Method of determining mhc-presented peptides

Abstract

A method, implemented by at least one processor, for determining whether or not a peptide query sequence is a major histocompatibility complex (MHC)-presented peptide, includes: collecting multiallelic data that includes (i) a plurality of known MHC-presented peptide sequences and (ii) a list of MHC-alleles; training a machine learning model using the multiallelic data to generate a trained machine learning model; inputting the peptide query sequence into the trained machine learning model, the peptide query sequence corresponding to a peptide presented by a specific MHC-allele to output a score, the specific MHC-allele not being included in the list of MHC-alleles; based on the score being less than or equal to a predetermined value, determining that the peptide query sequence corresponds to an MHC- presented peptide; and based on the score greater less than the predetermined value, determining that the peptide query sequence does not correspond to an MHC-presented peptide.

Description

[0001] Method of determining MHC-presented peptides

[0002] The present disclosure relates to a method for determining that a peptide query sequence is a MHC-presented peptide.

[0003] Background

[0004] The major histocompatibility complex (MHC) plays a critical role in the immune system of vertebrates, including humans. The MHC genes encode proteins that are present on the surface of cells and are responsible for the presentation of antigens to the immune system. Proteins are constantly synthesized and proteosomally degraded within cells and short peptides derived from these degraded proteins are presented by the MHC molecules on the cell surface. These MHC-presented peptides can be detected by T lymphocytes with their T cell receptor (TCR) and non- self antigens may initiate immune reactions.

[0005] The entirety of MHC-presented peptides is referred to as the immunopeptidome, which is highly complex and dynamic, varying with the type of cell, its developmental stage, and the presence of diseases or infections. The total number of unique MHC-presented peptides of an individual is estimated to be in the millions.

[0006] MHC-presented peptides are of great interest as targets for immunotherapy to initiate an immune response against specific cells characterized by certain MHC-presented peptides. In this regard it may be desirable to identify peptides that are presented by the MHC.

[0007] Initially peptide-MHC binding affinity (BA) was determined to study binding preferences of different MHC molecules. BA assay are typically limited to a single peptide and are therefore laborious. Recent advances in liquid chromatography mass spectrometry (LC- MS / MS) have opened a new chapter in immunopeptidomics. Several thousand of MHC- associated peptides can be sequenced in a single experiment resulting in a significant increase in available immunopeptidomic data.

[0008] In addition to the experimental identification of MHC-presented peptides, there are also advances to predict if a peptide is a MHC-presented peptide based on its amino acid sequence. Experimental data is hereby used to train methods of predicting if a peptide is an MHC- presented peptide. Such predictors have previously been trained with data derived from BA assays and / or large scale immunopeptidomic data, e.g. in Alvarez B. et al. (Mol Cell Proteomics 18, 2459-2477 (2019)); Reynisson B. et al. (Nucleic Acids Res 48, gkaa379-); Bulik- Sullivan, B. etal. (Nat Biotechnol 37, 55-63 (2018)); Abelin, J. G. et al. (Immunity 46, 315-326 (2017)); Sarkizova, S. et al. (Nat Biotechnol 38, 199-209 (2019)); Rammensee, H. G. et al. (Curr Opin Immunol 41, 178-228 (1995)); Rammensee, H. et al. (Immunogenetics 50, 213-9 (1999)); O’Donnell, T. J. et al. (Cell Syst 7, 129-132 e4 (2018)); O’Donnell, T. J. et al. (Cell Syst 11, 42-48 e7 (2020)). For example, mono-allelic cell lines were used for prediction of HLA epitopes; see e.g. Sarkizova, S. et al. (Nat Biotechnol 38, 199-209 (2019). However, this approach requires the use of genetically engineered cells.

[0009] Summary

[0010] The disclosed embodiments provide methods for determining if a peptide query sequence is a MHC-presented peptide, wherein a predictor is trained only with multiallelic data on MHC-presented peptides. In the herein provided examples, it is shown that the herein provided methods provide a reliable prediction and even provide a more precise prediction compared to methods of the prior art. This reduction in false positive findings is of particular importance in a clinical setting wherein a false positive finding may have severe consequences.

[0011] The disclosed methods provide further advantages: The disclosed methods can be applied on typically obtained experimental datasets such as e.g. immunopeptidome data that is multiallelic data and that can be used directly for training of the mathematical model in the disclosed embodiments without requiring previous training with e.g. monoallelic data, i.e. the training can be made exclusively with multiallelic data. Thus, peptide sequences determined from samples obtained from patients (e.g. tissue samples), for example, cancer patients or healthy subjects, can be used for training of the herein provided methods. In addition, the exclusive use of multiallelic data without pre-processing (experimental or otherwise) has not previously been described or suggested for training a mathematical model that can determine if a peptide sequence is a MHC-presented peptide.

[0012] In a first aspect, the disclosed embodiments provide a method for determining that a peptide query sequence is a MHC-presented peptide, wherein the method comprises the following steps:

[0013] (I) obtaining a trained mathematical model, wherein the data used to train the mathematical model comprises positive data, which is indicative of MHC- presented peptides, wherein a list of MHC-alleles that may present the MHC- presented peptide is provided for each MHC-presented peptide, wherein the specific MHC-allele presenting the peptide is not provided and wherein the positive data indicative of MHC-presented peptides, that is used to train the mathematical model is multiallelic data, and wherein the positive data indicative of MHC-presented peptides contains only MHC-presented peptides identified in more than one sample;

[0014] (II) providing at least one peptide query sequence and utilizing the trained mathematical model to determine for each peptide query sequence a score indicating that the at least one peptide query sequence is an MHC-presented peptide.

[0015] In a second aspect, the disclosed embodiments provide a method for generating a mathematical model for determining that a peptide query sequence is a MHC-presented peptide, wherein the method comprises training a mathematical model with data that comprises positive data, which is indicative of MHC-presented peptides, wherein a list of MHC-alleles that may present the MHC-presented peptide is provided for each MHC-presented peptide, wherein the specific MHC-allele presenting the peptide is not provided and wherein the positive data indicative of MHC-presented peptides, that is used to train the mathematical model is multiallelic data and wherein the positive data indicative of MHC-presented peptides contains only MHC-presented peptides identified in more than one sample; wherein the trained mathematical model can be used to determine if the peptide query sequence is a MHC-presented peptide.

[0016] In a third aspect, the disclosed embodiments provide a method for preparing an immunogenic composition comprising at least one peptide determined to be a MHC-presented peptide, wherein the method comprises the following steps:

[0017] (A) obtaining a trained mathematical model, wherein the data used to train the mathematical model comprises positive data, which is indicative of MHC- presented peptides, wherein a list of MHC-alleles that may present the MHC- presented peptide is provided for each MHC-presented peptide, wherein the specific MHC-allele presenting the peptide is not provided and wherein the positive data indicative of MHC-presented peptides, that is used to train the mathematical model is multiallelic data and wherein the positive data indicative of MHC-presented peptides contains only MHC-presented peptides identified in more than one sample;

[0018] (B) providing at least one peptide query sequence and utilizing the trained mathematical model to determine if the at least one peptide query sequence is an MHC-presented peptide; (C) including at least one MHC-presented peptide identified in step (B) into the immunogenic composition.

[0019] Brief Description of the Drawings

[0020] In the following, the content of the figures comprised in this specification is described. In this context please also refer to the detailed description above and / or below.

[0021] Figure 1: Workflow for Bayesian approach refers to an outline of the Bayesian approach described in Example 3.

[0022] Figure 2: depicts a flowchart of the model architecture. Specifically, the trained neural network is indicated in the grey box.

[0023] Figure 3: depicts weight update steps during neural network model training for individual multi-allelic training examples.

[0024] Figure 4: refers to a comparison of the method with the state of the art method MHCflurry. The method is exemplified using a PSSM as a mathematical model (herein PSSM) and using a neural network as mathematical model (herein Neural Network). (A) Boxplot of the area-under-the-curve (AUC) for the receiver operator curve (ROC); (B) boxplot indicating the average precision of the compared methods; (C) boxplot indicating the PPV @ 40% Sensitivity metric demonstrating the average precision values at a specific recall threshold (PPV = positive-predictive value = precision, sensitivity = recall).

[0025] Figure 5: refers to Precision-Recall curves comparing the method with the state of the art method MHCflurry for selected HLA alleles (HLA-A*25:01, HLA-B*08:01, HLA- C*01:02).

[0026] Figure 6: refers to a table of underlying metric values for figure 5.

