Neural network-based neoantigen selection

Abstract

Disclosed herein are prediction models useful for determining a personalized prediction for an individual. For example, prediction models can predict whether a candidate neoantigen is likely to be immunogenic. Selected candidate neoantigens that are likely to be immunogenic can be included in a therapeutic vaccine for administering to the individual. To train a prediction model, a high-quality training dataset of neoantigens is constructed and further augmented to improve the performance of the prediction model. Prediction models are trained on individual-level labels, which are more readily obtainable and capture aggregate immune responses observed in individual patients.

Description

Attorney Docket No. N.093.W0.01NEURAL NETWORK-BASED NEOANTIGEN SELECTIONCROSS REFERENCE TO RELATED APPLICATIONS

[0001] This application claims the benefit of and priority to U.S. Provisional Patent Application No. 63 / 789,800, titled “Neural Network-Based Neoantigen Selection,” filed on April 16, 2025, Greek Provisional Patent Application No. 20250100386, titled “Neural Network-Based Neoantigen Selection,” filed on May 22, 2025, and U.S. Provisional Patent Application No. 63 / 827,608, titled “Neural Network-Based Neoantigen Selection,” filed on June 20, 2025, the entire contents of each of which are incorporated by reference herein.BACKGROUND

[0002] Personalized cancer vaccines are revolutionizing cancer treatment by offering tailored therapies that exploit the unique molecular profile of a patient’s tumor. At the core of these vaccines are neoantigens, which are tumor-specific peptides resulting from somatic mutations that are presented on major histocompatibility complexes (MIICs) and recognized by T cells as foreign. These neoantigens are valuable therapeutic targets because they are absent from normal tissues, reducing the risk of autoimmunity while enabling a precise immune attack on tumor cells. The therapeutic potential of personalized cancer vaccines depends on the accurate identification and prioritization of the most immunogenic neoantigens from the vast pool of mutations present in a tumor.

[0003] Accordingly, there is a need for improved computational frameworks for neoantigen prioritization that can effectively use available immune response data, process variable-size neoantigen collections, and generate clinically useful immunogenicity predictions for personalized cancer vaccine design.SUMMARY

[0004] Personalized cancer vaccines rely on accurately identifying immunogenic neoantigens that can trigger an immune response. However, predicting which neoantigens are truly immunogenic remains a major challenge. Conventional approaches to neoantigen prediction have relied on analyzing biochemical properties of individual candidate neoantigens, such as predicted binding affinity to MHC molecules and peptide-MHC complex stability. While these biochemical-property-based methods can provide usefulAttorney Docket No. N.093.W0.01signals, they often fail to capture the full complexity of in vivo immune responses. One challenge is the limited availability of experimentally validated data at the epitope level. Epitope-level annotations require labor-intensive assays for each candidate neoantigen, and as a result, training datasets with epitope-level labels tend to be small and incomplete. This scarcity of fine-grained labels constrains the performance of models trained to predict neoantigen immunogenicity on a per-epitope basis.

[0005] Another challenge arises from the inherent noise present in neoantigen datasets. For any given patient, the number of candidate neoantigens derived from somatic mutations can be large, yet immune responses to neoantigens exhibit inherent sparsity, as typically a small subset of candidate epitopes triggers an immune response. Including all candidate neoantigens during model training without appropriate filtering can introduce substantial noise, since the majority of epitopes in each patient’s set are likely non-immunogenic.Furthermore, each patient’s set of candidate neoantigens varies in size depending on the tumor mutational burden, which presents additional challenges for machine learning architectures that expect fixed-size inputs.

[0006] Disclosed herein is a computational framework for neoantigen prioritization that leverages patient- level immune response information, such as patient-level immunogenicity labels. Patient-level immunogenicity labels are more easily available and capture aggregate immune responses observed in individual patients. Each patient in the dataset has a set of epitopes or neoantigens associated with their somatic mutations, along with a label indicating whether the patient had an immune response to their tumor. The disclosed framework can use patient- level immune response information to train a prediction model and generate candidate-neoantigen-specific immunogenicity scores. By training a model on patient-level labels, informative predictions are made, even in the absence of extensive epitope- level annotations. In some embodiments, the disclosed framework uses a permutation-invariant architecture configured to process unordered neoantigen sets having variable cardinalities. The disclosed framework can generate predictions without requiring an artificial ordering of neoantigens or forcing patient-specific neoantigen collections into a rigid fixed-position input format

[0007] Disclosed herein is a pipeline involving a multiple step process including 1) a sampling step for generating higher quality training data by analyzing biochemical properties of neoantigens, 2) augmenting the training data, and 3) training a prediction model with the disclosed architecture using the higher quality training data. These steps can reduce trainingAttorney Docket No. N.093.W0.01noise and improve robustness of machine-learning training across neoantigen sets having different cardinalities, thereby enabling improved personalized predictions of candidate neoantigens during inference. Comparatively, the disclosed pipeline using the multiple-step process generates a prediction model that outperforms conventional processes that generate predictions for neoantigens based on biochemical properties of those neoantigens.

[0008] In some implementations, the disclosed prediction approach provides a technical improvement in computerized processing of biological sequence data by enabling machine analysis of unordered neoantigen sets having variable cardinalities. Shared sub-structures can be applied to respective neoantigens, and a pooling structure can transform multiple neoantigen representations into a fixed-length individual-level representation suitable for downstream machine classification. This framework can improve computerized handling of patient-specific neoantigen data structures relative to approaches that rely on order-dependent or fixed-length input representations.

[0009] Disclosed herein is a method for assessing immunogenicity of a candidate neoantigen for an individual, the method comprising: obtaining or having obtained a peptide sequence of the candidate neoantigen for the individual; analyzing the epitope using a prediction model, wherein the analyzing comprises: generating one or more candidate neoantigen representations of the epitope; pooling the one or more candidate neoantigen representations into a single individual-level representation; and generating a score for the candidate neoantigen using the single individual-level representation.

[0010] Further disclosed herein is a computer system for implementing the aforementioned method. In particular, disclosed herein is a computer system for assessing immunogenicity of a candidate neoantigen for an individual, the computer system comprising a processor and a memory. The memory stores instructions that, when executed by the processor, cause the computer system to: receive a peptide sequence of the candidate neoantigen for the individual; analyze the epitope using a prediction model stored on or accessed by the computer system, wherein the analyzing comprises: generating one or more candidate neoantigen representations of the epitope; pooling the one or more candidate neoantigen representations into a single individual-level representation; and generating a score for the candidate neoantigen using the single individual-level representation.

[0011] In various embodiments, the candidate neoantigen comprises a peptide sequence comprising between about 5 to about 50 amino acids, optionally wherein the peptide sequence comprises between about 10 to about 20 amino acids. In various embodiments,Attorney Docket No. N.093.W0.01obtaining or having obtained the peptide sequence of the candidate neoantigen for the individual comprises performing sequencing of nucleic acids from the individual and using the nucleic acid sequences to obtain the peptide sequence of the candidate neoantigen. In various embodiments, the sequencing is next generation sequencing (NGS) or any massively parallel sequencing approach. In various embodiments, performing sequencing of nucleic acids comprises performing sequencing on tumor tissue comprising the nucleic acids from the individual. In various embodiments, performing sequencing of nucleic acids comprises performing sequencing on nucleic acids in a sample obtained from the individual.

[0012] In various embodiments, the sample is one of a blood sample, a serum sample, a stool sample, a urine sample, a mucous sample, or a saliva sample. In various embodiments, performing sequencing of nucleic acids comprises performing one of targeted, exome, transcriptome, or whole genome sequencing. In various embodiments, methods further comprise generating a candidate neoantigen instance of the candidate neoantigen. In various embodiments, generating the candidate neoantigen instance comprises sampling copies of the candidate neoantigen. In various embodiments, generating one or more candidate neoantigen representations of the candidate neoantigen comprises applying one or more sub-structures of the prediction model to analyze the plurality of candidate neoantigen instances of the candidate neoantigen.

[0013] In various embodiments, the one or more sub-structures of the prediction model comprise one or more neural networks. In various embodiments, the one or more substructures of the prediction model comprise one or more multi-layer perceptrons (MLPs). In various embodiments, the one or more MLPs are two-layer MLPs. In various embodiments, the plurality of candidate neoantigen instances comprise N candidate neoantigen instances, where N is a total number of sub-structures of the prediction model. In various embodiments, the one or more candidate neoantigen representations of the candidate neoantigen comprise one or more candidate neoantigen embeddings. In various embodiments, pooling the one or more candidate neoantigen representations comprises performing a max pooling operation to downsample the one or more candidate neoantigen representations into the single individuallevel representation. In various embodiments, the generated score for the candidate neoantigen represents a measure of immunogenicity of the candidate neoantigen.

[0014] In various embodiments, the prediction model is trained by: obtaining or having obtained peptide sequences for a plurality of training neoantigens; sampling the plurality of training neoantigens to identify a set of training neoantigens representing a subset of theAttorney Docket No. N.093.W0.01plurality of training neoantigens; augmenting the set of training neoantigens to generate a plurality of training neoantigen instances; training a prediction model using the plurality of training neoantigen instances, wherein the training comprises: for each of one or more training neoantigen instances of the plurality of training neoantigen instances: generating training neoantigen representations for the training neoantigen instance; pooling the training neoantigen representations into a single individual-level representation; and generating a prediction for the training neoantigen using the single individual-level representation.

[0015] In various embodiments, training the prediction model comprises adjusting one or more parameters of the prediction model to minimize a loss function. In various embodiments, the loss function is a cross-entropy loss function. In various embodiments, the loss function represents a combination of the prediction for the training neoantigen and a ground truth label. In various embodiments, the plurality of training neoantigen instances are all associated with the ground truth label. In various embodiments, the ground truth label is an individual-level ground truth label for a training individual, wherein every training neoantigen from the individual is associated with the individual-level ground truth label. In various embodiments, each training neoantigen comprises a peptide sequence comprising between about 5 to about 50 amino acids, optionally wherein the peptide sequence comprises between about 10 to about 20 amino acids. In various embodiments, obtaining or having obtained peptide sequences for the plurality of training neoantigens comprises performing sequencing of nucleic acids from a training individual and using the nucleic acid sequences to obtain the peptide sequences for the plurality of training neoantigens. In various embodiments, obtaining sequencing data for the plurality of training neoantigens comprises performing sequencing of nucleic acids from the training individual. In various embodiments, the sequencing is next generation sequencing (NGS) or any massively parallel sequencing approach. In various embodiments, performing sequencing of nucleic acids comprises performing sequencing on tumor tissue comprising the nucleic acids from the training individual. In various embodiments, performing sequencing of nucleic acids comprises performing sequencing on nucleic acids in a sample obtained from the training individual. In various embodiments, the sample is one of a blood sample, a serum sample, a stool sample, a urine sample, a mucous sample, or a saliva sample.

[0016] In various embodiments, performing sequencing of nucleic acids comprises performing one of targeted, exome, transcriptome, or whole genome sequencing. In various embodiments, the set of training neoantigens comprises between 100 and 300 trainingAttorney Docket No. N.093.W0.01neoantigens, optionally wherein the set of training neoantigens comprises 200 training neoantigens. In various embodiments, sampling the plurality of training neoantigens comprises analyzing biochemical properties of the training neoantigens and assigning immunogenicity scores based on the analyzed biochemical properties. In various embodiments, the set of training neoantigens have assigned immunogenicity scores reflecting stronger likelihood of immunogenicity in comparison to other neoantigens not included in the set of training neoantigens. In various embodiments, augmenting the set of training neoantigens to generate a plurality of training neoantigen instances comprises: iteratively generating training neoantigen instances of the plurality, wherein each iteration generates a training neoantigen instance that samples a larger number of training neoantigens in comparison to a training neoantigen instance of a prior iteration. In various embodiments, for a / -th iteration, generating the training neoantigen instance comprises samples nj = 20 + 5j training neoantigens. In various embodiments, methods further comprise associating the plurality of training neoantigen instances with a common ground truth label.

[0017] Additionally disclosed herein is a circuit for a prediction model, the circuit comprising: one or more sub-circuitries each designed for a sub-structure for analyzing a candidate neoantigen instance and generating a candidate neoantigen representation, each of the one or more sub-circuitries comprising: a plurality of neurons comprising a plurality of microprocessors; a plurality of synaptic circuits including memories for storing synaptic weights, wherein each neuron is connected to at least one other neuron of the plurality of neurons via one of the plurality of synaptic circuits; a sub-circuitry designed for a pooling structure for pooling candidate neoantigen representations into a single individual-level representation, the sub-circuitry for the pooling structure comprising: a plurality of neurons comprising a plurality of microprocessors; a plurality of synaptic circuits including memories for storing synaptic weights, wherein each neuron is connected to at least one other neuron of the plurality of neurons via one of the plurality of synaptic circuits; and a sub-circuitry designed for a classification stmcture for generating a classification using the single individual-level representation.BRIEF DESCRIPTION OF THE DRAWINGS

[0018] These and other features, aspects, and advantages of the present invention will become better understood with regard to the following description and accompanying drawings. It is noted that wherever practicable, similar or like reference numbers may beAttorney Docket No. N.093.W0.01used in the figures and may indicate similar or like functionality. For example, a letter after a reference numeral, such as “third party entity 155A,” indicates that the text refers specifically to the element having that particular reference numeral. A reference numeral in the text without a following letter, such as “third party entity 155,” refers to any or all of the elements in the figures bearing that reference numeral (e.g. “third party entity 155” in the text refers to reference numerals “third party entity 155A” and / or “third party entity 155B” in the figures).

[0019] Figure (FIG.) 1 A depicts an overall system environment for generating a personalized prediction (e.g., for a personalized cancer vaccine), in accordance with an embodiment.

[0020] FIG. IB depicts a system environment involving a prediction system, in accordance with the embodiment shown in FIG. 1A.

[0021] FIG. 2A depicts an overall flow process for generating a personalized prediction during inference, in accordance with an embodiment.

[0022] FIG. 2B depicts generation of neoantigen instances during training, in accordance with an embodiment.

[0023] FIG. 2C further discloses training of a prediction model, in accordance with the embodiment shown in FIG. 2B.

[0024] FIG. 3 depicts an example architecture of the prediction model, in accordance with an embodiment.

[0025] FIG. 4A is a flow diagram for generating a personalized prediction for a patient, in accordance with an embodiment.

[0026] FIG. 4B is a flow diagram for training a prediction model, in accordance with an embodiment.

[0027] FIG. 5 illustrates an example computer for implementing the entities shown in FIGs.1A-1B, 2A-2C, 3, and 4A-4B.

[0028] FIG. 6 shows an exemplary diagram for classifying responders and non-responders using the methods disclosed herein.

[0029] FIG. 7 shows an exemplary ranking of candidate neoantigens.

[0030] FIG. 8A shows performance of a methodology for predicting positive candidate neoantigens.

