediction, or an immunogenicity prediction for a corresponding mutant peptide-IPC combination; and receiving a report generated based on the result; and selecting, based on the report, a subset of the plurality of mutant peptides for use in treating the subject.

[0477] Embodiment 86. The method includes obtaining a peptide sequence that characterizes a mutant peptide, wherein the peptide sequence comprises a variant relative to a corresponding reference sequence; obtaining an MHC sequence identified for a major histocompatibility complex (MHC); processing the peptide sequence and the MHC sequence using different processing paths in an attention-based machine learning model to generate a result, wherein the result provides information about immunological activity relative to both the mutant peptide and the MHC; generating a report based on the result.

[0478] Embodiment 87. The method of embodiment 86, wherein the processing comprises processing the peptide sequence through a peptide processing path in an attention-based machine learning model, wherein the peptide processing path comprises a first insertion unit and a first attention unit that comprises at least one self-attention layer; and

[0479] processing the MNN sequence through an MNN processing path in an attention-based machine learning model, wherein the MNN processing path comprises a second insertion unit and a second attention unit that comprises at least one self-attention layer.

[0480] Embodiment 88. The method of embodiment 87 further comprises obtaining a TCR sequence identified for a T cell receptor (TCR); and wherein the processing further comprises processing the TCR sequence through a TCR processing pathway in an attention-based machine learning model, wherein the TCR processing pathway comprises a third insertion unit and a third attention unit that comprises at least one self-attention layer.

[0481] Embodiment 89. The method of any one of embodiments 86-88, wherein the immunological activity comprises an immune response, and the information comprises predicting the ability of the mutant peptide to provoke an immune response.

[0482] Embodiment 90. The method of any one of embodiments 86-89, wherein the processing comprises generating a transformed peptide representation from a peptide sequence via a peptide processing pathway; generating a transformed MHC representation from the MHC sequence via an MHC processing pathway; generating a composite representation using the transformed peptide representation and the transformed MHC representation; processing the composite representation to generate a result.

[0483] Embodiment 91. The method of any one of embodiments 86-90, wherein the immunological activity comprises binding of the mutant peptide to MHC, and wherein the result comprises at least one of a first prediction corresponding to whether the mutant peptide binds to MHC, or a second prediction corresponding to an affinity associated with the binding.

[0484] Embodiment 92. The method of any one of embodiments 86-91 further comprises determining to include the mutant peptide as a target for immunotherapy based on the report.

[0485] Embodiment 93. The method of embodiment 92, wherein the immunotherapy is selected from the group consisting of T cell therapy, personalized cancer therapy, antigen-specific immunotherapy, antigen-dependent immunotherapy, a vaccine, and natural killer (NK) cell therapy.

[0486] Embodiment 94. The method of any one of embodiments 86-93 further comprises at least one of: determining to exclude the mutant peptide as a target for immunotherapy based on the report.

[0487] Embodiment 95. The method of embodiment 94, wherein the immunotherapy is selected from the group consisting of T cell therapy, personalized cancer therapy, antigen-specific immunotherapy, antigen-dependent immunotherapy, a vaccine, and natural killer (NK) cell therapy.

[0488] Embodiment 96. The method of any one of embodiments 86-95 further comprises determining, based on the report, the inclusion of at least one of the mutant peptide, the precursor of the mutant peptide, nucleic acids that encode the mutant peptide, or a plurality of cells that express the mutant peptide, in the treatment agent; and manufacturing the treatment agent.

[0489] Embodiment 97. The method of embodiment 96 further comprises treating the subject with the treatment agent.

[0490] Embodiment 98. The method of any one of embodiments 86-97, wherein a peptide sequence characterized by a mutant peptide was identified by sequencing a disease sample from a subject, wherein the peptide sequence has at least one sequence variation relative to a corresponding reference sequence, and wherein a treatment is developed for the subject based on the report.

[0491] Embodiment 99. The method includes obtaining a peptide sequence that is characterized by a mutant peptide, wherein the peptide sequence comprises a variant relative to a corresponding reference sequence; obtaining a TCR sequence identified for a T cell receptor (TCR); processing the peptide sequence and the TCR sequence using different processing paths in an attention-based machine learning model to generate a result, wherein the result provides information about immunological activity relative to both the mutant peptide and the TCR; generating a report based on the result.