[0027] Figure 7: refers to a comparison of the method of the present invention with the state of the art method MHCflurry. The method of the present invention is exemplified using a PSSM as a mathematical model (herein PSSM) and using a neural network as mathematical model (herein Neural Network). (A) Boxplot of the area-under-the-curve (AUC) for the receiver operator curve (ROC); (B) boxplot indicating the average precision of the compared methods; (C) boxplot indicating the PPV @ 40% Sensitivity metric demonstrating the average precision values at a specific recall threshold (PPV = positive-predictive value = precision, sensitivity = recall).

[0028] Figure 8: A to C refer to Precision-Recall curves comparing the method of the present invention with the state of the art method MHCflurry for selected HLA alleles ((A) HLA- A*02:01, (B) HLA-B*08:01, (C) HLA-C*07:04) Figure 9: refers to a table of underlying metric values for figure 8.

[0029] Detailed Description

[0030] Before the disclosed embodiments are described in detail below, it is to be understood that is the disclosed embodiments are not limited to the particular methodology, protocols and reagents described herein as these may vary. It is also to be understood that the terminology used herein is for the purpose of describing particular embodiments only, and is not intended to limit the scope of the disclosed embodiments which will be limited only by the appended claims. Unless defined otherwise, all technical and scientific terms used herein have the same meanings as commonly understood by one of ordinary skill in the art.

[0031] Preferably, the terms used herein are defined as described in "A multilingual glossary of biotechnological terms: (IUPAC Recommendations)", Leuenberger, H.G.W, Nagel, B. and Klbl, H. eds. (1995), Helvetica Chimica Acta, CH-4010 Basel, Switzerland).

[0032] Throughout this specification and the claims which follow, unless the context requires otherwise, the word "comprise", and variations such as "comprises" and "comprising", will be understood to imply the inclusion of a stated integer or step or group of integers or steps but not the exclusion of any other integer or step or group of integers or steps. In the following passages, different aspects of the disclosed embodiments are defined in more detail. Each aspect so defined may be combined with any other aspect or aspects unless clearly indicated to the contrary. In particular, any feature indicated as being optional, preferred or advantageous may be combined with any other feature or features indicated as being optional, preferred or advantageous.

[0033] Several documents are cited throughout the text of this specification. Each of the documents cited herein (including all patents, patent applications, scientific publications, manufacturer's specifications, instructions etc.), whether supra or infra, is hereby incorporated by reference in its entirety. Nothing herein is to be construed as an admission that the application is not entitled to antedate such disclosure by virtue of prior invention. Some of the documents cited herein are characterized as being “incorporated by reference” . In the event of a conflict between the definitions or teachings of such incorporated references and definitions or teachings recited in the present specification, the text of the present specification takes precedence.

[0034] In the following, the elements of the disclosed embodiments will be described. These elements are listed with specific embodiments; however, it should be understood that they may be combined in any manner and in any number to create additional embodiments. The variously described examples and preferred embodiments should not be construed to limit the disclosed embodiments to only the explicitly described embodiments. This description should be understood to support and encompass embodiments which combine the explicitly described embodiments with any number of the disclosed and / or preferred elements. Furthermore, any permutations and combinations of all described elements in this application should be considered disclosed by the description of the present application unless the context indicates otherwise.

[0035] Definitions

[0036] In the following, some definitions of terms frequently used in this specification are provided. These terms will, in each instance of its use, in the remainder of the specification have the respectively defined meaning and preferred meanings.

[0037] As used in this specification and the appended claims, the singular forms "a", "an", and "the" include plural referents, unless the content clearly dictates otherwise.

[0038] The term “about” when used in connection with a numerical value is meant to encompass numerical values within a range having a lower limit that is 5% smaller than the indicated numerical value and having an upper limit that is 5% larger than the indicated numerical value.

[0039] The “major histocompatibility complex” (MHC) as used in this disclosure is a set of cell surface proteins essential for the acquired immune system to recognize foreign molecules in vertebrates, which in turn determines histocompatibility. The main function of MHC molecules is to bind to protein fragments and display them on the cell surface for recognition by the appropriate T cells. The human MHC is also called the HLA (human leukocyte antigen) complex (often just the HLA). Preferably the MHC in the disclosed method is HLA. The MHC gene family is divided into three subgroups: class I, class II, and class III. Complexes of peptide and MHC class I are recognized by CD8-positive T cells bearing the appropriate T cell receptor (TCR), whereas complexes of peptide and MHC class II molecules are recognized by CD4- positive-helper-T cells bearing the appropriate TCR. Since both types of response, CD8 and CD4 dependent, contribute jointly and synergistically, the identification and characterization of MHC-presented peptides and corresponding T cell receptors is important in the development of immunotherapies such as vaccines and cell therapies. HLA-class I (or MHC-class I) are subdivided in A, B and C (i.e. HLA- A, HLA-B and HLA-C). Since each HLA has a different affinity for peptides of certain structures, greater variety of HLAs means greater variety of antigens to be 'presented' on the cell surface. Each individual can express up to two types of HLA-A (i.e. HLA-A alleles), one from each of their parents. Some individuals will inherit the same HLA-A allele from both parents, decreasing their individual HLA diversity; however, the majority of individuals will receive two different copies of HLA-A. In other words, every single person can only express either one or two of the over 7000 known HLA-A alleles. This same pattern follows for all HLA groups (e.g. HLA-B and -C).

[0040] A preferred MHC class I HLA protein may be an HLA-A, HLA-B or HLA-C protein, more preferably HLA-A protein, even more preferably the allotype group HLA-A*02, most preferably the specific allotype HLA-A*02:01. Other preferred HLA-A allotypes include HLA- A*24:02, HLA-A*01:01, HLA-A*03:01, HLA-B*07:02, HLA-B*08:01, HLA-B*44:02 and HLA-B*44:03.

[0041] “HLA-A*02:01” signifies a specific HLA allele or allotype, wherein the letter A signifies the gene and the suffix “02” indicates the allele or allotype group and “01” a specific HLA allele or allotype. The rules for nomenclature of HLA proteins is well known in the art and can for example be found at https: / / hla.alleles.org / nomenclature / naming.html.

[0042] In the MHC class I dependent immune reaction, peptides not only have to be able to bind to certain MHC class I molecules, they subsequently also have to be recognized by T cells bearing specific T cell receptors (TCR).

[0043] A “MHC-presented peptide” is thus a peptide that is presented by a MHC molecule. Preferably a MHC class I presented peptide has a length of 8 to 11 amino acids, preferably 9 to 10, most preferably 9 amino acids. Preferably a MHC class II presented peptide has a length of 13 to 25 amino acids.

[0044] The term “sample” as used herein refers to a sample of a donor. A donor may be a human, in particular a patient (e.g. a subject suffering from cancer and / or tumor) or a healthy subject, wherein a healthy subject may also be a subject that does not suffer from cancer and / or tumorous disease / disorder but may suffer from other diseases. Preferably, the sample may refer to a tissue sample, e.g. a tumor sample, blood sample, or a multiallelic cell line.

[0045] The term “immunopeptidome” as used herein refers to the set of peptides presented by major histocompatibility complex (MHC) molecules on the surface of cells to enable T-cell immuno surveillance. The immunopeptidome is thus a collection of all MHC-presented peptides in a particular sample or organism. The immunopeptidome can be further subdivided into sets of peptides presented by specific MHC molecules (or HLA- molecules in the human context). For example, the HLA-A*02 immunopeptidome would only encompass the peptides presented by HLA-A*02 molecules. Likewise, the HLA-A*02:01 immunopeptidome would only encompass the peptides presented by HLA-A*02:01 molecules. In a preferred embodiment the immunopeptidome refers to all MHC (or HLA) presented peptides. In another preferred embodiment the immunopeptidome is restricted to peptides presented by the HLA molecule selected from: HLA-A*02:01, HLA-A*24:02, HLA-A*01:01, HLA-A*03:01, HLA-B*07:02, HLA-B*08:01, HLA-B*44:02 and HLA-B*44:03.

[0046] The term “mathematical model” as used herein refers to mathematical models (e.g., machine learning models) that can be trained to recognize specific conditions (e.g. being MHC- presented peptides) in a query dataset. Therefore, the mathematical model may be trained first to be able to recognize these specific conditions. This is typically achieved by providing a training dataset that contains examples of the specific conditions to be recognized (i.e. positive dataset). Some mathematical models (e.g. neural networks) the training dataset may further include negative data that are not indicative of the specific condition to be recognized. The disclosed mathematical model can be trained to identify peptides that are MHC presented peptides. Therefore, the mathematical model is trained with positive data of MHC-presented peptides and optionally with negative data not indicative or associated with MHC-presented peptides. The training then allows the mathematical model to identify criteria in the peptides of the training dataset that are indicative of MHC-presented peptides. These criteria can then be applied by the trained mathematical model to a peptide sequence to determine if this peptide is a MHC-presented peptide. Non-limiting examples of mathematical models that can be used in the disclosed methods are a position- specific scoring matrix (PSSM) (preferably a PSSM trained by an iterative Bayesian approach)), deep learning algorithms (preferably neural networks), support- vector machines and / or random forests.