[0031] FIG. 8B shows performance of the disclosed pipeline for predicting positive candidate neoantigens.Attorney Docket No. N.093.W0.01DETAILED DESCRIPTIONDefinitions

[0032] Terms used in the claims and specification are defined as set forth below unless otherwise specified.

[0033] The term “antigen” refers to a substance that induces an immune response.

[0034] The term “neoantigen” is an antigen that has at least one alteration that makes it distinct from the corresponding wild-type, parental antigen, e.g., via mutation in a tumor cell or post -translational modification specific to a tumor cell. A neoantigen can include a polypeptide sequence. A mutation can include a frameshift or nonframeshift indel, missense or nonsense substitution, splice site alteration, genomic rearrangement or gene fusion, or any genomic or expression alteration giving rise to a neoORF. A mutation can also include a splice variant.

[0035] The terms “subject,” “patient,” and “individual” are used interchangeably and encompass a cell, tissue, or organism, human or non-human, whether in vivo, ex vivo, or in vitro, male or female. The term “subject” is inclusive of mammals including humans. The term “mammal” encompasses both humans and non-humans and includes but is not limited to humans, non-human primates, canines, felines, murines, bovines, equines, and porcines.

[0036] The phrase “candidate neoantigen” is a mutation or other aberration giving rise to a new sequence that may represent a neoantigen.

[0037] The term “sample” can include a single cell or multiple cells or fragments of cells or an aliquot of body fluid, such as a blood sample, taken from a subject, by means including venipuncture, excretion, ejaculation, massage, biopsy, needle aspirate, lavage sample, scraping, surgical incision, or intervention or other means known in the art. Examples of an aliquot of body fluid include amniotic fluid, aqueous humor, bile, lymph, breast milk, interstitial fluid, blood, blood plasma, cerumen (earwax), Cowper’s fluid (pre-ejaculatory fluid), chyle, chyme, female ejaculate, menses, mucus, saliva, urine, vomit, tears, vaginal lubrication, sweat, serum, semen, sebum, pus, pleural fluid, cerebrospinal fluid, synovial fluid, intracellular fluid, and vitreous humour.

[0038] The term “obtaining a peptide sequence” encompasses obtaining a peptide sequence that is determined from at least one sample e.g., from an individual. Obtaining a peptide sequence encompasses obtaining a sample and processing the sample to experimentally determine the peptide sequence (e.g., by performing sequencing of nucleic acids and using the sequenced nucleic acids to determine the peptide sequence). The phrase alsoAttorney Docket No. N.093.W0.01encompasses receiving the peptide sequence, e.g., from a third party that has processed the sample to experimentally determine the peptide sequence.

[0039] The term “exome” is a subset of the genome that codes for proteins. An exome can be the collective exons of a genome.

[0040] fhe term “neural network” is a machine learning model for classification or regression consisting of multiple layers of linear transformations followed by element- wise nonlinearities typically trained via stochastic gradient descent and back-propagation.

[0041] It must be noted that, as used in the specification, the singular forms “a,” “an” and “the” include plural referents unless the context clearly dictates otherwise.

[0042] All references, issued patents and patent applications cited within the body of the specification are hereby incorporated by reference in their entirety, for all purposes.Overview

[0043] Figure (FIG.) 1 A depicts an overall system environment for generating a personalized prediction (e.g., useful for generating a personalized vaccine), in accordance with an embodiment.

[0044] FIG. 1 A introduces an individual 110. In various embodiments, the individual 110 has been diagnosed with a cancer. In various embodiments, the individual 110 may be suspected of having cancer, but may not yet have been previously diagnosed. In various embodiments, the individual is healthy and is not yet suspected of having cancer. In such embodiments, the individual may be undergoing a screen for early detection of cancer.

[0045] In various embodiments, methods include obtaining a sample from the individual 110 and performing an assay 120 to analyze the sample. In various embodiments, a sample can include any of a blood sample, a serum sample, a stool sample, a urine sample, a mucous sample, or a saliva sample. The sample can be obtained by the individual 110 or by a third party, e.g., a medical professional. Examples of medical professionals include physicians, emergency medical technicians, nurses, first responders, psychologists, phlebotomist, medical physics personnel, nurse practitioners, surgeons, dentists, and any other obvious medical professional. In various embodiments, the one or more samples can be obtained from the individual 110 by a reference lab.

[0046] As shown in FIG. 1A, methods involve performing an assay 120. In various embodiments, the assay 120 involves a sequencing assay to generate sequencing data of nucleic acids of the individual 110. For example, the sequencing assay can include any of aAttorney Docket No. N.093.W0.01targeted sequencing assay, an exome sequencing assay, a transcriptome sequencing assay, or a whole genome sequencing assay. The assay 120 processes the sample from the individual 110 to generate data (e.g., sequencing data of nucleic acids) that can be used to identify one or more candidate neoantigens of the individual 110. Further details of the assay 120 are described herein.

[0047] Following performance of the assay 120, the data generated as a result of the assay 120 is provided to the prediction system 130 for generating a personalized prediction 140. In various embodiments, the data from the assay 120 can be used to determine one or more candidate neoantigens of the individual 110. For example, the data from the assay 120 can include sequencing data of nucleic acids from the individual. One or more mutations may be present in the sequencing data, which lead to a candidate neoantigen (e.g., an antigen that has at least one mutation that makes it distinct from the corresponding wild-type, parental antigen). For example, the candidate neoantigen may include one or more substituted amino acids that correspond to the one or more mutations present in the sequencing data.

[0048] Generally, the prediction system 130 implements a prediction model, which is described in further detail herein. During inference, the prediction system 130 deploys a prediction model which analyzes the candidate neoantigens of the individual 110 and generates a personalized prediction 140, such as a prediction of whether one or more of the candidate neoantigens are likely to be immunogenic (e.g., trigger an immune response in the individual 110). During training, the prediction system 130 trains the prediction model using training neoantigens derived from training individuals with corresponding individual- level ground truth labels. Here, individual-level immunogenicity labels are more easily available (in comparison to epitope-level or neoantigen-level immunogenicity labels) and further capture aggregate immune responses observed in individuals. Thus, prediction models are trained to make informative personalized predictions, even in the absence of extensive epitope-level or neoantigen-level immunogenicity labels.

[0049] As shown in FIG. 1A, the prediction system 130 generates a personalized prediction 140. In various embodiments, the personalized prediction 140 is a predicted score. For example, the predicted score can represent a predicted level of immunogenicity of the candidate neoantigen. In various embodiments, during inference, the prediction system 130 analyzes a plurality of candidate neoantigens derived from the individual 110 and generates a personalized prediction 140 for each of the candidate neoantigens of the plurality. In various embodiments, the prediction system 130 analyzes 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15,Attorney Docket No. N.093.W0.0116, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 110, 120. 130, 140, 150, 160, 170, 180, 190, 200, 250, 300, 400, 500, 600, 700. 800, 900, 1000, or over 1000 candidate neoantigens. In various embodiments, the prediction system 130 analyzes between 2 and 1000 candidate neoantigens. In various embodiments, the prediction system 130 analyzes between 5 and 500 candidate neoantigens, between 10 and 200 candidate neoantigens, between 20 and 100 candidate neoantigens, or between 50 and 75 candidate neoantigens. Thus, the prediction system 130 can generate a personalized prediction for each of the candidate neoantigens for the individual 110.

[0050] In various embodiments, methods can further include selecting the candidate neoantigens that are predicted to have the highest level of immunogenicity in the individual. For example, selecting the candidate neoantigens predicted to have the highest level of immunogenicity in the individual can include ranking the candidate neoantigens according to their personalized predictions. In various embodiments, a personalized prediction is a score that is indicative of the immunogenicity of the candidate neoantigen. Therefore, ranking the candidate neoantigens can involve ranking the candidate neoantigens based on their assigned scores. For example, the highest ranked candidate neoantigen can have the highest assigned score, indicating a likely strong immunogenicity of the candidate neoantigen.

[0051] The selected candidate neoantigens can be expressed and / or synthesized and manufactured in a composition, such as a therapeutic vaccine. In various embodiments, methods include administering the therapeutic vaccine including one or more of the selected candidate neoantigens that are predicted to have the highest level of immunogenicity in the individual. Thus, upon administration, the therapeutic vaccine is more likely to cause an immune response in the individual, thereby treating a cancer in the individual.

[0052] FIG. IB depicts a system environment involving a prediction system 130, in accordance with the embodiment shown in FIG. 1 A. The system environment includes a prediction system 130 for at least performing one or more steps shown in FIG. 1A, and one or more third party entities 155A and 155B in communication with one another through a network 160. FIG. IB depicts one embodiment of the system environment in which two third party entities 155 A and 155B are involved. In other embodiments, additional or fewer third party entities 155 in communication with the prediction system 130 can be included. TheAttorney Docket No. N.093.W0.01third party entities 155 may communicate with the prediction system 130 to enable the prediction system 130 to perform the methods disclosed herein, e.g., for generating a personalized prediction for an individual, such as by analyzing one or more neoantigens and determining the one or more neoantigens that are likely to cause an immunogenic reaction in the individual.Third Party Entity

[0053] A third party entity 155 represents a partner entity of the prediction system 130 that can operate upstream, downstream, or both upstream and downstream of the operations of the prediction system 130. In some embodiments, the third party entity 155 may operate upstream of the prediction system 130 and provide samples obtained from individuals and / or data from samples obtained from individuals to the prediction system 130. Thus, the prediction system 130 can perform methods disclosed herein, e.g., methods for analyzing one or more neoantigens and determining the one or more neoantigens that are likely to cause an immunogenic reaction in the individual.

[0054] As another example, the third party entity 155 may process samples obtained from patients by performing one or more assays on the samples to generate data. For instance, a third party entity 155 may perform sequencing of nucleic acids of a sample obtained from an individual and provide the sequencing data to the prediction system 130. Thus, the third party entity 155 can provide the data derived from the assays to the prediction system 130 such that the prediction system 130 can perform methods disclosed herein, e.g., methods for analyzing one or more neoantigens and determining the one or more neoantigens that are likely to cause an immunogenic reaction in the individual.

[0055] In other embodiments, the third party entity 155 may operate downstream of the prediction system 130. In this scenario, the prediction system 130 may generate a personalized prediction for an individual, such as whether a neoantigen is likely to cause an immunogenic reaction in the individual. The prediction system 130 can provide an indication to the third party entity 155 that identifies one or more neoantigens that are likely to cause an immunogenic reaction in the individual. The third party entity 155 can prepare a composition, such as a vaccine comprising the one or more neoantigens, that is useful for administering to the individual for treating a cancer in the individual.

[0056] In this manner, the third party entity 155 A and the third party entity 155B may each serve distinct roles in the workflow, with one third party entity operating upstream to supplyAttorney Docket No. N.093.W0.01assay data and another third party entity operating downstream to prepare therapeutic compositions based on the predictions generated by the prediction system 130.Network

[0057] This disclosure contemplates any suitable network 160 that enables connection between the prediction system 130 and third party entities 155. The network 160 may have any combination of local area and / or wide area networks, using both wired and / or wireless communication systems. In one embodiment, the network 160 uses standard communications technologies and / or protocols. For example, the network 160 includes communication links using technologies such as Ethernet, 802.11, worldwide interoperability for microwave access (WiMAX), 3G, 4G, code division multiple access (CDMA), digital subscriber line (DSL), etc. Examples of networking protocols used for communicating via the network 160 include multiprotocol label switching (MPLS), transmission control protocol / Intemet protocol (TCP / IP), hypertext transport protocol (HTTP), simple mail transfer protocol (SMTP), and file transfer protocol (FTP). Data exchanged over the network 160 may be represented using any suitable format, such as hypertext markup language (HTML) or extensible markup language (XML). In some embodiments, all or some of the communication links of the network 160 may be encrypted using any suitable technique or techniques.Inference

[0058] FIG. 2A depicts an overall flow process for generating a personalized prediction during inference, in accordance with an embodiment. FIG. 2A may depict steps performed by the prediction system 130 described in reference to FIG. 1 A. FIG. 2A introduces a candidate neoantigen 210 that is analyzed using a prediction model 230. The inference process depicted in FIG. 2A may deploy a trained prediction model (e.g., 230) to analyze the candidate neoantigen 210 of an individual and generate a personalized prediction 140. As described herein, the candidate neoantigen 210 may include one or more substituted amino acids in comparison to a wild-type, parental antigen, where the one or more substituted amino acids correspond to one or more mutations present in sequencing data generated by performing a sequencing assay on nucleic acids of an individual.

[0059] Although FIG. 2A shows a single candidate neoantigen 210, in various embodiments during inference, there is significantly more than a single candidate neoantigen. In various embodiments, during inference, the process can include analyzing 2, 3, 4, 5, 6, 7, 8, 9, 10, 11,Attorney Docket No. N.093.W0.0112, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99. 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 250, 300, 400. 500, 600, 700, 800, 900, 1000, or over 1000 candidate neoantigens for an individual. In various embodiments, during inference, the process includes analyzing between 2 and 1000 candidate neoantigens. In various embodiments, during inference, the process includes analyzing between 5 and 500 candidate neoantigens, between 10 and 200 candidate neoantigens, between 20 and 100 candidate neoantigens, or between 50 and 75 candidate neoantigens.

[0060] As shown in FIG. 2A, a prediction model 230 is implemented to analyze the candidate neoantigen 210. In various embodiments, the peptide sequence of the candidate neoantigen 210 can be provided as input to the prediction model 230. In various embodiments, the candidate neoantigen 210 is processed to generate an input structure, also referred herein as a candidate neoantigen instance, that can be provided as input to the prediction model 230. A neoantigen instance refers to a data structure including peptide sequences of one or more neoantigens. In various embodiments, a neoantigen instance refers to a vector or matrix in which each entry of the vector or matrix includes a peptide sequence of a neoantigen. In various embodiments, generating the candidate neoantigen instance includes sampling copies of the candidate neoantigen. For example, given a candidate neoantigen, sampling copies of the candidate neoantigen involves including copies of the peptide sequence of the candidate neoantigen in the candidate neoantigen instance. In various embodiments, a neoantigen instance includes a concatenation of peptide sequences of one or more neoantigens. In various embodiments, a neoantigen instance corresponds to a particular candidate neoantigen and is embodied within a larger data structure, such as a matrix. For example, a neoantigen instance can be a row of the larger data structure. In such embodiments, each row of the larger data structure can include values (e.g., feature values) for a particular corresponding candidate neoantigen.

[0061] In some embodiments, the neoantigen instance includes the peptide sequence of only the candidate neoantigen 210 without any other neoantigen peptide sequences. Each individual candidate neoantigen can be treated as a set containing a single epitope, and the set is passed through the prediction model 230 to output a neoantigen immunogenicityAttorney Docket No. N.093.W0.01prediction. In this scenario, the prediction model 230 generates a prediction based on the neoantigen instance including the peptide sequence of only the candidate neoantigen 210.