[0492] Embodiment 100. The method of embodiment 99, wherein the processing comprises processing a peptide sequence through a peptide processing pathway in an attention-based machine learning model, wherein the peptide processing pathway comprises a first insertion unit and a first attention unit; and processing a TCR sequence through a TCR processing pathway in an attention-based machine learning model, wherein the TCR processing pathway comprises a second insertion unit and a second attention unit.

[0493] Embodiment 101. The method of embodiment 100 further comprises obtaining an MHC sequence identified for a major histocompatibility complex (MHC); and wherein the processing further comprises processing the MHC sequence through an MHC processing path in an attention-based machine learning model, wherein the MHC processing path comprises a third insertion block and a third MHC block.

[0494] Embodiment 102. The method of any one of embodiments 99-101, wherein the immunological activity comprises an immune response, and the information comprises predicting the ability of the mutant peptide to provoke an immune response.

[0495] Embodiment 103. The method of any one of embodiments 99-102, wherein the processing comprises generating a converted peptide representation from a peptide sequence via a peptide processing pathway; generating a converted TCR representation from a TCR sequence via a TCR processing pathway; generating a composite representation using the converted peptide representation and the converted TCR representation; processing the composite representation to generate a result.

[0496] Embodiment 104. The method of any one of embodiments 99-103, wherein the immunological activity comprises binding of the mutant peptide to MHC, and wherein the result comprises at least one of a first prediction corresponding to whether the mutant peptide binds to MHC, or a second prediction corresponding to the affinity associated with the binding.

[0497] Embodiment 105. The method of any one of embodiments 99-104 further comprises determining to include the mutant peptide as a target for immunotherapy based on the report.

[0498] Embodiment 106. The method of embodiment 105, wherein the immunotherapy is selected from the group consisting of T cell therapy, personalized cancer therapy, antigen-specific immunotherapy, antigen-dependent immunotherapy, a vaccine, and natural killer (NK) cell therapy.

[0499] Embodiment 107. The method of any one of embodiments 99-106 further comprises at least one of: determining to exclude the mutant peptide as a target for immunotherapy based on the report.

[0500] Embodiment 108. The method of embodiment 107, wherein the immunotherapy is selected from the group consisting of T cell therapy, personalized cancer therapy, antigen-specific immunotherapy, antigen-dependent immunotherapy, a vaccine, and natural killer (NK) cell therapy.

[0501] Embodiment 109. The method of any one of embodiments 99-108 further comprises determining, based on the report, the inclusion of at least one of the mutant peptide, the precursor of the mutant peptide, nucleic acids that encode the mutant peptide, or a plurality of cells that express the mutant peptide, in the treatment agent; and manufacturing the treatment agent.

[0502] Embodiment 110. The method of embodiment 109 further comprises treating the subject with the treat...

Claims

1. A computer-implemented method for generating predictions regarding whether peptides of interest will experience a target interaction with an immunoglobulin complex (IPC), comprising: accessing a set of peptide sequences characterizing a set of peptides, wherein each peptide sequence in the set of peptide sequences was identified by processing a disease sample from a subject; accessing an immunoglobulin complex (IPC) sequence identified for an immunoglobulin complex (IPC) of a subject, wherein the IPC comprises a major histocompatibility complex (MHC); processing a set of peptide representations that represents a set of peptide sequences using a first attention unit in an initial attention subsystem of an attention-based machine learning model, and an immunoglobulin complex (IPC) representation that represents an IPC sequence using a second attention unit in the initial attention subsystem, to generate a result, wherein the result includes at least one of an interaction prediction or an interaction affinity prediction for a corresponding peptide-IPC combination, wherein the corresponding peptide-IPC combination includes a peptide from the set of peptides and an MHC, the interaction affinity prediction for the corresponding peptide-IPC combination predicts the binding affinity between the peptide and the MHC, and the interaction prediction for the corresponding peptide-IPC combination predicts whether the MHC will present the peptide on the cell surface; and generating a report based on the result.

2. The method according to claim 1, wherein at least one peptide sequence from the set of peptide sequences contains a variant-encoding sequence that contains a variant relative to the corresponding reference sequence.

3. A method according to any of the preceding paragraphs, wherein the processing comprises: obtaining a peptide representation from a set of peptide representations for a corresponding peptide sequence from a set of peptide sequences and transforming the peptide representation by a first attention block into a transformed peptide representation, wherein the first attention block comprises a set of attention sub-blocks, in which each attention sub-block of the set of attention sub-blocks comprises an internal attention layer.