[0047] Disclosed embodiments may involve use of one or more computers. A computer as used in this disclosure may include a general purpose computer, a personal computer, a workstation, a mainframe computer, a notebook, a global positioning device, a laptop computer, a smart phone, a personal digital assistant, a network server, and any other electronic device that may interact with a user to develop programming code.

[0048] In some embodiments, any of the one or more computers may include at least one processor, a memory, and other components including those components that facilitate electronic communication. Other components may include user interface devices such as an input and output devices. Any of the one or more computers may include computer hardware components such as a combination of Central Processing Units (CPUs) or processors, buses, memory devices, storage units, data processors, input devices, output devices, network interface devices, and other types of components that will become apparent to those skilled in the art. Any of the one or more computers may further include application programs that may include software modules, sequences of instructions, routines, data structures, display interfaces, and other types of structures that execute operations of the present disclosure.

[0049] The at least one processor may be implemented in hardware, firmware, or a combination of hardware and software. The at least one processor may be one or more of a central processing unit (CPU), a graphics processing unit (GPU), an accelerated processing unit (APU), a microprocessor, a microcontroller, a digital signal processor (DSP), a field-programmable gate array (FPGA), an application- specific integrated circuit (ASIC), and another type of processing component. The at least one processor may include one or more processors capable of being programmed to perform a function.

[0050] The memory may include a random access memory (RAM), a read only memory (ROM), and / or another type of dynamic or static storage device (e.g., a flash memory, a magnetic memory, and / or an optical memory) that stores information and / or instructions for use by the at least one processor. The memory may also store information and / or software related to the operation and use of the one or more computers. For example, the memory may include a hard disk (e.g., a magnetic disk, an optical disk, a magneto-optic disk, and / or a solid state disk), a compact disc (CD), a digital versatile disc (DVD), a floppy disk, a cartridge, a magnetic tape, and / or another type of non-transitory computer-readable medium, along with a corresponding drive.

[0051] The one or more computers may perform one or more processes described herein. The one or more computers may perform operations based on the at least one processor executing software instructions stored by a non-transitory computer-readable medium, such as the memory. A computer-readable medium is defined herein as a non-transitory memory device. A memory device includes memory space within a single physical storage device or memory space spread across multiple physical storage devices.

[0052] Software instructions may be read into the memory from another computer-readable medium or from another device via, for example, one or more transceivers. When executed, software instructions stored in the memory may cause the at least one processor 118 to perform one or more processes described herein.

[0053] Additionally, or alternatively, hardwired circuitry may be used in place of or in combination with software instructions to perform one or more processes described herein. Thus, embodiments described herein are not limited to any specific combination of hardware circuitry and software.

[0054] Disclosed embodiments may involve receiving through a communication network. A communication network as used in this disclosure may include a set of computers (such at least some of the one or more computers) sharing resources located on or provided by network nodes. This set of computers may use common communication protocols over digital interconnections to communicate with each other. These interconnections may be made up of telecommunication network technologies, based on physically wired, optical, and wireless radio -frequency methods that may be arranged in a variety of network topologies. For example, these interconnections may take place through databases, servers, RF (radio frequency) signals, cellular technology, Ethernet, telephone, “TCP / IP” (transmission control protocol / internet protocol), and any other electronic communication format. For example, the network 130 may include a cellular network (e.g., a fifth generation (5G) network, a long-term evolution (LTE) network, a third generation (3G) network, a code division multiple access (CDMA) network, etc.), a public land mobile network (PLMN), a local area network (LAN), a wide area network (WAN), a metropolitan area network (MAN), a telephone network (e.g., the Public Switched Telephone Network (PSTN)), a private network, an ad hoc network, an intranet, the Internet, a fiber optic-based network, or the like, and / or a combination of these or other types of networks.

[0055] The number and arrangement of computers and networks may be adjustable.

[0056] In some embodiments, the communications network may be set up as a neural network. A neural network may be based on a collection of connected units or nodes called artificial neurons, which loosely model the neurons in a biological brain. Each connection, like the synapses in a biological brain, may transmit a signal to other neurons. An may artificial neuron receive signals to process and may then signal other neurons connected to it. These signals at a connection may be real numbers, and the output of each neuron may be computed by some nonlinear function of the sum of its inputs. These connections may be edges. Neurons and edges may have a weight that adjusts as learning proceeds. The weight may increase or decrease the strength of the signal at a connection. Neurons may have a threshold such that a signal may be sent only if the aggregate signal crosses that threshold. Neurons may be aggregated into layers. Different layers may perform different transformations on their inputs. Signals may travel from a first layer (e.g., an input layer), to a last layer (e.g., an output layer), through potential intermediate layers and may do so multiple times.

[0057] In some embodiments the mathematical model comprises more than one (e.g. 1, 2, 3, etc.) neural networks.

[0058] The term “positive data” as used herein refers to training data that is representative for the condition(s) that the mathematical model should be trained on. The positive data used in the disclosed methods is multiallelic data, i.e. the positive data contains information on MHC- presented peptides of different MHC-alleles. Typically, the multiallelic data is not deconvo luted, i.e. the MHC-presented peptides are not assigned to a specific MHC-allele.

[0059] Positive data can be ‘indicative’ of the condition to be identified by the trained mathematical model, which means that it contains data indicating the condition to be identified (e.g. being MHC-presented peptides). For example, a mathematical model trained to identify if a specific peptide is a MHC-represented peptide can be trained with positive data that is indicative of MHC-presented peptides. Such positive data would contain a multitude of peptide sequences that are known to be MHC-presented peptides (i.e. indicative of being MHC represented). The indicative positive data also contains a list of MHC-alleles that may present the peptides in the positive data. This list of MHC-alleles does not contain all possible MHC- alleles, but a selection thereof. This selection of MHC-alleles can be determined by a number of alternative methods. One example is to provide MHC-typing of the donor samples (i.e. determined the MHC-alleles present in this sample) in which the positive data is determined from. In a preferred embodiment, the sample in which the positive data is determined is subject to a pre-selection by MHC-affinity purification. Preferably, the MHC-affinity purification is achieved by using antibodies specific for a certain type or (sub)class of MHC-alleles (i.e. MHC- class I or II, HLA-A, -B, -C etc.). Preferably the allele coverage of the positive data being indicative of MHC-presented peptides is about the same as the allele distribution within the population of which the peptide query sequence is obtained from.

[0060] Another type of positive data that can be used in addition to indicative positive data is positive data ‘associated’ with MHC-presented peptides. Such associated positive data would not refer to peptides known to be MHC-represented themselves, but would rather be associated with the indicative positive data. For example, a certain peptide sequence is comprised within the indicative positive data and the associated positive data would contain a quantification of this peptide or transcriptomic data of the indicative positive data. Associated positive data could also refer to other properties of the biological samples of which the indicative positive data is derived from. Non-limiting examples are MHC-alleles present, disease status of biological sample; ploidy of biological sample, mutation status of particular genes, etc.

[0061] The term “immunoproteasome” refers to a specialized version of the proteasome, a cellular complex responsible for degrading proteins that are no longer needed or are damaged. By generating peptides that can be presented by major histocompatibility complex (MHC) molecules, the immunoproteasome plays a critical role in antigen processing and presentation to T cells. The term “position- specific scoring matrix (PSSM)” as used herein, is a motif descriptor. The PSSM is trained with positive data indicative of MHC-presented peptides to capture the intrinsic variability characteristic of sequence patterns (e.g. such as those of MHC presented peptides). Thereby the trained PSSM can be used to determine if a given peptide query sequence is a MHC-presented peptide. A PSSM can also be trained iteratively by applying several consecutive rounds of learning to the PSSM. Typically, the PSSM resulting from a previous round of training is further trained with a positive dataset to refine the PSSM.

[0062] The term “neural network” as used herein refers to a mathematical data modeling tool. A neural network can be trained with complex data to ‘learn’ intrinsic patterns in training data (e.g. such as those of MHC-presented peptides). This is particularly useful for large and / or complex data, which makes it impractical or even impossible to identify intrinsic patterns by hand.