[0062] The prediction model 230 analyzes the neoantigen instance and generates the personalized prediction 140. As described herein, the personalized prediction 140 can indicate the immunogenicity of the candidate neoantigen 210. Therefore, the personalized prediction 140 can be useful for selecting the candidate neoantigen 210 for inclusion in a therapeutic vaccine e.g., for administering to an individual.

[0063] In various embodiments, the prediction model 230 includes one or more substructures. Thus, deploying the prediction model 230 can involve applying the one or more sub-structures of the prediction model 230 to analyze one or more candidate neoantigen instances of the candidate neoantigen. In various embodiments, the candidate neoantigen instances have N candidate neoantigen instances, where N is also a total number of substructures of the prediction model 230. In general, a neoantigen instance is a data structure representing one or more neoantigen peptide sequences as an unordered set for input to the prediction model, such that the model can generate pooled patient-level or candidate-level immunogenicity predictions. In some embodiments, the prediction model 230 generates one or more candidate neoantigen representations and pools the one or more candidate neoantigen representations through a pooling operation to downsample the representations into a single individual-level representation. The prediction model 230 further analyzes the single individual-level representation to generate a prediction, an example of which is a score. The score can be indicative of the predicted immunogenicity of the candidate neoantigen. Further details of the prediction model 230, the sub-structures, and the pooling operation are described below with respect to FIG. 3.

[0064] Additionally, the personalized prediction 140 can be useful for selecting the candidate neoantigen 210 for inclusion in a therapeutic vaccine, such as for administering to an individual. As described with respect to FIG. 1 A, the candidate neoantigens that are predicted to have the highest level of immunogenicity can be ranked according to their personalized predictions and selected for expression, synthesis, and manufacture in a composition, such as a therapeutic vaccine. Thus, the inference process shown in FIG. 2A provides a mechanism by which each candidate neoantigen derived from the individual is individually assessed by the prediction model 230 to generate a corresponding personalized prediction 140 that quantifies the predicted immunogenicity of that candidate neoantigen.Attorney Docket No. N.093.W0.01Training

[0065] Methods disclosed herein involve training a prediction model 230 that can be deployed for determining a personalized prediction for an individual. As described herein, methods for training a prediction model 230 can use individual-level labels as opposed to epitope-level or neoantigen-level immunogenicity labels. An individual-level label refers to a label that indicates whether an individual experienced an immune response. An epitope-level immunogenicity label refers to a label that indicates whether an immune response was induced specifically due to a particular epitope. A neoantigen- level immunogenicity label refers to a label that indicates whether an immune response was induced specifically due to a particular neoantigen. Generally, individual-level immunogenicity labels are more easily obtainable (in comparison to epitope-level or neoantigen-level immunogenicity labels) and further capture aggregate immune responses observed in individuals. Thus, a prediction model 230 disclosed herein can be trained to make informative personalized predictions, even in the absence of extensive epitope-level or neoantigen-level immunogenicity labels.

[0066] In various embodiments, training the prediction model 230 involves a step of generating a set of high quality training data of training neoantigens associated with individual-level ground truth labels. Here, generating the set of high quality training data of training neoantigens reduces noise in the training dataset by filtering out weak or non-immunogenic epitopes and / or neoantigens that would otherwise lead to a poorly trained model. In various embodiments, the high quality training data of training neoantigens are further augmented. By augmenting the training dataset, this enables the prediction model 230 to learn more robust representations of training neoantigens across different dataset sizes. This leads to further improvements to the performance of the prediction model 230.

[0067] Reference is now made to FIG. 2B, which depicts generation of neoantigen instances during training, in accordance with an embodiment. Reference will further be made to FIG.2C, which further discloses training of a prediction model 230, in accordance with the embodiment shown in FIG. 2B. Ilere, the neoantigen instances 235 refer to the augmented training dataset that are provided to train the prediction model 230.

[0068] FIG. 2B illustrates a multi-stage data preparation pipeline, e.g., preparing training data for the prediction model 230 in FIG. 2A. The training process may begin with an initial set of neoantigens 220, proceed through a sampling module 225 to produce a refined set of neoantigens 226, and further proceed through an augmentation module 228 to generate neoantigen instances 235 for training the prediction model. The pipeline depicted in FIG. 2BAttorney Docket No. N.093.W0.01addresses the challenge of noise in neoantigen datasets by first filtering the initial set of neoantigens to retain the neoantigens with stronger predicted immunogenic potential, and then augmenting the filtered set to produce training instances of varying sizes that enable the prediction model to learn robust representations across different set sizes.

[0069] In some embodiments, FIG. 2B begins with a set of neoantigens 220 (e.g., neoantigen 220A...neoantigen 220N...neoantigen 220X). In various embodiments, the set of neoantigens 220 may include 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60. 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 150, 200, 250, 300, 350, 400, 450, 500, 600, 700, 800, 900, 1000, 2000, 3000, 4000, 5000, 6000, 7000, 8000, 9000, 10,000, 20,000, 30,000, 40,000, 50,000, 60,000, 70,000, 80,000, 90,000, 100,000, or more than 100,000 neoantigens. In various embodiments, the set of neoantigens 220 includes between 2 and 100,000 neoantigens. In various embodiments, the set of neoantigens 220 includes between 100 and 50,000 neoantigens, between 500 and 10,000 neoantigens, between 1,000 and 5,000 neoantigens, or between 2,000 and 4,000 neoantigens. Here, these neoantigens 220 used for training may be referred to herein as training neoantigens.

[0070] In various embodiments, these neoantigens 220 are derived from individuals, also referred to herein as training individuals. In various embodiments, the peptide sequences of the neoantigens 220 (e.g., training neoantigens) are obtained by performing sequencing of nucleic acids from one or more training individuals. The nucleic acid sequences can then be used to obtain the peptide sequences for neoantigens 220 (e.g., training neoantigens). In various embodiments, the sequencing includes next generation sequencing (NGS) or any massively parallel sequencing approach. In various embodiments, the sequencing is performed on tumor tissue comprising the nucleic acids from the training individuals. In various embodiments, the sequencing is performed on nucleic acids in samples obtained from the training individuals.

[0071] FIG. 2B introduces a sampling module 225, which performs a sampling step to generate a high-quality training dataset including a reduced number of neoantigens 226. In various embodiments, the sampling module 225 samples the initial set of neoantigens 220 and analyzes their biochemical properties to determine which neoantigens are most likely to be immunogenic. In various embodiments, the sampling module 225 applies a trainedAttorney Docket No. N.093.W0.01machine learning model that predicts neoantigens that are most likely to be immunogenic. A trained machine learning model can analyze sequences of neoantigens and consider likely biochemical properties to determine which neoantigens are likely to be immunogenic.Example trained machine learning models include NetMHCpan and NetMHCStab.NetMHCpan is described in further detail in Hoof, 1., et al., NetMHCpan, a method for MHC class I binding prediction beyond humans. Immunogenetics 61, 1-13 (2009), which is incorporated by reference in its entirety. NetMHCStab is described in further detail in Jorgensen, K.W., et al. (2014), NetMHCstab - predicting stability of peptide-MHC-I complexes: impacts for cytotoxic T lymphocyte epitope discovery. Immunology, 141: 18-26, which is incorporated by reference in its entirety. A further example trained machine learning model can be trained using mass spectrometry data and used to predict likely immunogenic neoantigens. An example trained machine learning model trained using mass spectrometry data is described in WO2017106638, titled NEO ANTIGEN IDENTIFICATION, MANUFACTURE, AND USE, filed December 16, 2016, which is incorporated by reference in its entirety. Additional example trained machine learning models can be trained using in vivo immune response data (e.g., in vivo responses to immunotherapy treatment) which can be more informative for neoantigen prioritization compared to other e.g., in vitro approaches.

[0072] In various embodiments, the sampling module 225 assigns scores to the neoantigens that reflect the likely immunogenicity of the neoantigens. The sampling module 225 samples the initial set of neoantigens 220 and selects a threshold number of neoantigens that are most likely to be immunogenic. In various embodiments, the sampling module 225 selects the top X neoantigens that are most likely to be immunogenic, where a value of X is any one of 5, 10, 15, 20, 25, 30, 35, 40, 45, 50, 55, 60, 65, 70, 75, 80, 85, 90, 95, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 225, 250, 275, 300, 350, 400, 450, 500, 550, 600, 650, 700, 750, 800, 850, 900, 950, or 1000. In various embodiments, the value of Xis between 5 and 500. In various embodiments, the value of X is between 50 and 400, between 100 and 300, between 150 and 250, between 175 and 225, or between 190 and 210. In particular embodiments, the value of X is 200. Here, the top X neoantigens represent the neoantigens that are most likely to be immunogenic and therefore, the set of top X neoantigens excludes less immunogenic or non-immunogenic neoantigens. In various embodiments, the selected set of training neoantigens each have assigned immunogenicity scores that reflect strongerAttorney Docket No. N.093.W0.01likelihood of immunogenicity in comparison to other neoantigens not included in the selected set of training neoantigens.

[0073] In various embodiments, the sampling module 225 may perform a training-time signal-enrichment operation prior to input of neoantigen data into the prediction model 230. Because a large fraction of candidate neoantigens associated with a patient may be weakly immunogenic or non-immunogenic, direct use of the full neoantigen set can dilute the patient-level learning signal and increase computational burden. By assigning immunogenicity-related scores and selecting a refined subset of neoantigens, the sampling module 225 reduces input-set noise, decreases the dimensional complexity of the training input, and concentrates model training on neoantigens more likely to contain informative response-associated features.

[0074] In various embodiments, the sampling operation functions as a front-end filtering stage that is distinct from the prediction model 230 and that improves downstream machinelearning performance by removing low-information candidates before representation learning is performed by the shared sub-structures of the prediction model 230 (as described below with respect to FIG. 3). This staged arrangement can improve training efficiency and improve convergence behavior relative to approaches that indiscriminately process all candidate neoantigens for each training individual.

[0075] As shown in FIG. 2B, the high-quality training dataset includes the top X neoantigens of neoantigen 226A and neoantigen 226N. Neoantigen 226A and neoantigen 226N represent a subset of the initial set of neoantigens 220 selected by the sampling module 225 based on their predicted immunogenicity. The refined set of neoantigens 226, including neoantigen 226A through neoantigen 226N, forms a candidate pool of training neoantigens from which the subsequent augmentation step draws samples.

[0076] FIG. 2B introduces an augmentation module 228, which augments the high-quality training dataset to generate neoantigen instances of different sizes. Each of these neoantigen instances (e.g., neoantigen instance 235) can be associated with a common ground truth label. For example, if all of the initial set of neoantigens 220 are from one or more individuals who experienced an immune reaction, then the common ground truth label can be an indication of immunogenicity. As another example, if all of the initial set of neoantigens 220 are from one or more individuals who did not experience an immune reaction, then the common ground truth label can be an indication of a lack of immunogenicity.Attorney Docket No. N.093.W0.01

[0077] In various embodiments, a label can be a binary value. Therefore, a first binary value can be indicative of an immune response, whereas a second binary value can be indicative of a lack of an immune response. As a specific example, a binary value of “1” can be indicative of an immune response and a binary value of “0” can be indicative of a lack of an immune response. In various embodiments, a label can be a continuous value. For example, the label can be a continuous value between two numbers (e.g., between 0 and 1), where a first number is indicative of an immune response and a second number is indicative of lack of an immune response. Thus, a continuous value closer to the first number can be reflective of an immune response whereas a continuous value closer to the second number can be reflective of a lack of an immune response.

[0078] In various embodiments, the augmentation module 228 augments the neoantigens 226 to generate training neoantigen instances through an iterative process. In various embodiments, each iteration generates a training neoantigen instance that samples a different number of training neoantigens in comparison to a training neoantigen instance of a prior iteration. By doing so, this creates training instances of different sizes such that the prediction model 230 can learn more robust representations across the different sizes. In various embodiments, each iteration generates a training neoantigen instance that samples a larger number of training neoantigens in comparison to a training neoantigen instance of a prior iteration. In various embodiments, the augmentation module 228 generates two training neoantigen instances. In various embodiments, the augmentation module 228 generates three training neoantigen instances, four training neoantigen instances, five training neoantigen instances, six training neoantigen instances, seven training neoantigen instances, eight training neoantigen instances, nine training neoantigen instances, or ten training neoantigen instances. In particular embodiments, the augmentation module 228 generates five training neoantigen instances by sampling different quantities of training neoantigens. For example, the augmentation module 228 may sample 20, 25, 30, 35, and 40 training neoantigens to generate five neoantigen training instances.

[0079] In various embodiments, for a j-th iteration, the augmentation module 228 generates a neoantigen instance 235 which includes Tij = A + B * j training neoantigens sampled from the high-quality dataset, where A and B are positive constant values. In various embodiments, A is a value of 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, or 50. In various embodiments, A is between 1 and 50. In variousAttorney Docket No. N.093.W0.01embodiments, the value of A is between 5 and 40, between 10 and 30, between 15 and 25, between 18 and 22, or between 19 and 21. In various embodiments, B is a value of 1, 2, 3, 4, 5, 6, 7, 8, 9, or 10. In particular embodiments, A is a value of 20 and B is a value of 5.

[0080] By creating multiple views of the same patient with different neoantigen subsets, the augmented training set represented by neoantigen instances 235 enables the prediction model to learn more robust representations across different set sizes. The usage of immunogenicity scores from the sampling module 225 for the initial ranking ensures that neoantigens with stronger predicted immunogenic potential are prioritized, while random sampling within the refined set of neoantigens 226 maintains diversity across the augmented training instances. The neoantigen instances 235 generated by the augmentation module 228 are subsequently provided as input to the prediction model 230 for training, as further described with respect to FIG. 2C.

[0081] In various embodiments, the augmentation module 228 may generate multiple stochastic views of the same training individual by repeatedly sampling different neoantigen subsets from the refined set of neoantigens 226 while maintaining the same individual-level ground truth label. Each generated training neoantigen instance therefore constitutes a different set-valued representation of the same underlying patient immune-response state. Training on these multiple views can regularize the prediction model 230 against overfitting to any single subset configuration and can improve robustness to variation in neoantigen-set composition during inference.

[0082] In various embodiments, the iterative increase in the number of sampled neoantigens across augmentation iterations trains the prediction model 230 to operate across neoantigen sets of different cardinalities. This is technically beneficial because patients may present substantially different tumor mutational burdens and, correspondingly, different numbers of candidate neoantigens. Exposure during training to multiple set sizes allows the prediction model 230 to learn cardinality-tolerant patient-level representations while preserving permutation invariance with respect to the order of neoantigens within each set. In some embodiments, the augmentation process also improves data efficiency by reusing patientlevel supervision across multiple distinct training neoantigen instances derived from the same training individual without requiring additional experimentally generated epitope-level labels.