4. A method according to any of the preceding paragraphs, wherein the processing comprises: getting IPC representation and transforming the IPC representation by a second attention unit into a transformed IPC representation, wherein the second attention unit comprises a set of attention sub-units, in which each attention sub-unit of the set of attention sub-units comprises an internal attention layer.

5. The method according to any one of the preceding claims, wherein at least a portion of the peptide representation corresponds to a monomer in the peptide sequence and at least a portion of the IPC representation corresponds to a monomer in the IPC sequence; and wherein the processing comprises: generating a transformed representation of the peptide based on the representation of the peptide using a first attention block and a first set of weights; generating a transformed IPC representation based on the IPC representation using the second attention unit and the second set of weights; and generating a composite representation using the converted peptide representation and the converted MHC representation.

6. A method according to any of the preceding paragraphs, which further comprises: or inserting a peptide sequence from a set of peptide sequences to generate an inserted peptide representation for the peptide sequence; and encoding, positionally, the inserted peptide representation for the peptide sequence to generate a peptide representation from the set of peptide representations that represents the peptide sequence; or inserting an IPC sequence to generate an inserted IPC representation of the IPC sequence; and encoding, positionally, the inserted IPC representation from the IPC sequence to generate an IPC representation.

7. A method according to any of the preceding paragraphs, in which: the first attention block contains a set of attention subblocks; and each attention subunit of the set of attention subunits comprises a neural network that comprises at least one self-attention layer; and / or the second attention block contains a set of attention subblocks; and Each attention subunit of the set of attention subunits contains a neural network that contains at least one self-attention layer.

8. A method according to any of the preceding paragraphs, in which: a peptide representation from the set of peptide representations forms a first part of the aggregate representation processed by the first attention unit; and the second part of the aggregate representation represents at least one of the N-flanking sequence or the C-flanking sequence.

9. A method according to any of the preceding paragraphs, in which: a peptide sequence from a set of peptide sequences forms the first part of the sequence of the aggregate; and the second portion of the aggregate sequence comprises at least one of an N-flanking sequence or a C-flanking sequence; and an attention-based machine learning model comprises a representation unit that receives and processes an aggregate sequence to form an aggregate representation that comprises a peptide representation from a set of peptide representations corresponding to the peptide sequence, wherein the aggregate representation is processed using a first attention unit.

10. The method according to any one of the preceding claims, wherein the attention-based machine learning model comprises a plurality of self-attention layers and, for each of the plurality of self-attention layers, a corresponding top-down feedforward neural network.

11. A method according to any of the preceding paragraphs, in which: the first attention unit comprises a first neural network configured to receive and process a peptide representation from a set of peptide representations to generate a transformed peptide representation; and the second attention unit comprises a second neural network configured to receive and process the IPC representation to generate a transformed IPC representation; and wherein each of the first neural network and the second neural network comprises at least one self-attention layer; and wherein the attention-based machine learning model is configured to generate a composite representation using the transformed peptide representation and the transformed IPC representation.

12. The method according to any one of the preceding claims, wherein the attention-based machine learning model further comprises: an attention unit for composites that comprises a neural network configured to receive and process a representation of the composite, wherein the neural network comprises a self-attention layer.

13. The method according to any one of the preceding claims, wherein the attention-based machine learning model further comprises: an attention unit for composites that comprises a set of attention sub-units, wherein each attention sub-unit of the set of attention sub-units comprises a neural network that comprises at least one self-attention layer.

14. The method according to any one of the preceding paragraphs, wherein the attention-based machine learning model is trained using a set of training data that comprises at least one of experimental data for interaction affinity or experimental data for interaction for a plurality of training peptide sequences and a set of training MHC sequences.

15. A method according to any of the preceding paragraphs, wherein the processing comprises: processing the set of peptide representations using a first attention unit and the IPC representation using a second attention unit to generate a set of composite representations for the set of peptide-IPC combinations; processing a set of composite representations to generate a set of results; selecting a subset from the set of peptide-IPC combinations, wherein the set of selected interactions is more likely to occur with each peptide-IPC combination of the subset compared to the remaining subset from the set of peptide-IPC combinations, where the report identifies each peptide in the subset.

16. A method according to any of the preceding paragraphs, wherein: each peptide from the set of peptides is used to form a set of peptide-MHC combinations; and an attention-based machine learning model is configured to generate an interaction prediction for each peptide-MHC combination of a set of peptide-MHC combinations, wherein the interaction prediction for a peptide-MHC combination of a set of peptide-MHC combinations is a prediction of whether a peptide in the peptide-MHC combination will be presented by an MHC on the cell surface.