[0063] The anti-HLA antibodies W6 / 32, BB7.2, GAP- A3 and B.1.23.2 can be used within the disclosed methods, in particular for affinity purification of samples in which HLA- (or MHC-) presented peptides are determined (i.e. the indicative positive data). These antibodies are further described in Apps et al (Immunology; 2009 May; 127(l):26-39), Parham & Brodsky (Hum. Immol; 1981 Dec;3(4):277-99) and Berger et al (Hybridoma. 1982;l(2):87-90), which are incorporated herein by reference. Antibody W6 / 32 is a pan antibody that binds all class I HLA- alleles. Antibody BB7.2 specifically binds to all HLA-A*02 and -A*96 alleles. Antibody GAP- A3 specifically binds to all HLA-A*03 alleles. Antibody B.1.23.2 specifically binds to all HLA- B and HLA-C alleles as well as the following HLA-A alleles: A*23, A*25, A*26, A*29, A*30, A*31, A*32, A*33, A*34, A*43, A*66, A*74; (A*24 and A*80 weakly).

[0064] The term “deconvolution” or “motif deconvolution“ as used herein refers to the process of assigning the accurate MHC allele to each MHC-presented peptide, preferably for mass spectrometry derived data of MHC-presented peptides. Preferably the methods disclosed herein are deconvolution- free methods, i.e. they do not rely on deconvo luted data in particular for training of the mathematical model.

[0065] “Pharmaceutically acceptable” means suitable for use in animals, and more particularly in humans (e.g., as may be evidenced by being approved by a regulatory agency of the Federal or a state government or listed in the U.S. Pharmacopeia, European Pharmacopeia (Ph. Eur.) or other generally recognized pharmacopeia).

[0066] The term “carrier”, as used herein, refers to a diluent, adjuvant, excipient, or vehicle with which the therapeutic agent is administered. Such pharmaceutical carriers can be sterile liquids, such as saline solutions in water and oils, including those of petroleum, animal, vegetable or synthetic origin, such as peanut oil, soybean oil, mineral oil, sesame oil and the like. A saline solution is a preferred carrier when the pharmaceutical composition is administered intravenously. Saline solutions and aqueous dextrose and glycerol solutions can also be employed as liquid carriers, particularly for injectable solutions. Suitable pharmaceutical excipients include starch, glucose, lactose, sucrose, gelatin, malt, rice, flour, chalk, silica gel, sodium stearate, glycerol monostearate, talc, sodium chloride, dried skim milk, glycerol, propylene, glycol, water, ethanol and the like. The composition, if desired, can also contain minor amounts of wetting or emulsifying agents, or pH buffering agents. These compositions can take the form of solutions, suspensions, emulsions, tablets, pills, capsules, powders, sustained-release formulations and the like. The composition can be formulated as a suppository, with traditional binders and carriers such as triglycerides. The compounds can be formulated as neutral or salt forms. Pharmaceutically acceptable salts include those formed with free amino groups such as those derived from hydrochloric, phosphoric, acetic, oxalic, tartaric acids, etc., and those formed with free carboxyl groups such as those derived from sodium, potassium, ammonium, calcium, ferric hydroxides, isopropylamine, triethylamine, 2- ethylamino ethanol, histidine, procaine, etc. Examples of suitable pharmaceutical carriers are described in "Remington's Pharmaceutical Sciences" by E. W. Martin. Such compositions will contain a therapeutically effective amount of the compound, preferably in purified form, together with a suitable amount of carrier so as to provide the form for proper administration to the patient. The formulation should suit the mode of administration.

[0067] Embodiments

[0068] In the following, different aspects of the disclosed embodiments are described in more detail. Each aspect so described may be combined with any other aspect or aspects unless clearly indicated to the contrary. In particular, any feature indicated as being preferred or advantageous may be combined with any other feature or features indicated as being preferred or advantageous.

[0069] In a first aspect, the disclosed embodiments provide a method for determining that a peptide query sequence is a MHC-presented peptide, wherein the method comprises the following steps:

[0070] (I) obtaining a trained mathematical model, wherein the data used to train the mathematical model comprises positive data, which is indicative of MHC- presented peptides, wherein a list of MHC-alleles that may present the MHC- presented peptide is provided for each MHC-presented peptide, wherein the specific MHC-allele presenting the peptide is not provided and wherein the positive data indicative of MHC-presented peptides, that is used to train the mathematical model is multiallelic data and wherein the positive data indicative of MHC-presented peptides contains only MHC-presented peptides identified in more than one sample;

[0071] (II) providing at least one peptide query sequence and utilizing the trained mathematical model to determine for each peptide query sequence a score indicating that the at least one peptide query sequence is an MHC-presented peptide.

[0072] In some embodiments, the list of MHC-alleles may or may not present the MHC- presented peptide.

[0073] In a preferred embodiment the mathematical model is trained only with multiallelic data. The present method does not rely on monoallelic or binding data for the training of the mathematical model. In a preferred embodiment the mathematical model is trained only with multiallelic data.

[0074] In a preferred embodiment, the method does not require deconvolution of the data used to train the mathematical model.

[0075] In a preferred embodiment the mathematical model is trained only with mutliallelic data that has not been deconvo luted (i.e. assigned an MHC-allele).

[0076] In a second aspect, the disclosed embodiments provide a method for generating a mathematical model for determining that a query sequence is a MHC-presented peptide, wherein the method comprises training a mathematical model with data that comprises positive data, which is indicative of MHC-presented peptides, wherein a list of MHC-alleles that may present the MHC-presented peptide is provided for each MHC-presented peptide, wherein the specific MHC-allele presenting the peptide is not provided and wherein the positive data indicative of MHC-presented peptides, that is used to train the mathematical model is multiallelic data and wherein the positive data indicative of MHC-presented peptides contains only MHC-presented peptides identified in more than one sample; wherein the trained mathematical model can be used to determine if the peptide query sequence is a MHC-presented peptide.

[0077] In a preferred embodiment the mathematical model is trained only with multiallelic data.

[0078] The present method does not rely on monoallelic or binding data for the training of the mathematical model. In a preferred embodiment the mathematical model is trained only with multiallelic data. In a preferred embodiment, the method does not require deconvolution of the data used to train the mathematical model.

[0079] In a preferred embodiment the mathematical model is trained only with mutliallelic data that has not been deconvo luted (i.e. assigned an MHC-allele).

[0080] In a third aspect, the disclosed embodiments provide a method for preparing an immunogenic composition comprising at least one peptide determined to be a MHC-presented peptide, wherein the method comprises the following steps:

[0081] (A) obtaining a trained mathematical model, wherein the data used to train the mathematical model comprises positive data, which is indicative of MHC- presented peptides, wherein a list of MHC-alleles that may present the MHC- presented peptide is provided for each MHC-presented peptide, wherein the specific MHC-allele presenting the peptide is not provided and wherein the positive data indicative of MHC-presented peptides, that is used to train the mathematical model is multiallelic data and wherein the positive data indicative of MHC-presented peptides contains only MHC-presented peptides identified in more than one sample;

[0082] (B) providing at least one peptide query sequence and utilizing the trained mathematical model to determine if the at least one peptide query sequence is an MHC-presented peptide;

[0083] (C) including at least one MHC-presented peptide identified in step (B) into the immunogenic composition.

[0084] In some embodiments, the mathematical model does not require the use of monoallelic data for training of the mathematical model.

[0085] In other words, the trained mathematical model provides an output score for each input peptide query sequence. The score may be used to determine the likelihood that the peptide query sequence corresponds to a MHC-presented peptide.

[0086] In a preferred embodiment the mathematical model is trained only with multiallelic data. The present method does not rely on monoallelic or binding data for the training of the mathematical model. In a preferred embodiment the mathematical model is trained only with multiallelic data.

[0087] In a preferred embodiment, the method does not require deconvolution of the data used to train the mathematical model.

[0088] In a preferred embodiment the mathematical model is trained only with mutliallelic data that has not been deconvo luted (i.e. assigned an MHC-allele). In some embodiments, the disclosed method of all aspects of the invention is a computer-implemented method.

[0089] In some embodiments, the positive data indicative of MHC-presented peptides contains only MHC-presented peptides identified in more than one, more than two, more than three, or more than four, more than five, more than 10, more than 20, more than 30, more than 40, more than 50 samples (preferably more than two samples). Multiple detections of the MHC-presented peptides in different samples may be used to be able to determine which MHC-allele is assigned to the MHC-presented peptide. In general higher numbers of samples used for deducing MHC annotation is considered beneficial.

[0090] In some embodiments, the MHC presented peptides contained in the positive data indicative of MHC-presented peptides can be assigned to a specific MHC-allele by comparing the list of MHC-alleles provided for each MHC-presented peptide. The mathematical model can utilize the information provided in the list of MHC-alleles for each MHC-presented peptides by per-peptide allele frequency statistics. During model training all lists of MHC- alleles for a specific MHC-presented peptide may be presented as positive training examples to the mathematical model to learn from.