[0083] FIG. 2C is a schematic diagram depicting training of the prediction model 230 using neoantigen instances 235 and ground truth labels across multiple training iterations 255, in accordance with the embodiment shown in FIG. 2B. As depicted, the neoantigen instancesAttorney Docket No. N.093.W0.01235 are provided as input to the prediction model 230 across multiple training iterations 255. At each training iteration, the prediction model 230 analyzes a neoantigen instance 235 and outputs a prediction 250. The prediction 250 is combined with a ground truth label 260 to determine a loss value 270 which is used to further train (e.g., adjuster the parameters of) the prediction model 230. The prediction 250 represents the output generated by the prediction model 230 for a given neoantigen instance during a given training iteration. In various embodiments, the prediction 250 is a score that can be combined with a value of the label 260.

[0084] In various embodiments, a loss function is employed when combining the prediction 250 and the ground truth label 260, and the loss value 270 of the loss function represents a penalty. The loss value 270 can be back-propagated, and the weights, parameters, and / or biases of the prediction model 230 are adjusted to minimize the loss value over subsequent training iterations. Example loss functions include a regression loss function, a mean squared error (MSE) loss function, a mean absolute error (MAE) loss function, a Huber loss function, a Log-Cosh loss function, or a cross-entropy loss function (e.g., a binary, categorical, or sparse loss function). In particular embodiments, the loss function is a cross-entropy loss function. An example cross entropy loss function can be expressed as:L(f(.Pi),yi) = ~[yt log log (f(pi')) + (1 - y;) log log (1 - )] where f(pi) is the model’s predicted immune response for patient Pi, and y; is the corresponding ground truth label.

[0085] As shown in FIG. 2C, at each training iteration of the training iterations 255, a respective neoantigen instance from the neoantigen instances 235 is selected and fed to the prediction model 230 to generate a prediction 250. The prediction 250 is compared against the label 260 to compute the loss value 270, and the parameters of the prediction model 230 are adjusted based on the loss value 270. This iterative process continues until the prediction model 230 has been trained across multiple neoantigen instances 235. Since the neoantigen instances 235 include neoantigen instances of different sizes as generated by the augmentation module 228 as shown in FIG. 2B, the prediction model 230 is exposed to neoantigen sets of varying cardinality during training, which enables the prediction model 230 to learn robust representations that generalize across different set sizes. Upon completion of the training iterations, the prediction model 230 can be deployed during inference to analyze candidate neoantigens of an individual.Attorney Docket No. N.093.W0.01Prediction Model

[0086] As disclosed herein, prediction models can be trained and deployed for determining a personalized prediction for an individual. In various embodiments, a prediction model can include any one of a regression model (e.g., linear regression, logistic regression, or polynomial regression), decision tree, random forest, support vector machine, Naive Bayes model, k-means cluster, or neural network (e.g., feed- forward networks, convolutional neural networks (CNN), deep neural networks (DNN), autoencoder neural networks, generative adversarial networks, or recurrent networks (e.g., long short-term memory networks (LSTM), bi-directional recurrent networks, deep bi-directional recurrent networks).

[0087] The prediction model can be trained using a machine learning implemented method, such as any one of a linear regression algorithm, logistic regression algorithm, decision tree algorithm, support vector machine classification, Naive Bayes classification, K-Nearest Neighbor classification, random forest algorithm, deep learning algorithm, gradient boosting algorithm, and dimensionality reduction techniques such as manifold learning, principal component analysis, factor analysis, autoencoder regularization, and independent component analysis, or combinations thereof. In various embodiments, the prediction model is trained using supervised learning algorithms, unsupervised learning algorithms, semi-supervised learning algorithms (e.g., partial supervision), weak supervision, transfer, multi-task learning, or any combination thereof.

[0088] In various embodiments, the prediction model has one or more parameters, such as hyperparameters or model parameters. Hyperparameters are generally established prior to training. Examples of hyperparameters include the learning rate, depth or leaves of a decision tree, number of hidden layers in a deep neural network, number of clusters in a k-means cluster, penalty in a regression model, and a regularization parameter associated with a cost function. Model parameters are generally adjusted during training. Examples of model parameters include weights associated with nodes in layers of neural network, support vectors in a support vector machine, and coefficients in a regression model. The model parameters of the prediction model are trained (e.g., adjusted) using the training data to improve the predictive power of the prediction model.

[0089] FIG. 3 depicts an example architecture of the prediction model, in accordance with an embodiment. The prediction model 230 may be configured to receive, as input, a neoantigen instance 335. In various embodiments, a neoantigen instance includes peptide sequences of one or more neoantigens. During inference, the neoantigen instance includesAttomey Docket No. N.093.W0.01peptide sequences of one or more candidate neoantigens from an individual. During training, the neoantigen instance includes peptide sequences of one or more training neoantigens. In various embodiments, the prediction model 230 may be deployed to analyze candidate neoantigens of an individual during inference and to process training neoantigen instances for parameter adjustment during training.

[0090] For example, as shown in FIG. 3, neoantigen instance 335 may include peptide sequences of N different neoantigens e.g., including peptide sequence of neoantigen 310A and up to peptide sequence of neoantigen 3 ION. In various embodiments, N can be any of 1, 2, 3, 4, 5, 6, 7, 8. 9, 10, 11, 12, 13. 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104. 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, or 200 different neoantigens. In various embodiments, the N is between 1 and 200. In various embodiments, the value of N is between 5 and 150, between 10 and 100, between 20 and 80, between 30 and 70, or between 40 and 60.

[0091] The prediction model 230 may include one or more sub-structures that generate neoantigen representations, a pooling structure that pools the neoantigen representations into a single individual-level representation, and a classification structure that generates the prediction 250 from the single individual-level representation. In various embodiments, the sub-structures may share one or more parameters such that the same learnable transformation is applied across different neoantigen positions within a neoantigen instance. For example, the prediction model 230 can be trained such that the shared sub-structures learn a common element-wise transformation for individual neoantigens, while the pooling structure and classification structure learn how to aggregate those transformed neoantigen features into a patient-level immune-response prediction. Because the same learnable transformation is applied across different neoantigen positions, the number of learned parameters need not scale with the number of neoantigens in a training neoantigen instance. This shared-Attorney Docket No. N.093.W0.01parameter training configuration improves data efficiency, supports permutation invariance, and avoids parameter growth with set size.

[0092] In various embodiments, the pooling operation may be selected to enable the prediction model 230 to detect response-associated signal arising from a sparse subset of neoantigens within a larger candidate set. For example, where only one or a few neoantigens in a patient are truly immunogenic, a max pooling operation can preserve one or more salient activation values generated by the shared sub-structures and propagate those values into the individual-level representation used for patient- level classification. This training behavior is particularly useful in neoantigen datasets in which immune responses are sparse relative to the total number of candidate neoantigens.

[0093] In various embodiments, training with an individual-level ground truth label can cause the prediction model 230 to leam whether the training neoantigen instance, considered as a set, is associated with an immune response, rather than requiring supervision identifying which specific neoantigen within the set is immunogenic. This training paradigm allows the prediction model 230 to exploit clinically available responder / non-responder information while still learning set-level feature patterns that are later usable for candidate-neoantigen prioritization during inference.

[0094] Referring back to FIG. 3, the prediction model 230 can include one or more substructures configured to analyze peptide sequences of neoantigens of the neoantigen instance. In various embodiments, the prediction model 230 includes N different sub-structures configured to analyze peptide sequences of the N neoantigens of the neoantigen instance. As shown in FIG. 3, each sub-structure 312 analyzes a peptide sequence of a particular neoantigen 310. In various embodiments, the one or more sub-structures 312 include one or more machine learning models. In various embodiments, each sub-structure 312 is a machine learning model. In various embodiments, the one or more sub-structures 312 include one or more neural networks. In various embodiments, each sub-structure 312 is a neural network. In various embodiments, the one or more sub-structures 312 have one or more multi-layer perceptrons (MLPs). In various embodiments, each sub-structure 312 is a MLP. In various embodiments, each MLP is a two-layer MLP.

[0095] Although FIG. 3 shows individual substructures 312 (e.g., sub-structure 312A to substructure 312N), in various embodiments, the different sub-structures can be embodied as a single machine learning model (e.g., a single neural network). In various embodiments, the different sub-structures may share one or more parameters and therefore, the different subAttorney Docket No. N.093.W0.01structures can be jointly trained by adjusting the shared one or more parameters during training.

[0096] The output of each sub-structure 312 is a neoantigen representation 315. Here, a neoantigen representation refers to a transformed representation of a peptide sequence of a neoantigen. For example, if the sub-structure 312 is a machine learning model, then the substructure 312 transforms the peptide sequence of the neoantigen 310 to generate a transformed neoantigen representation 315. In various embodiments, the transformation is a learnable function < >: Rd-> Rk. In various embodiments, the neoantigen representation is a neoantigen embedding (e.g., a representation of discrete data as a vector).

[0097] The neoantigen representations 315 outputted by the sub-structures 312 are dimensionally reduced. In various embodiments, the dimensional reduction has a pooling operation. As shown in FIG. 3, the prediction model 230 can include a pooling structure 320 that performs a pooling operation to downsample the neoantigen representations 315. The output of the pooling structure 320 is a single individual-level representation 325. Here, the single individual-level representation 325 is a dimensionally reduced representation in comparison to the neoantigen representations 315. In various embodiments, the pooling structure 320 performs a max pooling operation.

[0098] The individual-level representation 325 is provided as input into a classification structure 330 that generates a prediction 350. In various embodiments, the prediction 350 shown in FIG. 3 is the personalized prediction described herein (e.g., personalized prediction 140 shown in FIG. 1 A). In various embodiments, the classification structure 330 is a machine learning model. In various embodiments, the classification structure 330 is a neural network or a layer of a neural network. In various embodiments, the classification structure 330 applies a learnable function p: Rk-> R, which processes the single individual-level representation 325 and produces a final patient-level prediction e.g., prediction 350.

[0099] In various embodiments, the prediction model employs a permutation-invariant deep learning architecture. A permutation-invariant deep learning architecture can be well-suited for modeling unordered nature of neoantigen sets. An example of a permutation-invariant model is DeepSets. Further details regarding an example DeepSets model is described in Zaheer, M., et al., (2017). Deep sets. Advances in neural information processing systems, 30, which is incorporated by reference in its entirety.

[0100] Upon completion of training, the prediction model 230 has learned a set-processing function that maps a variable-cardinality neoantigen set to a patient-level response-associatedAttorney Docket No. N.093.W0.01representation. Since the trained function is permutation-invariant and tolerant of different set sizes, the same trained prediction model 230 may be deployed during inference using a set containing a single candidate neoantigen, thereby converting patient-level supervision obtained during training into candidate-neoantigen-level scoring during inference.

[0101] Accordingly, the disclosed training framework does not merely train a generic classifier, but instead trains a set-based prediction architecture that is structurally configured to (i) suppress noise through preselection of training neoantigens, (ii) improve cardinality robustness through augmentation, (iii) learn shared neoantigen representations through parameter sharing, and (iv) aggregate sparse response-associated features into a patient-level prediction through pooling. The resulting trained model is particularly suited for neoantigen prioritization in personalized cancer vaccine workflows.

[0102] In various embodiments, the prediction model disclosed herein can be embodied in the fomr of a physical circuit. For example, each of the sub-structure 312, pooling structure 320, and classification structure 330 can be embodied as physical circuitry for performing the functions described herein. In various embodiments, each of the sub-structure 312, pooling structure 320, and classification structure 330 perform the functionality of machine learning models, such as neural networks. Thus, the circuitry of the prediction model can include neurons arranged in layers that are interconnected through synaptic circuits. The parameters of the neural networks can be learned and stored in memory as synaptic weights.

[0103] Disclosed herein is a circuit for a prediction model, the circuit includes one or more sub-circuitries each designed for a sub-structure for analyzing a candidate neoantigen instance and generating a candidate neoantigen representation. In various embodiments, each of the one or more sub-circuitries may include a plurality of neurons, a plurality of synaptic circuits, a sub-circuitry designed for a pooling structure, and a sub-circuitry designed for a classification structure. The neurons include a plurality of microprocessors. The synaptic circuits include memories for storing synaptic weights, where each neuron is connected to at least one other neuron of the neurons via one of the synaptic circuits.

[0104] In various embodiments, the sub-circuitry designed for a pooling structure can be used to pool candidate neoantigen representations into a single individual-level representation. The sub-circuitry for the pooling structure may include a plurality of neurons comprising a plurality of microprocessors, and a plurality of synaptic circuits including memories for storing synaptic weights. Each neuron is connected to at least one other neuron of the plurality of neurons via one of the plurality of synaptic circuits. In variousAttorney Docket No. N.093.W0.01embodiments, a sub-circuitry designed for a classification structure can be used to generate a classification using the single individual-level representation.

[0105] In some implementations, the disclosed framework provides a technical improvement in computerized analysis of biological sequence data by enabling a machinelearning system to process patient-specific neoantigen collections as unordered sets having variable cardinalities. Conventional order-dependent or fixed-length input models may require artificial padding, truncation, or positional encoding that does not reflect the biological structure of neoantigen collections. By contrast, the disclosed shared-sub-structure and pooling architecture converts variable-cardinality neoantigen inputs into a fixed-length individual-level representation for downstream computation, thereby improving machine processing of neoantigen data structures and improving robustness across patients having different numbers of candidate neoantigens.Example Flow Process for Generating a Personalized Prediction

[0106] FIG. 4A is a flow diagram for generating a personalized prediction for a patient, in accordance with an embodiment. Specifically, FIG. 4A shows the process during inference, which includes deploying a trained prediction model to analyze one or more candidate ncoantigcns derived from an individual and to generate a personalized prediction indicative of the immunogenicity of each candidate neoantigen.

[0107] Step 410 involves obtaining sequencing data of a candidate neoantigen for an individual. For example, the step 410 can include performing sequencing of nucleic acids from the individual and using the nucleic acid sequences to obtain a peptide sequence of the candidate neoantigen. The sequencing can include next generation sequencing (NGS) or any massively parallel sequencing approach.

[0108] Step 420 involves analyzing the candidate neoantigen using a prediction model. As shown in FIG. 4A, step 420 involves substeps 425, 430, and 440.

[0109] Step 425 involves generating one or more candidate neoantigen representations for the candidate neoantigen. In various embodiments, step 425 involves generating the one or more candidate neoantigen representations using a sub-structure of a prediction model, such as a neural network structure of the prediction model. In various embodiments, the substructure may include one or more MLPs. In some embodiments, each MLP may be a two-layer MLP. In some embodiments, the one or more candidate neoantigen representations may include one or more candidate neoantigen embeddings, where each embedding is aAttorney Docket No. N.093.W0.01transformed representation of the peptide sequence of the candidate neoantigen produced by a learnable function applied by the sub-structure.