17. The method of claim 16, wherein the report identifies a subset of peptides from the set of peptides with an increased probability of representing MHC relative to the remainder of the set of peptides.

18. A method according to any of the preceding paragraphs, wherein: a peptide sequence from a set of peptide sequences is a variant coding sequence characterizing a mutant peptide, wherein the variant coding sequence comprises: a first portion identifying the sequence at the N-terminus of the mutant peptide; and a second part identifying the epitope sequence of the mutant peptide; and processing includes: processing a first representation of a first portion of the variant encoding sequence using a first internal attention layer of the initial attention subsystem; and processing the second representation of the second part of the encoding variant sequence using the second layer of internal attention of the original attention subsystem.

19. The method of any one of the preceding claims, wherein the attention-based machine learning model comprises one or more transformer encoders, wherein each of the one or more transformer encoders comprises an internal attention layer.

20. The method of any one of the preceding claims, wherein the IPC sequence is identified using a disease sample and / or a biological sample from the subject.

21. A method according to any of the preceding paragraphs, in which: the IPC sequence contains the MHC sequence; and The IPC representation contains the MHC representation.

22. The method of claim 21, wherein the MHC comprises a class I MHC molecule.

23. The method of claim 21, wherein the MHC comprises a class II MHC molecule.

24. A method according to any of the preceding paragraphs, in which: The subject's IPC contains a T cell receptor (TCR); the IPC sequence contains the TCR sequence; and The IPC view contains a TCR view.

25. The method according to any one of claims 2-24, wherein each of the at least one set of variant coding sequences is based on RNA sequences of a disease sample.

26. The method according to any of the preceding paragraphs, which further comprises: receiving input data entered by a user, wherein the input data corresponds to the subject; wherein the set of peptide sequences and the IPC sequence are accessed in response to receiving input data by retrieval from a data array; and wherein the report identifies a subset of peptides from the set of peptides for inclusion in a personalized vaccine for treating a medical condition of a subject.

27. The method according to paragraph 26, which further comprises: generating a treatment recommendation for the subject that includes a personalized vaccine.

28. The method according to any of the preceding paragraphs, which further comprises: receiving input data entered by a user, wherein the input data corresponds to the subject; wherein the set of peptide sequences and the IPC sequence are accessed in response to receiving input data by retrieval from a data array; and identifying a set of peptides for treatment to be included in a personalized vaccine based on the report; and initiating an action that facilitates the production of a personalized vaccine that includes the set of peptides for treatment.

29. A method according to any of the preceding paragraphs, wherein the processing comprises: obtaining from the insertion unit into the attention-based machine learning model a representation that comprises a plurality of elements, wherein the representation is either a peptide representation from a set of peptide representations that represents a peptide sequence in a set of peptide sequences, or an IPC representation that represents an IPC sequence; and where each element in the multi-element data set corresponds to a monomer in either the peptide sequence or the IPC sequence; determining for each element of the plurality of elements a key vector, a value vector, and a query vector based on a set of key weights, a set of value weights, and a set of query weights, respectively, associated with an internal attention layer of the attention-based machine learning model; performing a transformation on a plurality of elements to form a plurality of modified elements, wherein the transformation is performed using attention indicators generated for the plurality of elements and a value vector determined for each of the plurality of elements; and generating a result based on a set of modified elements.

30. The method of claim 29, wherein performing the transformation for a selected element from a plurality of elements comprises: determining an attention score of a selected element using a key vector and a query vector of the element, wherein the remaining portion of the set of elements other than the selected element forms a set of remaining elements; determining an additional attention metric for each remaining element from the set of remaining elements using the key vector of the remaining element and the query vector of the selected element to form a set of additional attention metrics; and generating a modified element using the attention metric, the set of additional attention metrics, and the value vector of each element from the plurality of elements.

31. The method according to any one of the preceding claims, wherein the IPC is a human leukocyte antigen (HLA) molecule.

32. A prediction system for generating predictions regarding whether peptides of interest will experience a target interaction with an immunoglobulin complex (IPC), comprising: one or more data processing devices and a non-volatile computer-readable storage medium containing instructions, the execution of which on one or more data processing devices causes the one or more data processing devices to perform the method according to any one of paragraphs 1-31.

33. A machine-readable storage medium containing instructions, the execution of which on one or more data processing devices causes the one or more data processing devices to perform the method according to any one of paragraphs 1-31.

Images