[0091] As shown in the appended examples and in preferred embodiments of the positive data indicative of MHC-presented peptides that is used to train the mathematical model, the specific MHC-presented peptides are identified in the samples of the donors. For clarity, each sample is obtained from one donor. The expressed MHC-alleles of the donors are determined, in other words MHC-typing is performed on the samples. The MHC-presented peptide is identified in more than one sample of the donors and the same MHC allele is expressed in more than one donor of the samples, wherein the MHC-presented peptides are identified. The more often the specific MHC-presented peptide is identified in the samples with the same MHC-alleles expressed, the more likely the specific MHC-presented peptide is presented by the MHC-allele.

[0092] In some embodiments, the positive data indicative of MHC-presented peptides that is used to train the mathematical model is obtained by a method comprising: a) Determining the expressed MHC-alleles of the donors of the samples, b) Identifying the MHC-presented peptides in the samples of the donors, c) Determining the specific MHC-presented peptides that are (i) identified in more than one sample of the donors, and preferably (ii) wherein the same MHC allele is expressed in more than one of the donors, and d) Selecting the specific MHC-presented peptide(s) of step (c) as the positive data. In a preferred embodiment, the score determined in step (II) is a percentile rank. In a preferred embodiment the percentile rank indicates a comparison between the peptide query sequence and a list of randomly chosen peptides (preferably at least 50,000, at least 100,000, at least 200,000, at least 500,000, at least 1,000,000 randomly chosen peptides). For example, a percentile rank of 0.5 would indicate that 99.5% of the random peptides have a lower rank. Likewise, a percentile rank of 2 or 5 would indicate that 98% or 95% of the random peptides have a lower rank. The percentile rank may therefore indicate if a particular peptide query sequence is a strong or a weak binder. In a preferred embodiment, a rank of 0.5 or less indicates a strong binder. In another preferred embodiment, a percentile rank of 2 to 0.5 indicates a normal binder. In another preferred embodiment, a percentile rank of 5 to 2 indicates a weak binder. In another preferred embodiment a percentile rank of more than 5 indicates that the peptide query sequence is not a binder. Preferably, the percentile rank indicates if the peptide query sequence is a MHC-presented peptide. In a preferred embodiment, a percentile rank of 0.5 or less, 2 or less, 5 or less (preferably 0.5 or less) indicates that the peptide query sequence is a MHC-presented peptide.

[0093] In a preferred embodiment, in step (II) the trained mathematical model determines for each MHC-allele included in the trained mathematical model a score that the at least one peptide query sequence is presented by said MHC-allele.

[0094] In a preferred embodiment the peptide query sequence has the same length as the MHC- presented peptides in the indicative positive data used for training. In another preferred embodiment the peptide query sequence has a length of 8 to 11 amino acids, preferably 9 to 10, most preferably 9 amino acids. In yet another preferred embodiment the peptide query sentence has a length of 13 to 25 amino acids. In another embodiment the peptide query sequence has a length of 9 to 25 amino acids. In a preferred embodiment, the peptide query sequence length is determined by the mathematical model used. Preferably the mathematical model requires a query sequence length of 8, 9, 10, 11, 12 or 13 amino acids, preferably 9, 10, 11 or 12 amino acids.

[0095] In some embodiments, the peptide query sequence is associated with a specific disease. Preferably the disease is cancer.

[0096] In a preferred embodiment the mathematical model (preferably a neural network) can be used with a peptide query sequence being shorter than required by the mathematical model, wherein the peptide query sequence is elongated by adding one or more space holders to the peptide query sequence to match the required peptide query sequence length. In a preferred embodiment the one or more space holders are added to the N-terminus of the peptide query sequence to match the required peptide query sequence length. In another preferred embodiment the one or more space holders are added to the C-terminus of the peptide query sequence to match the required peptide query sequence length. In yet another the one or more space holders are added in the middle of the peptide query sequence to match the required peptide query sequence length. In yet another embodiment, the one or more space holders are added to the C-terminus of the peptide query sequence to match the length required by the mathematical and separately the same numbers of space holders are added to the N-terminus of the unmodified peptide query sequence and these two modified peptide query sequences are then concatenated. For example, the peptide query sequence YLKVLPQEL (SEQ ID NO: 1) is modified to XXYLKVLPQEL (SEQ ID NO: 2) and YLKVLPQELXX (SEQ ID NO: 3). These two sequences are then concatenated to the sequence XXYLKVLPQELYLKVLPQELXX (SEQ ID NO: 4). Concatenating the two sequences with space holders is advantageous as it allows the model to learn fixed-position sub-motifs from both termini (e.g. L at second position and V at the last, independent of peptide length), without breaking the sequence apart.

[0097] In some embodiments, the list of MHC-alleles of step (III) is determined by MHC- typing of the specific sample. Methods to determine the MHC-alleles in a biological sample are known in the art.

[0098] In some embodiments, the method further comprises the steps:

[0099] (III) providing a list of MHC-alleles present in a specific sample;

[0100] (IV) combining the scores of step (II) with the list of MHC-alleles present in the specific sample to determine if the at least one peptide query sequence is an MHC-presented peptide in the specific sample.

[0101] In some embodiments, step (IV) comprises combining the scores of step (II) with the list of step (III) to determine if the at least one peptide query sequence is an MHC-presented peptide in the specific sample. Step (II) provides for each peptide query sequence scores for each MHC-allele included in the trained mathematical model that the peptide query sequence is presented by the MHC-allele. Step (IV) takes into consideration the MHC-alleles present in a specific sample, allowing to determine if the peptide query sequence is a MHC-presented in the specific sample. In other words, if the MHC-allele determined in step (II) is not on the list of MHC-alleles of step (III) the peptide query sequence may be a MHC-presented peptide but it is not presented in the specific sample due to the lack of the necessary MHC-allele in said specific sample. In some embodiments, step (B) further comprises to determine which MHC-allele is presenting the at least one peptide query sequence and providing a list of MHC-alleles present in a specific sample and selecting peptide query sequences that are presented by a MHC-allele present in the specific sample, wherein in step (C) only selected peptide query sequences are considered. This embodiment is particularly useful if the specific sample is obtained from a subject to be treated with the immunogenic composition to determine if the at least one peptide query is a MHC-presented peptide in this particular subject.

[0102] The disclosed method(s) can be used in many different settings. In a preferred embodiment the method(s) can be used for deconvolution of multiallelic data by assigning to each peptide in a multiallelic dataset an HL A- allele.

[0103] In a preferred embodiment, the disclosed method(s) can be used to assign a MHC-allele to each peptide query sequence. In a preferred embodiment, step (II) further comprises selecting for each peptide query sequence the MHC-allele with the score indicating the highest probability and assigning this MHC-alleles to the peptide query sequence.

[0104] In yet another preferred embodiment, the disclosed method(s) can be used in personalized immunotherapy. Preferably such a use encompasses a step (Ila) comprising providing further information of a specific patient. In a preferred embodiment step (Ila) comprises providing genomic, proteomic and / or transcriptomic data of a particular subject to determine if the peptide query sequence is a MHC-presented peptide in the particular subject.

[0105] In an alternative aspect an in vitro method is provided for determining if a peptide query sequence is an MHC-presented peptide in a subject. The method comprises the following steps:

[0106] (1) obtaining a biological sample from said subject;

[0107] (2) determine genomic, transcriptomic and / or proteomic data in said biological sample;

[0108] (2) obtaining a trained mathematical model according to the method of the second aspect of the invention (i.e. claim 3);

[0109] (3) providing at least one peptide query sequence and utilizing the trained mathematical model to determine if the at least one peptide query sequence is an MHC- presented peptide;

[0110] (4) determine if the at least one peptide query sequence being an MHC-presented peptide of step (3) is an MHC-presented peptide in said subject by utilizing the data obtained in step (2).

[0111] The genomic, transcriptomic and / or proteomic data of step (2) is typically comprising data that is suitable to determine if the at least one peptide query sequence is present in said subject. In a preferred embodiment the peptide query sequence is associated with a particular disease, preferably cancer.

[0112] In a preferred embodiment the mathematical model is trained only with multiallelic data. The present method does not rely on monoallelic or binding data for the training of the mathematical model. In a preferred embodiment the mathematical model is trained only with multiallelic data.

[0113] In a preferred embodiment, the method does not require deconvolution of the data used to train the mathematical model.