[0110] Step 430 involves pooling the one or more candidate neoantigen representations into a single individual-level representation. Here, the pooling operator can perform a dimensional reduction of the candidate neoantigen representations into a single individuallevel representation. In various embodiments, the pooling operation can include performing a max pooling operation to downsample the one or more candidate neoantigen representations into the single individual-level representation. The resulting single individual-level representation may constitute a fixed-length patient- level feature vector derived from the pooled candidate neoantigen representations.

[0111] Step 440 involves generating a prediction for the individual using the single individual-level representation. In various embodiments, the prediction for the individual is a measure of the immunogenicity of the candidate neoantigen. In various embodiments, the prediction for the individual is a score that is indicative of the immunogenicity of the candidate neoantigen.

[0112] In various embodiments, if the prediction for the individual indicates that the candidate neoantigen is likely immunogenic, the candidate neoantigen can be selected e.g., for inclusion in a vaccine for administration to the individual. In various embodiments, inference is performed for each of candidate neoantigens for the individual by repeating the steps 410-440, and the resulting scores may be used to rank the candidate neoantigens according to their predicted immunogenicity. Thus, in such embodiments, the vaccine comprising the candidate neoantigen can be therapeutically administered for stimulating an immune response and treating a cancer in the individual.Example Flow Process for Training a Prediction Model

[0113] FIG. 4B is a flow diagram for training a prediction model, in accordance with an embodiment. It is noted that the training of a prediction model disclosed herein can train a set-based architecture on noisy, sparse, patient-derived neoantigen collections using signal-enriched sampling and multi-view augmentation so that patient-level labels can later drive neoantigen- level ranking.

[0114] Step 455 involves obtaining sequencing data of a plurality of training neoantigens, e.g., using the techniques described above.Attorney Docket No. N.093.W0.01

[0115] Step 460 involves sampling the plurality of training neoantigens to identify a set of candidate training neoantigens. In various embodiments, step 460 involves analyzing the plurality of training neoantigens based on their biochemical properties to evaluate their likely immunogenicity. The set of candidate training neoantigens represents a refined set of neoantigens for training the prediction model. This sampling process at step 460 helps reduce noise in the dataset by filtering out weak or non-immunogenic neoantigens.

[0116] Step 465 may be optional as indicated by the dotted lines shown in FIG. 4B. In other words, in some embodiments the process for training a prediction model shown in FIG. 4B includes step 465, whereas in other embodiments the process shown in FIG. 4B omits step 465. Step 465 involves augmenting the set of candidate training neoantigens to generate a plurality of training neoantigen instances. In various embodiments, step 465 involves iteratively generating candidate training neoantigen instances by sampling the pool of candidate training neoantigens in ever-increasing numbers. In particular embodiments, step 465 involves iteratively generating candidate training neoantigen instances, where each iteration generates a candidate training neoantigen instance that samples a larger number of candidate training neoantigens in comparison to a candidate training neoantigen instance of a prior iteration. For example, step 465 may generate five training neoantigen instances by sampling 20, 25, 30, 35, and 40 training neoantigens from the candidate pool. In some embodiments, for a j-th iteration, step 465 generates a training neoantigen instance that includes A+Bx / training neoantigens sampled from the candidate pool, where A and B are positive constant values. Each augmented training neoantigen instance generated at step 465 is associated with a common ground truth label inherited from the training individual. By creating multiple views of the same patient with different neoantigen subsets, the augmented training set enables the prediction model to learn robust representations across different set sizes.

[0117] In various embodiments, the training neoantigen instances generated at step 465 are distinct set- valued training examples corresponding to different sampled views of the same patient-derived neoantigen pool. By associating each of these distinct training neoantigen instances with the same individual-level ground truth label, the disclosed method trains the prediction model to recognize patient-level immune-response patterns that persist across multiple neoantigen subset compositions.

[0118] Step 470 involves training a prediction model using the plurality of training neoantigen instances. As shown in FIG. 4B, step 470 can involve substeps 475, 480 and 485.Attorney Docket No. N.093.W0.01

[0119] Step 475 involves generating a candidate training neoantigen representation for a training neoantigen instance. Step 475 can be repeated for each training neoantigen instance. In various embodiments, step 475 involves analyzing each training neoantigen instance using a sub-structure of a prediction model, such as a neural network structure of the prediction model. In some embodiments, each sub-structure may include one or more MLPs. The substructures can be embodied as a single machine learning model with shared parameters that are jointly trained across different neoantigen positions, consistent with the permutationinvariant architecture described with respect to FIG. 3.

[0120] In various embodiments, the sub-structures used at step 475 share one or more parameters across different neoantigen positions within the training neoantigen instance. During training, backpropagation adjusts the shared parameters using gradients contributed by multiple neoantigens from multiple training neoantigen instances, thereby learning a common sequence-to-representation transformation that is reusable across unordered neoantigen sets.

[0121] Step 480 involves pooling candidate training neoantigen representations into a single individual-level representation. Here, the pooling operation can perform a dimensional reduction of the candidate training neoantigen representations into a single individual level representation. In some embodiments, the pooling operation comprises performing a max pooling operation to aggregate the candidate training neoantigen representations into a fixed-length patient-level feature vector. The resulting single individual-level representation captures aggregated information from across the neoantigen representations within the training neoantigen instance.

[0122] In various embodiments, the pooling operation at step 480 generates a fixed-length individual-level representation regardless of the number of neoantigens included in the training neoantigen instance. This allows the classification stage to operate on a consistent representation space even when the upstream neoantigen set cardinality varies from one training neoantigen instance to another.

[0123] Step 485 involves generating a prediction for the training neoantigen using the single individual-level representation. The prediction model is further trained using the generated score. For example, the parameters of the prediction model are adjusted to minimize a loss function. Such a loss function may involve a combination of the prediction for the candidate training neoantigen and a ground truth label. In some embodiments, the loss function is a cross-entropy loss function. The ground truth label is an individual-level ground truth labelAttorney Docket No. N.093.W0.01for a training individual, and each training neoantigen from that individual is associated with the individual-level ground truth label. The loss value computed from the combination of the prediction and the ground truth label is back-propagated, and the weights, parameters, and / or biases of the prediction model are adjusted to minimize the loss value over subsequent training iterations.

[0124] In various embodiments, the prediction generated at step 485 reflects whether a particular training neoantigen instance, as an unordered set of neoantigens, is associated with an immune response for the corresponding training individual. Repeated optimization across augmented training neoantigen instances causes the prediction model to learn to identify response-associated features that are reproducible across different subset samplings of the patient-derived neoantigen pool.

[0125] Upon completion of the training process depicted in FIG. 4B, the trained prediction model may be deployed during inference to analyze candidate neoantigens of an individual as shown in FIG. 4A.Assays

[0126] Methods disclosed herein involve performing one or more assays (e.g., assay 120 described in FIG. 1 A) for generating data (e.g., sequencing data) that can be analyzed by the prediction system 130 to generate a personalized prediction 140. In particular embodiments, the one or more assays include a sequencing assay. The sequencing assay can include any of a targeted sequencing assay, an exome sequencing assay, a transcriptome sequencing assay, or a whole genome sequencing assay. A targeted sequencing assay refers to a sequencing assay that amplifies target sequences of interest and sequences amplicons derived from the target sequences of interest. An exome sequencing assay refers to a sequencing assay that generates sequencing data of nucleic acids from exomes. A transcriptome sequencing assay refers to a sequencing assay that generates sequencing data of the transcriptome (e.g., expressed RNA transcripts including coding a non-coding transcripts). A whole genome sequencing assay refers to a sequencing assay that generates sequencing data across the whole genome.

[0127] In various embodiments, performing the assay includes performing nucleic acid amplification. Examples of such assays include, but are not limited to performing PCR assays, Real-time PCR assays, Quantitative real-time PCR (qPCR) assays, digital PCR (dPCR), Allele-specific PCR assays, Reverse-transcription PCR assays and reporter assays.Attorney Docket No. N.093.W0.01For example, a PCR assay can be performed to amplify target sequences to generate amplicons of the target sequences. PCR primers are added to initiate the amplification. In various embodiments, the PCR primers are whole genome primers that enable whole genome amplification. In various embodiments, the PCR primers are gene-specific primers that result in amplification of sequences of specific genes. In various embodiments, the PCR primers are allele-specific primers.

[0128] Various different sequencing methods can be implemented for sequencing nucleic acids (either DNA or RNA). For example, for DNA sequencing any one of whole genome sequencing, whole exome sequencing, or a targeted panel sequencing can be conducted. For RNA, RNA-seq (RNA Sequencing), an example of which includes Whole Transcriptome Shotgun Sequencing (WTSS), is a technology that utilizes the capabilities of next-generation sequencing to reveal a snapshot of RNA presence and quantity from a genome at a given moment in time. The transcriptome of a cell is dynamic; it continually changes as opposed to a static genome. The recent developments of Next-Generation Sequencing (NGS) allow for increased base coverage of a DNA sequence, as well as higher sample throughput. This facilitates sequencing of the RNA transcripts in a cell, providing the ability to look at alternative gene spliced transcripts, post-transcriptional changes, gene fusion, mutations / SNPs and changes in gene expression. In addition to mRNA transcripts, RNA-Seq can look at different populations of RNA to include total RNA, small RNA, such as miRNA, tRNA, and ribosomal profiling. RNA-Seq can also be used to determine exon / intron boundaries and verify or amend previously annotated 5' and 3' gene boundaries, Ongoing RNA-Seq research includes observing cellular pathway alterations during infection, and gene expression level changes in cancer studies. Prior to NGS, transcriptomics and gene expression studies were previously done with expression microarrays, which contain thousands of DNA sequences that probe for a match in the target sequence, making available a profile of all transcripts being expressed.Neoantigens

[0129] Disclosed herein are methods for determining a personalized prediction for an individual, e.g., by analyzing one or more neoantigens and determining the one or more neoantigens are likely to cause an immunogenic reaction in the individual. In various embodiments, neoantigens can include polypeptides and therefore, neoantigens including polypeptide sequences may be useful in vaccines for administering to the individual. InAttorney Docket No. N.093.W0.01various embodiments, one or more neoantigens can be immunogenic in an individual having a tumor, and therefore, the one or more neoantigens may be capable of eliciting a T cell response or a B cell response in the individual.

[0130] In various embodiments, a neoantigen has about 5, about 6, about 7, about 8, about 9, about 10, about 11, about 12, about 13, about 14, about 15, about 16, about 17, about 18, about 19, about 20, about 21, about 22, about 23, about 24, about 25, about 26, about 27, about 28, about 29, about 30, about 31, about 32, about 33, about 34, about 35, about 36, about 37, about 38, about 39, about 40, about 41, about 42, about 43, about 44, about 45, about 46, about 47, about 48, about 49, about 50, about 60, about 70, about 80, about 90, about 100, about 110, about 120 or greater amino molecule residues, and any range derivable therein. In particular embodiments, a neoantigen has a peptide sequence between about 5 to about 50 amino acids. In particular embodiments, a neoantigen has a peptide sequence between about 10 to about 20 amino acids.

[0131] Neoantigens can be: for MHC Class 1 15 residues or less in length and usually consist of between about 8 and about 11 residues, particularly 9 or 10 residues; for MHC Class II, 15-24 residues.

[0132] In various embodiments, neoantigens do not induce an autoimmune response and / or invoke immunological tolerance when administered to an individual.

[0133] Also disclosed herein are compositions comprising two or more neoantigens. In some embodiments the composition contains at least two distinct peptide sequences. At least two distinct peptide sequences can be derived from the same polypeptide. The phrase distinct polypeptides encompasses polypeptides that vary by length, amino acid sequence, or both. The peptides are derived from any polypeptide known to or have been found to contain a tumor specific mutation. Suitable polypeptides from which neoantigens can be derived can be found, for example, in the COSMIC database. COSMIC curates comprehensive information on somatic mutations in human cancer.

[0134] Proteins or peptides, such as neoantigen polypeptides, can be made by any technique known to those of skill in the art, including the expression of proteins, polypeptides or peptides through standard molecular biological techniques, the isolation of proteins or peptides from natural sources, or the chemical synthesis of proteins or peptides. The nucleotide and protein, polypeptide and peptide sequences corresponding to various genes have been previously disclosed, and can be found at computerized databases known to those of ordinary skill in the art. One such database is the National Center for BiotechnologyAttorney Docket No. N.093.W0.01Information’s Genbank and GenPept databases located at the National Institutes of Health website. The coding regions for known genes can be amplified and / or expressed using the techniques disclosed herein or as would be known to those of ordinary skill in the art.Alternatively, various commercial preparations of proteins, polypeptides and peptides are known to those of skill in the art.Vaccine Compositions

[0135] Also disclosed herein is an immunogenic composition, e.g., a vaccine composition, capable of raising a specific immune response, c.g., a tumor-specific immune response. Vaccine compositions may have one or more neoantigens. In particular embodiments, the one or more neoantigens included in a vaccine composition are selected using the methods disclosed herein e.g., methods for determining a personalized prediction for an individual e.g., by analyzing one or more neoantigens and determining the one or more neoantigens are likely to cause an immunogenic reaction in the individual.

[0136] In various embodiments, a vaccine can contain between 1 and 30 peptides, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, or 30 different peptides, 6, 7, 8, 9, 10 11, 12, 13, or 14 different peptides, or 12, 13 or 14 different peptides. In various embodiments, a vaccine can contain between 1 and 30 neoantigen sequences, 2. 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25.26. 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41. 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94,95, 96, 97, 98, 99, 100 or more different neoantigen sequences, 6, 7, 8, 9, 10 11, 12, 13, or 14 different neoantigen sequences, or 12, 13 or 14 different neoantigen sequences.

[0137] In various embodiments, a vaccine composition comprising one or more neoantigens can be capable of raising a specific cytotoxic T cells response and / or a specific helper T cell response. In various embodiments, a vaccine composition can further have an adjuvant and / or a carrier. Examples of useful adjuvants and carriers are given herein below. A composition can be associated with a carrier such as e.g. a protein or an antigen-presenting cell such as e.g. a dendritic cell (DC) capable of presenting the peptide to a T cell.Adjuvants are any substance whose admixture into a vaccine composition increases or otherwise modifies the immune response to a neoantigen. Carriers can be scaffold structures, for example a polypeptide or a polysaccharide, to which a neoantigen, is capable of being associated. Optionally, adjuvants are conjugated covalently or non-covalently. The ability ofAttorney Docket No. N.093.W0.01an adjuvant to increase an immune response to an antigen is typically manifested by a significant or substantial increase in an immune-mediated reaction, or reduction in disease symptoms. For example, an increase in humoral immunity is typically manifested by a significant increase in the titer of antibodies raised to the antigen, and an increase in T cell activity is typically manifested in increased cell proliferation, or cellular cytotoxicity, or cytokine secretion. An adjuvant may also alter an immune response, for example, by changing a primarily humoral or T helper cell (Th) response into a primarily cellular, or Th response.