[0114] In a preferred embodiment the mathematical model is trained only with multiallelic data that has not been deconvo luted (i.e. assigned an MHC-allele).

[0115] In yet another preferred embodiment, the disclosed method(s) can be used to control the validity of an experimental dataset. An experimental dataset comprising MHC-presented peptides, such as an immunopeptidomic dataset, can be used as peptide query sequences. In case a significant portion of the MHC-presented peptides in the experimental dataset is not predicted by the trained model to be presented by the MHC-alleles present in the experimental sample, this indicates an error in the experimental dataset and / or the typing of the MHC-alleles in the experimental sample. In a preferred embodiment, the at least one peptide query sequence is an experimental dataset (preferably immunopeptidomic dataset). In a preferred embodiment the immunopeptidomic dataset is a mass spectrometry dataset.

[0116] In yet another preferred embodiment, the disclosed method(s) can be used to determine if a peptide query sequence is not an MHC-presented peptide. Therefore it is predicted if MHC- presented peptides detected in a multitude of experimental datasets can bind to the MHC-alleles present in the multitude of experimental datasets. In case the MHC-presented peptides are predicted for the majority of cases to not bind to the MHC-alleles present in the experimental datasets, this is indicative of a false positive, i.e. that the peptide found in the experimental datasets is actually not a MHC-presented peptide. In a preferred embodiment, the at least one peptide query sequence of step (II) has been determined to be a MHC-presented peptide in a multitude of experimental datasets (preferably immunopeptidomic data) and wherein in step (III) lists of the MHC-alleles present in the multitude of experimental datasets (preferably immunopeptidomic data) are provided, wherein in step (IV) the scores of step (II) with the lists of step (III) are combined to determine if the at least one peptide query sequence is an MHC- presented peptide in the multitude of experimental datasets (preferably immunopeptidomic data) and if in the majority of experimental datasets the at least one peptide query sequence is determined not to be a peptide presented by the present MHC-alleles, this is indicative of a false positive result in the multitude of experimental datasets (preferably immunopeptidomic data).

[0117] In yet another preferred embodiment, the disclosed method(s) can be used to determine off-target toxicity of compounds that target a specific MHC-presented peptide. In a preferred embodiment the peptide query sequence of step (II) is a variation of an MHC-presented peptide targeted by a compound, wherein if the peptide query sequence is determined to be a MHC- presented peptide (preferably in the specific sample) this is indicative of an off-target toxicity of the compound.

[0118] In yet another preferred embodiment, the disclosed method(s) can be used to determine the MHC-alleles present in a dataset (preferably an experimental dataset). In a preferred embodiment, the at least one peptide query sequence in step (II) is a multitude of MHC- presented peptide sequences of an experimental dataset (preferably an immunopeptidomic dataset), wherein for each peptide query sequence the MHC-allele presenting the peptide is determined and the 1-2 (preferably two) MHC-alleles of each MHC gene with the most peptide sequences assigned to are considered to be the 1-2 (preferably two) MHC-alleles present in the experimental dataset.

[0119] In yet another preferred embodiment, the disclosed method(s) can be used to predict if a compound comprising a peptide sequence can elicit an immune response, such as for example anti-drug antibodies. In a preferred embodiment, the at least one peptide query sequence of step (II) is one or more peptide sequence comprised within a compound, wherein if at least one of the peptide query sequences is determined to be a MHC-presented peptide, this is indicative of the compound eliciting an immune response, optionally further comprising steps (III) and (IV), wherein the specific sample is obtained from a subject and if the at least one peptide query sequence is an MHC-presented peptide in the specific sample this is indicative of the compound eliciting an immune response in the subject of which the specific sample was obtained.

[0120] In some embodiments, the mathematical model is selected from a neural network or a position- specific- scoring matrix (PSSM). In a preferred embodiment the mathematical model is a PSSM. In another preferred embodiment the mathematical model is a neural network.

[0121] A further preferred embodiment of the first aspect of the disclosed embodiments relates to a method for determining that a peptide query sequence is a MHC-presented peptide, wherein the method comprises the following steps:

[0122] (I) obtaining a trained mathematical model, wherein the data used to train the mathematical model comprises positive data, which is indicative of MHC- presented peptides, wherein a list of MHC-alleles that may present the MHC- presented peptide is provided for each MHC-presented peptide, wherein the specific MHC-allele presenting the peptide is not provided and wherein the positive data indicative of MHC-presented peptides that is used to train the mathematical model is multiallelic data, wherein preferably the mathematical model is selected from PSSM and / or a neural network (more preferably PSSM), and

[0123] (II) providing at least one peptide query sequence and utilizing the trained mathematical model to determine for each peptide query sequence a score indicating that the at least one peptide query sequence is an MHC-presented peptide.

[0124] A further preferred embodiment of the second aspect of the disclosed embodiments relates to a method for generating a mathematical model for determining that a peptide query sequence is a MHC-presented peptide, wherein the method comprises training a mathematical model with data that comprises positive data, which is indicative of MHC-presented peptides, wherein a list of MHC-alleles that may present the MHC-presented peptide is provided for each MHC-presented peptide, wherein the specific MHC-allele presenting the peptide is not provided and wherein the positive data indicative of MHC-presented peptides, that is used to train the mathematical model is multiallelic data; wherein the trained mathematical model can be used to determine if the peptide query sequence is a MHC-presented peptide, and wherein preferably the mathematical model is selected from PSSM and / or a neural network (more preferably PSSM).

[0125] A further preferred embodiment of the third aspect of the disclosed embodiments relates to a method for preparing an immunogenic composition comprising at least one peptide determined to be a MHC-presented peptide, wherein the method comprises the following steps:

[0126] (A) obtaining a trained mathematical model, wherein the data used to train the mathematical model comprises positive data, which is indicative of MHC- presented peptides, wherein a list of MHC-alleles that may present the MHC- presented peptide is provided for each MHC-presented peptide, wherein the specific MHC-allele presenting the peptide is not provided and wherein the positive data indicative of MHC-presented peptides, that is used to train the mathematical model is multiallelic data and wherein preferably the mathematical model is selected from PSSM and / or a neural network (more preferably PSSM); (B) providing at least one peptide query sequence and utilizing the trained mathematical model to determine if the at least one peptide query sequence is an MHC-presented peptide;

[0127] (C) including at least one MHC-presented peptide identified in step (B) into the immunogenic composition; optionally further including a pharmaceutically acceptable carrier.

[0128] In some embodiments, the mathematical model is a neural network comprising one or more (preferably 1-5, 2-4, 2-3) hidden layers.

[0129] In some embodiments, the mathematical model comprises more than one neural network. Preferably, 1-5, 2-4, 2-3 neural networks.

[0130] In some embodiments, the mathematical model is a neural network, wherein the input layer comprises one or more (preferably 10-500, 20-50, 20-21) nodes per amino acid position in the peptide query sequence. In another preferred embodiment the input layer comprises further nodes for associated positive data.

[0131] In some embodiments, the mathematical model is a neural network e.g. consisting of one input layer, one hidden layer and one output layer. In a preferred embodiment the hidden layer of the neural network has 8192 nodes. In yet another preferred embodiment the trained neural network consists of a single fully connected hidden layer of 8192 ReLU units (without bias), followed by a fully connected allele score layer with one sigmoid unit (with bias) for each of the MHC-alleles expressed in the training samples.

[0132] In some embodiments, the neural network is a feedforward artificial neural network, preferably a multilayer perceptron.

[0133] In some embodiments, a portion (preferably 1% to 25%, preferably 5%-20%, 10%-20% ) of the positive training data (indicative positive data and optionally associated positive data) is not used for training of the mathematical model but is instead used for validation of the trained mathematical model.

[0134] In a preferred embodiment, the portion of positive data not used for training is mixed with negative data, preferably in a ratio that reflects the rarity of peptides being presented by a specific MHC-allele. In a preferred embodiment, the ratio of positive to negative data in the validation data is 1:50, 1:75 or 1:100. In another preferred embodiment the portion of the positive training data not used for training is determined by the fold of cross-validation, which is preferably 3 to 10 fold cross validation. For example, 4-fold cross-validation would require that one quarter (25%) of the positive data used for training is used for validation of the trained model. Accordingly 5-fold cross validation would require to use 20% of the positive data for validation. In another preferred embodiment, the mathematical models (preferably neural networks) obtained from cross validation are combined into an ensemble model. Preferably an average over the predictions of different mathematical models is computed.

[0135] In some embodiments, the mathematical model is a neural network, trained with multi- allelic positive and negative data. In a preferred embodiment, the positive data is indicative of MHC-presented peptides and optionally further comprises positive data associated with MHC- presented peptides. In another preferred embodiment the multiallelic negative data is not indicative of or associated with the MHC-presented peptides.