[0138] Suitable adjuvants include, but are not limited to 1018 ISS, alum, aluminium salts, Amplivax, AS15, BCG, CP-870,893, CpG7909, CyaA, dSLIM, GM-CSF, IC30, IC31, Imiquimod, ImuFact IMP321, IS Patch, ISS, ISCOMATRIX, Juvlmmune, LipoVac, MF59, monophosphoryl lipid A, Montanide IMS 1312, Montanide ISA 206, Montanide ISA 50V, Montanide ISA-51, OK-432, OM-174, OM-197-MP-EC, ONTAK, PepTel vector system, PLG microparticles, resiquimod, SRL172, Virosomes and other Virus-like particles, YF-17D, VEGF trap, R848, beta-glucan, Pam3Cys, Aquila's QS21 stimulon (Aquila Biotech, Worcester, Mass., USA) which is derived from saponin, mycobacterial extracts and synthetic bacterial cell wall mimics, and other proprietary adjuvants such as Ribi's Detox. Quil or Superfos. Adjuvants such as incomplete Freund's or GM-CSF are useful. Several immunological adjuvants (e.g., MF59) specific for dendritic cells and their preparation have been described previously (Dupuis M, et al., Cell Immunol. 1998; 186(1): 18-27; Allison A C; Dev Biol Stand. 1998; 92:3-11). Also, cytokines can be used. Several cytokines have been directly linked to influencing dendritic cell migration to lymphoid tissues (e.g., TNF-alpha), accelerating the maturation of dendritic cells into efficient antigen-presenting cells for T-lymphocytes (e.g., GM-CSF, IL-1 and IL-4) (U.S. Pat. No. 5,849,589, specifically incorporated herein by reference in its entirety) and acting as immunoadjuvants (e.g., IL-12) (Gabrilovich D I, et al., J Immunother Emphasis Tumor Immunol. 1996 (6) :414-418).

[0139] A carrier (or excipient) can be present independently of an adjuvant. The function of a carrier can for example be to increase the molecular weight of in particular mutant to increase activity or immunogenicity, to confer stability, to increase the biological activity, or to increase serum half-life. Furthermore, a carrier can aid presenting peptides to T cells. A earner can be any suitable carrier known to the person skilled in the art, for example a protein or an antigen presenting cell. A carrier protein could be but is not limited to keyhole limpet hemocyanin, serum proteins such as transferrin, bovine serum albumin, human serumAttorney Docket No. N.093.W0.01albumin, thyroglobulin or ovalbumin, immunoglobulins, or hormones, such as insulin or palmitic acid. For immunization of humans, the carrier is generally a physiologically acceptable carrier acceptable to humans and safe. However, tetanus toxoid and / or diptheria toxoid are suitable carriers. Alternatively, the carrier can be dextrans for example sepharose.

[0140] In various embodiments, neoantigens can be encoded by a nucleic acid included in viral vector-based vaccine platforms, such as vaccinia, fowlpox, self-replicating alphavirus, marabavirus, adenovirus (See, e.g., Tatsis et al., Adenoviruses, Molecular Therapy (2004) 10, 616 — 629), or lenti virus, including but not limited to second, third or hybrid second / third generation lentivirus and recombinant lentivirus of any generation designed to target specific cell types or receptors (See, e.g., Hu et al., Immunization Delivered by Lentiviral Vectors for Cancer and Infectious Diseases, Immunol Rev. (2011) 239(1): 45-61, Sakuma et al., Lentiviral vectors: basic to translational, Biochem J. (2012) 443(3):603-18, Cooper et al., Rescue of splicing-mediated intron loss maximizes expression in lentiviral vectors containing the human ubiquitin C promoter, Nucl. Acids Res. (2015) 43 (1): 682-690, Zufferey et aL, Self-Inactivating Lentivirus Vector for Safe and Efficient In Vivo Gene Delivery, J. Virol. (1998) 72 (12): 9873-9880). Dependent on the packaging capacity of the above mentioned viral vector-based vaccine platforms, this approach can deliver one or more nucleotide sequences that encode one or more neoantigen peptides. The sequences may be flanked by non-mutated sequences, may be separated by linkers or may be preceded with one or more sequences targeting a subcellular compartment (See, e.g., Gros et al., Prospective identification of neoantigen-specific lymphocytes in the peripheral blood of melanoma patients, Nat Med. (2016) 22 (4):433-8, Stronen et al., Targeting of cancer neoantigens with donor-derived T cell receptor repertoires, Science. (2016) 352 (6291): 1337-41 , Lu et al., Efficient identification of mutated cancer antigens recognized by T cells associated with durable tumor regressions, Clin Cancer Res. (2014) 20( 13) :3401 - 10). Upon introduction into a host, infected cells express the neoantigens, and thereby elicit a host immune (e.g., CTL) response against the peptide(s). Vaccinia vectors and methods useful in immunization protocols are described in, e.g., U.S. Pat. No. 4,722,848. Another vector is BCG (Bacille Calmette Guerin). BCG vectors are described in Stover et al. (Nature 351:456-460 (1991)). A wide variety of other vaccine vectors useful for therapeutic administration or immunization of neoantigens, e.g., Salmonella typhi vectors, and the like will be apparent to those skilled in the art from the description herein.Attorney Docket No. N.093.W0.01Cancers

[0141] The disclosure provides methods for determining a personalized prediction for an individual e.g., by analyzing one or more neoantigens and determining the one or more neoantigens are likely to cause an immunogenic reaction in the individual. In various embodiments, the individual is healthy and / or is not yet suspected of having a cancer. In certain embodiments, the individual is not yet diagnosed with cancer, but has a cancer and is prior to development of symptoms (e.g., such as an early-stage cancer). In various embodiments, in the individual is a cancer patient previously diagnosed with a cancer.

[0142] In various embodiments, the cancer is an early-stage cancer. In various embodiments, the cancer is a preclinical phase cancer. In various embodiments, the cancer is a stage I cancer. In various embodiments, the cancer is a stage II cancer. Thus, the methods disclosed herein enable the screening and diagnosis of an individual for an early stage or preclinical stage cancer.

[0143] In various embodiments, the cancer is any of an acute lymphoblastic leukemia, acute myeloid leukemia, adrenocortical carcinoma, soft tissue sarcoma, lymphoma, anal cancer, gastrointestinal cancer, brain cancer, skin cancer, bile duct cancer, bladder cancer, bone cancer, breast cancer, lung cancer, cardiac cancer, central nervous system cancer, cervical cancer, chronic lymphocytic leukemia, chronic myelogenous leukemia, chronic myeloproliferative neoplasms, colorectal cancer, uterine cancer, esophageal cancer, head and neck cancer, eye cancer, fallopian tube cancer, gallbladder cancer, gastric cancer, germ cell tumor, gestational trophoblastic cancer, hairy cell leukemia, liver cancer, Hodgkin lymphoma, intraocular melanoma, pancreatic cancer, kidney cancer, leukemia, mesothelioma, metastatic cancer, mouth cancer, multiple endocrine neoplasia syndromes, multiple myeloma neoplasms, myelodysplastic neoplasms, ovarian cancer, parathyroid cancer, penile cancer, pheochromocytoma, pituitary cancer, plasma cell neoplasm, primary peritoneal cancer, prostate cancer, rectal cancer, retinoblastoma, sarcoma, small intestine cancer, testicular cancer, throat cancer, thymoma and thymic carcinoma, thyroid cancer, urethral cancer, uterine cancer, vaginal cancer, and vulvar cancer.Computer Implementation

[0144] The methods disclosed herein, including the methods for determining a personalized prediction for an individual, are, in some embodiments, performed on one or more computers. In particular embodiments, the steps performed by the prediction system 130 (e.g., methods for deploying a prediction model during inference, and / or methods for trainingAttorney Docket No. N.093.W0.01a prediction model) are performed on one or more computers. The steps of performing an assay (e.g., assay 120 shown in FIG. 1 A) are not performed on one or more computers.

[0145] In various embodiments, the methods for determining a personalized prediction for an individual can be implemented in hardware, firmware, software, or any combination thereof. In one embodiment, a machine-readable storage medium is provided, the medium comprising a data storage material encoded with machine readable data which, when using a machine programmed with instructions for using said data, is capable of displaying data (e.g., methylation data) and results of the methods for determining a personalized prediction for an individual (e.g., an indication of the personalized prediction e.g., a responder or nonresponder and / or an indication of certain epitopes identified as likely to trigger an immune response in the individual). Such data can be used for a variety of purposes, such as for preparing a personalized therapy, such as a personalized vaccine including epitopes identified as likely to trigger an immune response in the individual. The methods disclosed herein can be implemented in computer programs executing on programmable computers, comprising a processor, a data storage system (including volatile and non-volatile memory and / or storage elements), a graphics adapter, a pointing device, a network adapter, at least one input device, and at least one output device. A display is coupled to the graphics adapter. Program code is applied to input data to perform the functions described above and generate output information. The output information is applied to one or more output devices, in known fashion. The computer can be, for example, a personal computer, microcomputer, or workstation of conventional design.

[0146] Each program can be implemented in a high level procedural or object oriented programming language to communicate with a computer system. However, the programs can be implemented in assembly or machine language, if desired. In any case, the language can be a compiled or interpreted language. Each such computer program is preferably stored on a storage media or device (e.g., ROM or magnetic diskette) readable by a general or special purpose programmable computer, for configuring and operating the computer when the storage media or device is read by the computer to perform the procedures described herein. The system can also be considered to be implemented as a computer-readable storage medium, configured with a computer program, where the storage medium so configured causes a computer to operate in a specific and predefined manner to perform the functions described herein.Attorney Docket No. N.093.W0.01

[0147] The signature patterns and databases thereof can be provided in a variety of media to facilitate their use. “Media” refers to a manufacture that contains the signature pattern information of the present invention. The databases of the present invention can be recorded on computer readable media, e.g. any medium that can be read and accessed directly by a computer. Such media include, but are not limited to: magnetic storage media, such as floppy discs, hard disc storage medium, and magnetic tape; optical storage media such as CD-ROM; electrical storage media such as RAM and ROM; and hybrids of these categories such as magnetic / optical storage media. One of skill in the art can readily appreciate how any of the presently known computer readable mediums can be used to create a manufacture comprising a recording of the present database information. “Recorded” refers to a process for storing information on computer readable medium, using any such methods as known in the art. Any convenient data storage structure can be chosen, based on the means used to access the stored information. A variety of data processor programs and formats can be used for storage, e.g. word processing text file, database format, etc.

[0148] In some embodiments, the methods disclosed herein, including methods for determining a personalized prediction for an individual are performed on one or more computers in a distributed computing system environment (e.g., in a cloud computing environment). In this description, “cloud computing” is defined as a model for enabling on-demand network access to a shared set of configurable computing resources. Cloud computing can be employed to offer on-demand access to the shared set of configurable computing resources. The shared set of configurable computing resources can be rapidly provisioned via virtualization and released with low management effort or service provider interaction, and then scaled accordingly. A cloud-computing model can be composed of various characteristics such as, for example, on-demand self-service, broad network access, resource pooling, rapid elasticity, measured service, and so forth. A cloud-computing model can also expose various service models, such as, for example, Software as a Service (“SaaS”), Platform as a Service (“PaaS”), and Infrastructure as a Service (“laaS”). A cloud-computing model can also be deployed using different deployment models such as private cloud, community cloud, public cloud, hybrid cloud, and so forth. In this description and in the claims, a “cloud-computing environment” is an environment in which cloud computing is employed.

[0149] In some embodiments, the output(s) from the prediction system 130 can be used to generate executable manufacturing control instructions identifying one or moreAttorney Docket No. N.093.W0.01candidate neoantigens and encoding manufacturing-relevant information. In some embodiments, the executable manufacturing control instructions may be generated locally or may be received from an external computing system. The executable manufacturing control instructions may be provided to a preparation, formulation, or manufacturing system to produce a personalized therapeutic vaccine composition comprising neoantigen components corresponding to the scored and / or identified candidate neoantigens. The manufacturing system may comprise automated synthesis, formulation, or packaging equipment controlled by the executable manufacturing instructions to produce the therapeutic composition. The identification of candidate neoantigens, generation of digital instructions, and execution of manufacturing operations may be performed by separate computing systems or at different times, provided that the operations are logically linked by the subject-specific or catalog-based target selection described herein.Example Computer

[0150] FIG. 5 illustrates an example computer for implementing the entities shown in FIGs.1 A-1B, 2A-2C, 3, and 4A-4B. In particular embodiments, the example computer 500 can represent computational system 202 described in FIG. 2A. The computer 500 includes at least one processor 502 coupled to a chipset 504. The chipset 504 includes a memory controller hub 520 and an input / output (I / O) controller hub 422. A memory 506 and a graphics adapter 512 are coupled to the memory controller hub 520, and a display 518 is coupled to the graphics adapter 512. A storage device 508, an input device 514, and network adapter 516 are coupled to the I / O controller hub 522. Other embodiments of the computer 500 have different architectures.

[0151] The storage device 508 is a non-transitory computer-readable storage medium such as a hard drive, compact disk read-only memory (CD-ROM), DVD, or a solid-state memory device. The memory 506 holds instructions and data used by the processor 502. The input interface 514 is a touch-screen interface, a mouse, track ball, or other type of pointing device, a keyboard, or some combination thereof, and is used to input data into the computer 500. In some embodiments, the computer 500 may be configured to receive input (e.g., commands) from the input interface 514 via gestures from the user. The graphics adapter 512 displays images and other information on the display 518. The network adapter 516 couples the computer 500 to one or more computer networks.

[0152] The computer 500 is adapted to execute computer program modules for providing functionality described herein. As used herein, the term “module” refers to computerAttorney Docket No. N.093.W0.01program logic used to provide the specified functionality. Thus, a module can be implemented in hardware, firmware, and / or software. In one embodiment, program modules are stored on the storage device 508, loaded into the memory 506, and executed by the processor 502. A module can be implemented as computer program code processed by the processing system(s) of one or more computers. Computer program code includes computerexecutable instructions and / or computer-interpreted instructions, such as program modules, which instructions are processed by a processing system of a computer. Generally, such instructions define routines, programs, objects, components, data structures, and so on, that, when processed by a processing system, instruct the processing system to perform operations on data or configure the processor or computer to implement various components or data structures in computer storage. A data structure is defined in a computer program and specifies how data is organized in computer storage, such as in a memory device or a storage device, so that the data can accessed, manipulated, and stored by a processing system of a computer.

[0153] The types of computers 500 used by the entities of FIGs. 1 A and IB can vary depending upon the embodiment and the processing power required by the entity. For example, the prediction system 130 can run in a single computer 500 or multiple computers 500 communicating with each other through a network such as in a server farm. The computers 500 can lack some of the components described above, such as graphics adapters 512, and displays 518.Systems

[0154] Further disclosed herein are systems for performing methods for determining a personalized prediction for an individual. In various embodiments, such a system can be configured to perform an assay (e.g., assay 120 disclosed in FIG. 1A). The data from the assay can then be used by the prediction system 130 for analysis and for generation of a personalized prediction. In various embodiments, systems disclosed herein can include a computer (e.g., computer 500), which can be communicatively coupled to an apparatus to obtain data from an assay.

[0155] The apparatus is configured to determine information from a sample obtained from an individual, such as sequencing information. For example, the apparatus can be configured to perform one or more of a nucleic acid amplification assay (e.g., polymerase chain reactionAttorney Docket No. N.093.W0.01assay) and / or nucleic acid sequencing (e.g., targeted gene sequencing, whole genome sequencing, or next generation sequencing).

[0156] In various embodiments, systems disclosed herein can be configured to enable determination of information (e.g., sequencing information) from the individual using one or more detection reagents. For example, detection reagents can be a set of primers that, when combined with a sample obtained from the individual, allows generation of sequence reads for sequencing.

[0157] The mixture of the detection reagents and sample may be presented to the apparatus through various conduits, examples of which include wells of a well plate (e.g., 96 well plate), a vial, a tube, and integrated fluidic circuits. As such, the apparatus may have an opening (e.g., a slot, a cavity, an opening, a sliding tray) that can receive the container including the reagent test sample mixture and perform a reading. Examples of an apparatus include one or more of a sequencer, an incubator, plate reader (e.g., a luminescent plate reader, absorbance plate reader, fluorescence plate reader), a spectrometer, or a spectrophotometer.

[0158] The computer system, such as example computer 500 described in FIG. 5, communicates with the apparatus to receive the information e.g., sequencing information. The computer system analyzes the information to determine a personalized prediction for an individual.EXAMPLES

[0159] Below are examples of specific embodiments for carrying out the present invention. The examples are offered for illustrative purposes only and are not intended to limit the scope of the present invention in any way. Efforts have been made to ensure accuracy with respect to numbers used (e.g., percentages, etc.), but some experimental error and deviation should be allowed for.Example 1: Deep Antigen Sets (DAS) Pipeline Successfully Identifies Higher Percentage of Immunogenic Neoantigens

[0160] This example introduces Deep Antigen Sets (DAS), a deep-learning framework for neoantigen prediction by directly modeling patient-level immune responses while making precise predictions at the individual neoantigen level. Using a permutationinvariant architecture, DAS models the set of potential neoantigens, which varies in size based on each patient’s unique tumor mutational profile, as a holistic input. This enables theAttorney Docket No. N.093.W0.01leveraging of patient-level clinical outcomes, bridging the gap between in vitro and in vivo predictions. Furthermore, a sampling strategy was implemented to mitigate dataset noise, focusing training on the most informative neoantigen candidates. After training on neoantigen sets from 543 in-house patients, DAS achieved 76% accuracy in predicting clinical response on the internal hold-out cohort. DAS was further validated on three independent clinical datasets: NCI, TESLA, and HiTIDE. The model’s ability to prioritize immunogenic neoantigens was evaluated by determining whether these neoantigens appear within the top 20, top 50, and top 100 predictions for each patient. DAS outperformed existing methods, achieving over 60%, 80%, and 90% recall at these respective ranking thresholds. DAS constitutes a paradigm shift by directly modeling in vivo neoantigen sets and patient responses to enhance the potential of personalized cancer vaccines. Its computational efficiency facilitates the rapid analysis of large patient cohorts, seamlessly integrating into clinical workflows. This powerful tool empowers oncologists and immunologists to design more effective targeted immunotherapies, ultimately improving patient outcomes.

[0161] Specifically disclosed in this example is a DAS pipeline that involved a multiple step process of 1) a sampling step for generating higher quality training data by analyzing biochemical properties of neoantigens 2) augmenting the training data, and 3) training a prediction model with the disclosed architecture using the higher quality training data. Using this multiple step process during training, the resulting prediction model achieved improved personalized predictions of candidate neoantigens during inference. This disclosed pipeline that uses the multiple step process outperforms conventional processes that generate predictions for neoantigens based on biochemical properties of those neoantigens.

[0162] FIG. 6 provides an overview of the disclosed pipeline 600 for neoantigen prioritization during training, where patient data was processed at least by sampling and through a prediction model to generate a classification result for each patent. As depicted, pipeline 600 starts at step 602, where each patient is represented as a set of epitope sequences derived from somatic mutations in the tumor. In some embodiments, the epitope set corresponded to the initial set of neoantigens 220 shown in FIG. 2B, which captured the full complement of candidate neoantigens identified for a given patient through sequencing of nucleic acids from the patient as shown in FIG. 1 A.

[0163] At step 604, a sampling process received an input epitope set and refined the selection for determining candidate neoantigens for training. Specifically, this epitopeAttorney Docket No. N.093.W0.01sampling at step 604 prioritized epitopes with higher predicted immunogenic potential based on known biochemical properties. This sampling process helped reduce noise in the dataset by filtering out weak or non-immunogenic epitopes. In some embodiments, at step 604, an input epitope set was provided as input into the trained machine learning model trained using in vivo immune response data. The trained machine learning model refined the selection of candidate neoantigens for training and outputted immunogenicity scores. The top-ranked neoantigens were selected to form a candidate pool. For example, the top 200 candidate neoantigens with the highest scores were selected.

[0164] The top-ranked (e.g., 200) candidate neoantigens, i.e., the refined set 606, visually conveyed that the sampling process had reduced the number of neoantigens by retaining those with stronger predicted immunogenic potential and excluding less immunogenic or non-immunogenic neoantigens. The refined set 606 had no inherent order and therefore, was provided as an unordered collection to a prediction model 608. In some embodiments, the refined set 606 was provided into a permutation-invariant architecture. The refined set 606 may correspond to the refined set of neoantigens 226 shown in FIG. 2B, and the prediction model 608 may correspond to the prediction model 230 described with respect to FIGs. 2 A, 2C, and 3.

[0165] Using the prediction model 608, individual epitopes were first embedded and processed through a shared MLP. A pooling mechanism then aggregated their representations through a pooling step 610 into a fixed-length patient-level feature vector, which was fed into a final classification layer to predict the likelihood of an immune response. In some embodiments, the classification output is compared against an individuallevel ground truth label indicating whether the patient experienced an immune response, as described in FIG. 2C. The responder versus non-responder distinction in classification output 612 may correspond to a binary ground truth label.

[0166] In a particular embodiment, the pipeline 600 includes a permutation-invariant deep learning architecture 608 well-suited for modeling the unordered nature of neoantigen sets associated with each patient. The disclosed pipeline was evaluated using three publicly available datasets: NCI, TESLA, and HiTIDE. The NCI dataset is further described in Muller, M. et al., Machine learning methods and harmonized datasets improve immunogenic neoantigen prediction. Immunity, 56(11 ):2650— 2663, 2023, which is incorporated by reference in its entirety. The TESLA dataset is further described in Parkhurst, M. et al., Unique neoantigens arise from somatic mutations in patients with gastrointestinal cancers.Attorney Docket No. N.093.W0.01Cancer discovery, 9(8): 1022-1035, 2019, which is incorporated by reference in its entirety. The HiTIDE dataset is described in Wells, D. et al., Key parameters of tumor epitope immunogenicity revealed through a consortium approach improve neoantigen prediction. Cell, 183(3):818- 834, 2020, which is incorporated by reference in its entirety. The disclosed pipeline outperforms existing approaches in accurately predicting immunogenic neoantigens in all datasets. The disclosed pipeline achieves higher prioritization accuracy, capturing a greater proportion of immunogenic neoantigens within the top-ranked predictions, as further demonstrated with respect to FIGs. 7, 8 A and 8B.Methods - Epitope Sampling

[0167] The high dimensionality and inherent noise in neoantigen datasets present significant challenges for model training. Immune responses to neoantigens exhibit inherent sparsity as only a small subset of candidate epitopes typically triggers an immune response in any given patient. Thus, including all epitopes during training introduces substantial noise, as the majority of epitopes in each set are likely non-immunogenic.

[0168] To address this challenge, an epitope sampling procedure is implemented. The epitope sampling procedure leverages the immunogenicity scores produced by the machine learning model. Formally, for a patient pt = {x: , x? , ... , x1}, rank all its epitopes according to their immunogenicity scores and select the top 200 epitopes to create a candidate pool. Next, k different training instances were generated for the patient, an increasing number of epitopes were assigned. Specifically, for the j -th instance (j = 0,. . . , k — 1), we sample n.j = 20 + 5j epitopes from the candidate pool. This creates a sequence of training instances with progressively larger epitope sets. Each instance maintained the original patient label, thus effectively augmenting the training data. By creating multiple views of the same patient with different epitope subsets, the augmented training set enables the model to learn more robust representations across different set sizes. The usage of immunogenicity scores for the initial ranking ensures that epitopes with stronger predicted immunogenic potential are prioritized, while random sampling within the top 200 maintains sufficient diversity in the training data.Methods

[0169] A permutation-invariant deep learning architecture designed for modeling sets was used. This deep learning model is well suited for this scenario, where each patient sample is represented as an unordered set of candidate neoantigens.Attorney Docket No. N.093.W0.01

[0170] Formally, a patient ptis represented as a set of epitopes:C 1 2 -4Pi = {xi,xi, ..., xiwhere xtJE Rdis a feature vector describing the j-th epitope of patient i. The goal is to learn a function / that maps a set of epitopes to a patient-level immune response score while ensuring pennutation invariance, meaning that the prediction should remain unchanged regardless of the order of elements in p . Moreover, since the number of epitopes per patient Tii varies, the model must be capable of handling sets of variable size. The model achieves this by learning a function f with the following structure:ftPi) =where < >: Rd-> Rkis a learnable function that is shared and acts on each individual elements and p: Rk-> R is another learnable function that processes the aggregated representation and produces a final patient-level prediction. Two-layer multi-layer perceptrons (MLPs) are used for < > and p and the max pooling operator for the aggregation.Training and Inference

[0171] The model was trained by minimizing the cross-entropy loss:= -[Ti log log + (1 - yj log log (1 - f(p;) )] where f(pt) is the model’s predicted immune response for patient pt, and ytis the corresponding ground truth label.

[0172] The main goal was to predict the immunogenicity of new neoantigens rather than the immune response of a patient. Therefore, at inference time, to achieve this, a patient was simulated for each individual candidate neoantigen by treating it as a patient with a single epitope. Specifically, for each candidate neoantigen, a set consisting solely of that single epitope was created and was passed through the model, generating a prediction that represents the probability that the given neoantigen is immunogenic. Consequently, the candidate neoantigens can be ranked according to their predicted immunogenicity, identifying the most promising neoantigens for further experimental testing or potential therapeutic development.

[0173] To assess model performance, the framework was evaluated for its ability to correctly prioritize immunogenic neoantigens across different ranking thresholds. FIG. 7 shows the inputting of candidate neoantigens during an inference phase into the deep learning model to generate predicted scores representing the immunogenicity of the epitopes.Attorney Docket No. N.093.W0.01Specifically, evaluation involved determining whether the truly immunogenic neoantigens appeared within the Top 20, Top 50, and Top 100 predictions for each patient of each dataset. These rankings represent practical thresholds for experimental validation, as selecting a manageable number of high-confidence neoantigens is valuable for designing personalized cancer vaccines.

[0174] In some embodiments, each candidate neoantigen in FIG. 7 was represented during inference as a single-element neoantigen set and was processed using the same permutationinvariant prediction architecture used during training. Specifically, the peptide sequence of the candidate neoantigen can be provided to one or more shared sub-structures that generate one or more transformed neoantigen representations. The one or more transformed neoantigen representations can be aggregated by a pooling structure into a fixed-length individual-level representation. A classification structure can generate a predicted immunogenicity score from the fixed- length individual-level representation. In this manner, the disclosed inference process can reuse the same set-based architecture during inference without requiring a separate candidate-specific prediction architecture.

[0175] In some embodiments, this inference process provides a technical mechanism by which a prediction model trained on patient- level immune response labels over unordered neoantigen sets is deployed to evaluate an individual candidate neoantigen without requiring retraining of the prediction model for candidate-level supervision. Because the trained architecture is permutation-invariant and tolerant of different set cardinalities, the prediction model can be executed on a single-element set while preserving compatibility with the setprocessing function learned during training. Thus, a deep learning (DL) model does not require neoantigen-level labels during training in order to generate candidate-neoantigen-level immunogenicity scores during inference.

[0176] In FIG. 7, two example parallel inference pathways, each including a candidate neoantigen provided as input to a respective instance 702 and 704 of a DL model, are illustrated. These pathways outputted a predicted score of 0.85 and 0.92, respectively, where the higher score reflected a stronger predicted immunogenicity for a particular candidate neoantigen. The candidate neoantigens that were predicted to have the highest level of immunogenicity can be selected based on their ranking for inclusion in a therapeutic vaccine or for further experimental validation.

[0177] In some embodiments, the disclosed process avoids requiring a rigid positiondependent input representation for each candidate neoantigen and instead reuses the sameAttorney Docket No. N.093.W0.01shared sub-structure and pooling architecture to generate a machine-processable fixed-length representation from each candidate input. This can improve computerized handling of candidate neoantigen evaluation across large patient-specific candidate pools because candidate neoantigens can be independently evaluated in a uniform set-based format irrespective of the total number of candidate neoantigens identified for the patient. The same trained machine function can thereby be repeatedly executed across many candidate neoantigens derived from the same sequencing workflow.

[0178] FIG. 8A shows the performance evaluation of the machine learning model alone on predicting immunogenic neoantigens across different datasets. In contrast, FIG. 8B shows the performance evaluation of the disclosed framework on predicting immunogenic neoantigens across different datasets.

[0179] FIG. 8A provides a performance baseline by depicting results achieved by a conventional methodology that generates predictions for neoantigens based on biochemical properties of those neoantigens without implementing the disclosed prediction model architecture. As depicted, three independent clinical datasets: NCI, TESLA, and HiTIDE, were used. For each dataset, three bars are displayed corresponding to three different ranking thresholds: Top 20, Top 50, and Top 100. For example, for the NCI dataset, the methodology in FIG. 8A achieved a percentage of predicted positives of 54.5 at the Top 20 ranking threshold, 77.2 at the Top 50 ranking threshold, and 89.7 at the Top 100 ranking threshold.

[0180] FIG. 8B depicts results achieved by the disclosed pipeline, which involves sampling training neoantigens based on biochemical properties to generate a refined candidate pool, augmenting the refined candidate pool to generate training neoantigen instances of varying sizes, and training a prediction model having a permutation-invariant deep learning architecture using the augmented training neoantigen instances. For each dataset of NCI, TESLA, and HiTIDE, three bars are displayed corresponding to the Top 20, Top 50, and Top 100 ranking thresholds. For example, for the NCI dataset, the disclosed pipeline achieved a percentage of predicted positives of 62.8 at the Top 20 ranking threshold, 81.9 at the Top 50 ranking threshold, and 91.5 at the Top 100 ranking threshold.