[0136] In a preferred embodiment of all aspects of the invention the mathematical model is trained only with multiallelic or non-allelic data. Non-allelic data includes any type of data that does not contain any information on MHC-alleles (or HLA- alleles in the human context). Nonlimiting examples on non-allelic is associated positive (or negative) data that does not contain information on MHC-alleles.

[0137] In some embodiments, step (IV) comprises the use of a weighted average based on the list of MHC-alleles in the specific sample, wherein the weighted average is applied to the score of step (II) to determine if the peptide query sequence is a peptide presented by a specific MHC- allele. In other words, the MHC-alleles that are detected in a specific sample are given more weight than MHC-alleles not being detected in the specific sample.

[0138] In some embodiments, the samples used to generate the positive data indicative of MHC-presented peptides are subject to MHC-affinity purification, wherein step (IV) comprises the use of a weighted average based on the MHC-alleles enriched by the MHC-affinity purification. In other words, MHC-affinity purification enriches certain MHC-alleles depending on the binding specificity of the used affinity binder (preferably an antibody). Those enriched MHC-alleles are then favored by using a weighted average that takes the enriched MHC-alleles into consideration.

[0139] This is based on the idea that that MHC-presented peptides that are presented by a MHC- allele not being detected in the specific sample is less likely to be an MHC-presented peptide in the specific sample.

[0140] In some embodiments, the mathematical model is a position- specific- scoring matrix (PSSM). The PSSM is trained with indicative positive data that is multiallelic. In a preferred embodiment the PSSM is trained with one positive dataset. In another preferred embodiment, an iterative Bayesian training is used for the PSSM. An iterative Bayesian training consists of several rounds of training, wherein the PSSM resulting from a first training is used for a further training. Thereby the trained PSSM is further improved with regard to its capability to detect the specific condition it is trained for (e.g. predicting MHC-presented peptides). In a preferred embodiment up to 10, up to 15 or up to 20 iterative rounds of training are applied to obtain a trained PSSM. In another preferred embodiment, 1-20, 2-15, 5-10 rounds of iterative training are performed. In yet another preferred embodiment 1-20 (e.g. 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19 or 20) rounds, preferably 15 rounds, of iterative training are performed. In general the number of iterative training rounds is a trade-off between runtime required for training and the performance of the trained mathematical model. In a preferred embodiment the PSSM is trained in iterative training rounds until the performance of the PSSM is no longer increased significantly by further iterative training rounds. Preferably the performance of the PSSM is not increase by more than 5%, 4%, 3%, 2%, 1% or less than 1% by the last iterative training round.

[0141] In a preferred embodiment wherein the mathematical model is a PSSM, the positive data, which is indicative of MHC-presented peptides, used for training the model is deconvo luted using an iterative Bayesian approach. In other words, the Bayesian approach uses the frequency of a peptide across all MHC alleles, i.e. the more often a peptide is found in a dataset with specific MHC allele(s), the more likely the presentation of this peptide by the respective MHC-allele is.

[0142] In a preferred embodiment wherein the mathematical model is a neural network, the positive data, which is indicative of MHC-presented peptides, used for training the model is not deconvo luted prior to training of the neural network.

[0143] In a preferred embodiment the positive data indicative of MHC-presented peptides used to train the mathematical model is exclusively multi-allelic data. In other words the positive data indicative of MHC-presented peptides does not contain mono-allelic data or mixed multi- allelic and mono-allelic data.

[0144] In a preferred embodiment, the method is using a monolithic architecture, i.e. comprising only a single neural network.

[0145] In some embodiments, the positive data comprises a list of MHC-alleles that may present the MHC-presented peptide of the positive data. The positive data is multiallelic data and the list of MHC-alleles provided is a selection of MHC-alleles that could potentially present the peptides included in the positive data. In a preferred embodiment, the list of MHC-alleles is derived from MHC-typing of the sample donors and optionally MHC-affinity purification of the sample. In a preferred embodiment, the list of MHC-alleles is derived from MHC-typing of the sample donors. In other words, in the sample of the donors used to determine the positive data, the present MHC-alleles are determined, which then represent the list of MHC-alleles that may present the MHC-presented peptide. In another preferred embodiment, the list of MHC- alleles is derived from a sample, wherein particular MHC-presented peptides are enriched by MHC-affinity purification. Therefore, the donor samples (or samples of the donors) used to determine the positive data is subject to MHC-affinity purification. Depending on the specificity of the affinity binder used (e.g. antibodies or antigen binding fragments thereof) a list of MHC-alleles can be determined for a given sample. For example, affinity purification with a binder (preferably an antibody or antigen binding fragment thereof) specifically binding MHC-class I alleles (or HLA-alleles in a human context) would eliminate all non-class I alleles from the list of MHC-alleles. Likewise affinity purification with a binder (preferably an antibody or antigen binding fragment thereof) specifically binding MHC-class A (or -B or -C) alleles would limit the list of MHC-alleles to those being recognized by the binder. In other words, the MHC-affinity purification eliminates all peptides that are bound to MHC-alleles that are not specifically bound by the used antibody / ies.

[0146] In a preferred embodiment the binders used in affinity purification are used to deplete the samples of specific MHC-alleles. In another preferred embodiment the binders used in affinity purification are used to enrich the samples of specific MHC-alleles. In a preferred embodiment, a multitude of binders with different specificity are used for affinity purification. For example, a binder (preferably an antibody or antigen binding fragment thereof) specific for MHC-class I alleles is used in combination with a binder (preferably an antibody or antigen binding fragment thereof) specific for HLA-A (or -B or -C) alleles. These different binders may be used independently of each other for depletion or enrichment. For example, a sample may first be depleted of specific MHC-alleles and then enriched for a different allele.

[0147] In some embodiments, the MHC-affinity purification is performed with binders (preferably antibodies or antigen binding fragments thereof), specific for certain MHC-alleles or sub-classes of MHC-alleles. In a preferred embodiment the binders (preferably an antibody or antigen binding fragment thereof) used for affinity purification are specific for MHC-class I or MHC-class II (or the corresponding HLA-classes in the human context). In a preferred embodiment the binders (preferably an antibody or antigen binding fragment thereof) used for affinity purification are specific for MHC-class I. In yet another preferred embodiment the binder (preferably an antibody or antigen binding fragment thereof) used for affinity purification is specific for HLA-A, -B or -C. In yet another preferred embodiment the binder (preferably an antibody or antigen binding fragment thereof) used for affinity purification is specific for HLA-A. In yet another preferred embodiment the binder (preferably an antibody or antigen binding fragment thereof) used for affinity purification is specific for HLA-A*02. In a preferred embodiment the antibody used for affinity purification is selected from w6 / 32, BB7.2, GAP- A3 and B.1.23.2. In a preferred embodiment the binders (preferably an antibody or antigen binding fragment thereof) specifically binds to HLA-A*02 alleles. In a preferred embodiment the binders (preferably an antibody or antigen binding fragment thereof) specifically binds to HLA-A*03 alleles. In a preferred embodiment the binders (preferably an antibody or antigen binding fragment thereof) specifically binds to HLA-A*02 alleles and HLA-A*96 alleles.

[0148] In some embodiments, the indicative positive data comprises the sequence of MHC- presented peptides. In a preferred embodiment the MHC-presented peptides have a length of 8 to 12 amino acids, preferably 9 to 11, most preferably 9 amino acids. In yet another preferred embodiment the MHC-presented peptides have a length of 13 to 25 amino acids. In another embodiment the MHC-presented peptides have a length of 9 to 25 amino acids.

[0149] In a preferred embodiment the MHC-presented peptides in the indicative positive data only contain standard amino acids (i.e. A, R, N, D, C, Q, E, G, H, I, L, K, M, F, P, S, T, W, Y, V).

[0150] In a preferred embodiment the MHC-presented peptides in the indicative positive data occur more than once in the indicative positive dataset. In a preferred embodiment the MHC- presented peptides in the indicative positive data occur more than 2, 3, 4, 5, 6, 7, 8, 9 or 10- times.