[0181] The results shown in FIGs. 8A and 8B demonstrate that the disclosed pipeline (including a multiple step process) shown in FIG. 6 consistently outperforms conventional processes that generate predictions for neoantigens based on biochemical properties of those neoantigens (but does not implement the disclosed prediction model using a permutationinvariant deep learning architecture). The disclosed pipeline achieves higher prioritizationAttorney Docket No. N.093.W0.01accuracy, capturing a greater proportion of truly immunogenic neoantigens within the topranked predictions.

[0182] More specifically, the performance improvements shown in FIG. 8B can arise from the combined effect of several technical features of the disclosed framework. First, the sampling stage reduces high-dimensional input noise by restricting training to a refined candidate pool enriched for neoantigens having stronger predicted immunogenic potential. Second, the augmentation stage generates multiple training neoantigen instances of different cardinalities from the same patient-derived neoantigen pool, thereby improving robustness of the trained prediction model across variable-size input sets. Third, the permutation-invariant prediction architecture applies shared sub-structures across individual neoantigens and aggregates the resulting neoantigen representations using a pooling operation into a fixed-length patient- level representation, thereby enabling machine learning over unordered neoantigen sets without dependence on input order.

[0183] In some implementations, the improved recall at the Top 20, Top 50, and Top 100 thresholds in FIG. 8B reflects a technical improvement in computerized neoantigen prioritization rather than merely presentation of a mathematical result. More specifically, the disclosed framework improves the ability of a computer-implemented prediction system to detect sparse response-associated neoantigen signals embedded within noisy, patient- specific candidate pools and to convert those signals into candidate-level rankings suitable for downstream therapeutic selection. The disclosed framework thereby improves computer-implemented ranking of candidate neoantigens derived from sequencing data.

[0184] In some implementations, the results shown in FIG. 8B demonstrate that the disclosed architecture more effectively processes unordered biological data structures having variable cardinalities than conventional biochemical-property-only ranking approaches. The disclosed framework therefore provides a technical improvement in machine-based prioritization of candidate neoantigens derived from biological samples and sequencing workflows, including workflows in which ranked candidate neoantigens are subsequently selected for experimental validation or for inclusion in personalized therapeutic compositions. The candidate neoantigens identified as having the highest predicted immunogenicity through the ranking process depicted in FIG. 7 can be selected for inclusion in a therapeutic vaccine for administration to an individual.

[0185] It is to be understood that both the general description and the detailed description provide examples that are explanatory in nature and are intended to provide an understandingAttorney Docket No. N.093.W0.01of the present disclosure without limiting the scope of the present disclosure. Various mechanical, compositional, structural, electronic, and operational changes may be made without departing from the scope of this description and the claims. In some instances, well-known circuits, structures, and techniques have not been shown or described in detail in order not to obscure the examples. Like numbers in two or more figures may represent the same or similar elements.

[0186] In addition, the singular forms “a”, “an”, and “the” are intended to include the plural forms as well, unless the context indicates otherwise. Moreover, the terms “comprises,” “comprising,” “includes,” and the like specify the presence of stated features, steps, operations, elements, and / or components but do not preclude the presence or addition of one or more other features, steps, operations, elements, components, and / or groups. Components described as coupled may be electronically or mechanically directly coupled, or they may be indirectly coupled via one or more intermediate components, unless specifically noted otherwise. Mathematical and geometric terms are not necessarily intended to be used in accordance with their strict definitions unless the context of the description indicates otherwise, because a person having ordinary skill in the art would understand that, for example, a substantially similar element that functions in a substantially similar way could easily fall within the scope of a descriptive term even though the term also has a strict definition.

[0187] Elements and their associated aspects that are described in detail with reference to one example may, whenever practical, be included in other examples in which they are not specifically shown or described. For example, if an element is described in detail with reference to one example and is not described with reference to a second example, the element may nevertheless be claimed as included in the second example.

[0188] Further modifications and alternative examples will be apparent to those of ordinary skill in the art in view of the disclosure herein. For example, the devices and methods may include additional components or steps that were omitted from the diagrams and description for clarity of operation. Accordingly, this description is to be constmed as illustrative only and is for the purpose of teaching those skilled in the art the general manner of carrying out the present teachings. It is to be understood that the various examples shown and described herein are to be taken as exemplary. Elements and materials, and arrangements of those elements and materials, may be substituted for those illustrated and described herein, parts and processes may be reversed, and certain features of the present teachings may be utilizedAttorney Docket No. N.093.W0.01independently, all as would be apparent to one skilled in the art after having the benefit of the description herein. Changes may be made in the elements described herein without departing from the scope of the present teachings and following claims.

[0189] It is to be understood that the particular examples set forth herein are non-limiting, and modifications to structure, dimensions, materials, and methodologies may be made without departing from the scope of the present teachings.

[0190] Other examples in accordance with the present disclosure will be apparent to those skilled in the art from consideration of the specification and practice of the disclosure herein. It is intended that the specification and examples be considered as exemplary only, with the following claims being entitled to their fullest breadth, including equivalents, under the applicable law.

[0191] While this specification contains many specific implementation details, these should not be construed as limitations on the scope of what may be claimed, but rather as descriptions of features that may be specific to particular embodiments. Certain features that are described in this specification in the context of separate embodiments can also be implemented in combination in a single embodiment. Conversely, various features that are described in the context of a single embodiment can also be implemented in multiple embodiments separately or in any suitable subcombination. Moreover, although features may be described above as acting in certain combinations and even initially claimed as such, one or more features from a claimed combination can in some cases be excised from the combination, and the claimed combination may be directed to a subcombination or variation of a subcombination.

[0192] Similarly, while operations are depicted in the drawings in a particular order, this should not be understood as requiring that such operations be performed in the particular order shown or in sequential order, or that all illustrated operations be performed, to achieve desirable results. In certain circumstances, multitasking and parallel processing may be advantageous. Moreover, the separation of various system components in the embodiments described above should not be understood as requiring such separation in all embodiments, and it should be understood that the described program components and systems can generally be integrated together in a single software product or packaged into multiple software products.

[0193] Particular embodiments of the subject matter have been described. Other embodiments are within the scope of the following claims. The disclosed embodimentsAttorney Docket No. N.093.W0.01constitute technical solutions implemented using computing systems, data processing pipelines, and controlled manufacturing equipment to solve technical problems in analyzing tumor-derived molecular data and preparing therapeutic compositions. For example, the actions recited in the claims can be perfomred in a different order and still achieve desirable results. As one example, the processes depicted in the accompanying figures do not necessarily require the particular order shown, or sequential order, to achieve desirable results. In certain implementations, multitasking and parallel processing may be advantageous. Other steps or stages may be provided, or steps or stages may be eliminated, from the described processes. Accordingly, other implementations are within the scope of the following claims.

Claims

Attorney Docket No. N.093.W0.01CLAIMS1. A method for assessing immunogenicity of a candidate neoantigen for an individual, the method comprising:obtaining or having obtained the candidate neoantigen for the individual; analyzing the candidate neoantigen using a trained prediction model by:generating one or more candidate neoantigen representations of the candidate neoantigen; andpooling the one or more candidate neoantigen representations into a single individual-level representation; andgenerating a score for the candidate neoantigen using the single individual-level representation.

2. The method of claim 1, wherein the candidate neoantigen comprises a peptide sequence comprising between about 5 to about 50 amino acids.

3. The method of claim 2, wherein the peptide sequence comprises between about 10 to about 20 amino acids.

4. The method of any one of claims 1-3, wherein obtaining or having obtained the candidate neoantigen for the individual comprises performing sequencing of nucleic acids from the individual and using nucleic acid sequences to obtain a peptide sequence of the candidate neoantigen.

5. The method of claim 4, wherein the sequencing is next generation sequencing (NGS) or a massively parallel sequencing approach.

6. The method of claim 4 or claim 5, wherein performing the sequencing of nucleic acids comprises performing sequencing on tumor tissue comprising the nucleic acids from the individual.

7. The method of any one of claims 4-6, wherein performing the sequencing of nucleic acids comprises performing sequencing on nucleic acids in a sample obtained from the individual.

8. The method of claim 7, wherein the sample is one of a blood sample, a serum sample, a stool sample, a urine sample, a mucous sample, or a saliva sample.Attorney Docket No. N.093.W0.

019. The method of any one of claims 4-8, wherein performing the sequencing of nucleic acids comprises performing one of targeted, exome, transcriptome, or whole genome sequencing.

10. The method of any one of claims 1-9, further comprising generating one or more candidate neoantigen instances of the candidate neoantigen.

11. The method of claim 10, wherein generating the candidate neoantigen instance comprises sampling copies of the candidate neoantigen.

12. The method of claim 10 or claim 11, wherein generating the one or more candidate neoantigen representations of the candidate neoantigen comprises applying one or more substructures of the prediction model to analyze the one or more candidate neoantigen instances of the candidate neoantigen.

13. The method of claim 12, wherein the one or more sub-structures of the prediction model comprise one or more neural networks.

14. The method of claim 12, wherein the one or more sub-structures of the prediction model comprise one or more multi-layer perceptrons (MLPs).

15. The method of claim 14, wherein the one or more MLPs are two-layer MLPs.

16. The method of any one of claims 10-15, wherein the one or more candidate neoantigen instances comprise N candidate neoantigen instances, where N is a total number of sub-structures of the prediction model.

17. The method of any one of claims 1-16, wherein the one or more candidate neoantigen representations of the candidate neoantigen comprise one or more candidate neoantigen embeddings.

18. The method of any one of claims 1-17, wherein pooling the one or more candidate neoantigen representations comprises performing a max pooling operation to downsample the one or more candidate neoantigen representations into the single individual-level representation.

19. The method of any one of claims 1-18, wherein the generated score for the candidate neoantigen represents a measure of immunogenicity of the candidate neoantigen.Attorney Docket No. N.093.W0.0120. The method of any one of claims 1-19, wherein the prediction model is trained by: obtaining or having obtained peptide sequences for a plurality of training neoantigens; sampling the plurality of training neoantigens to identify a set of training neoantigens representing a subset of the plurality of training neoantigens;augmenting the set of training neoantigens to generate a plurality of training neoantigen instances;training the prediction model using the plurality of training neoantigen instances, wherein the training comprises:for each of one or more training neoantigen instances of the plurality of training neoantigen instances:generating training neoantigen representations for the training neoantigen instance;pooling the training neoantigen representations into a single individuallevel representation; andgenerating a prediction for the training neoantigen instance using the single individual-level representation.

21. The method of claim 20, wherein training the prediction model comprises adjusting one or more parameters of the prediction model to minimize a loss function.

22. The method of claim 21, wherein the loss function is a cross-entropy loss function.

23. The method of claim 21 or claim 22, wherein the loss function represents a combination of the prediction for the training neoantigen and a ground truth label.

24. The method of any one of claims 20-23, wherein the plurality of training neoantigen instances is associated with the ground truth label.

25. The method of claim 23 or claim 24, wherein the ground truth label is an individuallevel ground truth label for a training individual, wherein each training neoantigen from the individual is associated with the individual-level ground truth label.

26. The method of any one of claims 20-25, wherein each training neoantigen comprises a peptide sequence comprising between about 5 to about 50 amino acids.Attorney Docket No. N.093.W0.0127. The method of any one of claims 20-26, wherein obtaining or having obtained the peptide sequences for the plurality of training neoantigens comprises performing sequencing of nucleic acids from a training individual and using nucleic acid sequences to obtain the peptide sequences for the plurality of training neoantigens.

28. The method of claim 27, wherein obtaining the peptide sequences for the plurality of training neoantigens comprises performing the sequencing of nucleic acids from the training individual.

29. The method of claim 27 or claim 28, wherein the sequencing is next generation sequencing (NGS) or a massively parallel sequencing approach.

30. The method of any one of claims 27-29, wherein performing sequencing of nucleic acids comprises performing sequencing on tumor tissue comprising the nucleic acids from the training individual.

31. The method of any one of claims 27-30, wherein performing the sequencing of nucleic acids comprises performing sequencing on nucleic acids in a sample obtained from the training individual.

32. The method of claim 31 , wherein the sample is one of a blood sample, a serum sample, a stool sample, a urine sample, a mucous sample, or a saliva sample.

33. The method of any one of claims 27-32, wherein performing the sequencing of nucleic acids comprises performing one of targeted, exome, transcriptome, or whole genome sequencing.

34. The method of any one of claims 20-33, wherein the set of training neoantigens comprises between 100 and 300 training neoantigens,35. The method of claim 34, wherein the set of training neoantigens comprises 200 training neoantigens.

36. The method any one of claims 20-35, wherein sampling the plurality of training neoantigens comprises analyzing biochemical properties of the training neoantigens and assigning immunogenicity scores based on the analyzed biochemical properties.Attorney Docket No. N.093.W0.0137. The method of claim 36, wherein the set of training neoantigens has assigned immunogenicity scores reflecting stronger likelihood of immunogenicity in comparison to other neoantigens not included in the set of training neoantigens.

38. The method of any one of claims 20-37, wherein augmenting the set of training neoantigens to generate the plurality of training neoantigen instances comprises:iteratively generating training neoantigen instances of the plurality, wherein each iteration generates a training neoantigen instance that samples a larger number of training neoantigens in comparison to a training neoantigen instance of a prior iteration.

39. The method of claim 38, wherein for a / th iteration, generating the training neoantigen instance comprises samples n;- = 20 + 5 / training neoantigens.

40. The method of any one of claims 20-39, further comprising associating the plurality of training neoantigen instances with a common ground truth label.

41. A method of preparing a vaccine composition based on the candidate neoantigen identified using the method of any one of claims 1-40, comprising:receiving a digital synthesis instruction identifying the candidate neoantigen; and manufacturing the vaccine composition comprising neoantigen components corresponding to the candidate neoantigen.

42. The method of claim 41, further comprising distributing the manufactured vaccine composition as a unit-dose preparation to a clinical site.

43. The method of claims 41 or claim 42, wherein the vaccine composition comprises a peptide vaccine, an mRNA vaccine, a DNA vaccine, or a combination thereof.

44. A circuit for a prediction model, the circuit comprising:one or more sub-circuitries each designed for a sub-structure for analyzing a candidate neoantigen instance and generating a candidate neoantigen representation, each of the one or more sub-circuitries comprising:a plurality of neurons comprising a plurality of microprocessors;a plurality of synaptic circuits including memories for storing synaptic weights, wherein each neuron is connected to at least one other neuron of the plurality of neurons via one of the plurality of synaptic circuits;Attorney Docket No. N.093.W0.01a sub-circuitry designed for a pooling structure for pooling candidate neoantigen representations into a single individual-level representation, the sub-circuitry for the pooling structure comprising:a plurality of neurons comprising a plurality of microprocessors; a plurality of synaptic circuits including memories for storing synaptic weights, wherein each neuron is connected to at least one other neuron of the plurality of neurons via one of the plurality of synaptic circuits; anda sub-circuitry designed for a classification structure for generating a classification using the single individual-level representation.

Images