[0151] In a preferred embodiment the indicative positive data consists of the sequence of MHC- presented peptides and a list of MHC-alleles that may present the MHC-presented peptides. The purpose of the positive data is to be used as a training dataset for the mathematical model used in the disclosed methods. Positive data indicating the amino acid sequence of MHC- presented peptides is therefore important so that the trained mathematical model is able to identify any pattern intrinsic to the amino acid sequence of MHC-presented peptides and to apply this to peptide query sequences to determine if these are MHC-presented peptides as well. In some instances associated positive data that goes beyond the amino acid sequence of MHC- presented peptides can further improve the training of the mathematical model. The associated positive data is determined within the same biological samples as the indicative positive data. In a preferred embodiment the associated positive data comprises quantitative data of proteins comprising at least one sequence of a MHC-presented peptide included in the indicative positive data. In other words, the (preferably absolute) quantification of the source protein of which one (or more) MHC-presented peptide in the indicative positive data originates from, can be used as associated positive data to train the mathematical model. In another preferred embodiment, the associated positive data comprises quantitative data on proteins involved in the antigen processing pathway, preferably selected from HLA, TAP, proteasome and immunoproteasome.

[0152] In another preferred embodiment, the associated positive data comprises quantitative data of nucleic acids encoding a MHC-presented peptide included in the indicative positive data. In a preferred embodiment, the nucleic acid is protein coding RNA. In another preferred embodiment the nucleic acid is genomic DNA, preferably genomic DNA encoding a MHC- presented peptide.

[0153] In another preferred embodiment, the associated positive data comprises or consists of one or more biological property of the sample in which the indicative positive data was determined. Preferably the one or more biological property is selected from MHC-alleles present, disease status of biological sample; ploidy of biological sample and mutation status of particular genes.

[0154] In some embodiments, the negative data used to train the mathematical model (preferably a neural network) are random peptide sequences obtained from a sequence database (preferably a proteome database), wherein a random list of MHC-alleles is provided. In other words, the data format of the negative data is to be the same as the positive data. In a preferred embodiment, the random peptide sequences obtained from a sequence database (preferably a proteome database) have a length or range of lengths as the MHC-presented peptides of the indicative positive data, wherein an MHC-allele is assigned randomly to the random peptide sequences. In another preferred embodiment the HLA-typing of a randomly selected donor is assigned to the negative data. In a preferred embodiment the random peptide sequences are filtered to not include sequences identical to the sequence of MHC-presented peptides (preferably the sequences of the indicative positive data used to train the mathematical model). In another preferred embodiment the random peptide sequences are not filtered to exclude sequences identical to the sequence of MHC-presented peptides (preferably the sequences of the indicative positive data used to train the mathematical model) and thus the negative data can by chance due to the random assignment also comprise MHC-presented peptides. This embodiment is particularly useful if the mathematical model is a neural network having a high noise resistance, since this increases the speed of the training process.

[0155] In some embodiments, the negative data used to train the mathematical model (preferably a neural network) further comprises random quantitative data of proteins comprising the random peptide sequences obtained from a sequence database (preferably a proteome database) or the quantitative data from the positive data assigned to the random peptide sequences. This embodiment provides the equivalent to associated positive data, by providing similar data for the negative data.

[0156] In some embodiments, the negative data used to train the mathematical model (preferably a neural network) further comprises random quantitative data of nucleic acids (preferably protein coding RNA) encoding the random peptide sequences obtained from a sequence database (preferably a proteome database) or the quantitative data from the positive data assigned to the random peptide sequences. This embodiment provides the equivalent to associated positive data, by providing similar data for the negative data.

[0157] In some embodiments, the negative data used to train the mathematical model (preferably a neural network) further comprises random biological properties (as defined for the associated positive data) randomly assigned to the random peptide sequences obtained from a sequence database (preferably a proteome database). This embodiment provides the equivalent to associated positive data, by providing similar data for the negative data.

[0158] In some embodiments, the negative data used to train the mathematical model (preferably a neural network) further comprises a random combination of the features of the positive data with the proviso that the data is not the positive data.

[0159] In some embodiments, the negative data used to train the mathematical model is derived from mismatched binders. In other words the negative data is not derived from random peptide sequences but actual sequences that are mismatched with regard to the HLA-allele.

[0160] In some embodiments, the MHC-allele(s) is (are) a MHC class I allele. Preferably in such an embodiment the trained mathematical model would be specific for MHC class I alleles, wherein preferably the training data (preferably the positive training data) comprises only MHC class I alleles.

[0161] In some embodiments, the MHC-allele(s) is (are) a MHC class II allele. Preferably in such an embodiment the trained mathematical model would be specific for MHC class II alleles, wherein preferably the training data (preferably the positive training data) comprises only MHC class II alleles.

[0162] In some embodiments, the positive and negative data is combined in a ratio that reflects the rarity of peptides being presented by a specific MHC-allele. In a preferred embodiment the ratio of positive data to negative data of 1:10 to 1:100, preferably 1:20 to 1:50. In another preferred embodiment the ratio of positive data to negative data is about 1:5 to 1:1000, preferably about 1:20 to 1:500, about 1:30 to 1:200, about 1:40 to 1:150, about 1:50 to 1:125. In some embodiments, the positive (and preferably also negative) data used for training of the mathematical model does not comprise peptide-MHC binding affinity (BA) assay data for training.

[0163] In some embodiments, the positive (and preferably also negative) data used for training of the mathematical model does not comprise data from mono-allelic samples (preferably mono-allelic cell lines).

[0164] An alternative aspect provides the trained mathematical model obtained by the methods disclosed herein.

[0165] Another alternative aspect relates to a computer readable storage medium comprising the trained mathematical model obtained by the methods disclosed herein.

[0166] Yet another alternative aspect relates to a computer system running the trained mathematical model obtained by the methods disclosed herein.

[0167] A further alternative aspect refers to a method for identifying an immunogenic composition, comprising the identification of at least one MHC-presented peptide present within a subject according to the methods disclosed herein, wherein at least one of those identified MHC-presented peptides is included in the immunogenic composition.

[0168] Some embodiments further regards the following items:

[0169] 1. A method, implemented by at least one processor, for determining whether or not a peptide query sequence is a major histocompatibility complex (MHC)-presented peptide, the method comprising: collecting multiallelic data that includes (i) a plurality of known MHC-presented peptide sequences and (ii) a list of MHC-alleles; training a machine learning model using the multiallelic data to generate a trained machine learning model; inputting the peptide query sequence into the trained machine learning model, the peptide query sequence corresponding to a peptide presented by a specific MHC- allele to output a score, the specific MHC-allele not being included in the list of MHC- alleles; based on the score being less than or equal to a predetermined value, determining that the peptide query sequence corresponds to an MHC-presented peptide; and based on the score greater less than the predetermined value, determining that the peptide query sequence does not correspond to an MHC-presented peptide. The method of item 1 further comprising: providing a list of MHC-alleles present in a specific sample; combining the scores with the list of MHC-alleles present in the specific sample to determine if the at least one peptide query sequence is an MHC-presented peptide in the specific sample. The method according to any one of items 1 to 2 further comprising:

[0170] (C) including at least one identified MHC-presented peptide into the immunogenic composition; optionally further including a pharmaceutically acceptable carrier. The method according to any one of items 1 to 2, wherein the mathematical model is selected from a neural network, a position- specific- scoring matrix (PSSM), machine learning methods (preferably support- vector machines or random forests). The method according to any one of items 1 to 2, wherein the mathematical model is a neural network, trained with multi-allelic positive and negative data, wherein the positive data is indicative of MHC-presented peptides and optionally further comprises positive data associated with MHC-presented peptides; and wherein the multi-allelic negative data is not indicative of or associated with the MHC- presented peptides. The method according to any one of items 2 and 4-5, wherein the combining the scores comprises the use of a weighted average based on the list of MHC-alleles present in the specific sample, wherein the weighted average is applied to the score to determine if the peptide query sequence is a peptide presented by a specific MHC-allele. The method according to any one of items 1 to 4, wherein the mathematical model is a position- specific- scoring matrix (PSSM), preferably wherein an iterative Bayesian method is used for training, wherein the PSSM resulting from a first training is used for a further training, preferably wherein up to 10, up to 15, up to 20 iterative rounds of training are applied to obtain a trained PSSM. The method according to any one of the preceding items, wherein the positive data comprises a list of MHC-alleles that may present the MHC-presented peptide of the positive data, wherein the list is derived from MHC-typing of the donors of the samples and optionally by MHC-affinity purification of the samples. The method according to item 8, wherein the MHC-affinity purification is performed with binders, preferably antibodies or antigen binding fragments thereof, specific for certain MHC-alleles or sub-classes of MHC-alleles. The method according to any one of the preceding items, wherein the positive data comprises:

[0171] (a) sequence of MHC-presented peptides;

[0172] (b) optionally, quantitative data of the MHC-presented peptides of (a) and / or proteins comprising at least one sequence of (a) determined in the same biological sample as (a);

[0173] (c) optionally, quantitative data of nucleic acids, preferably protein coding RNA, encoding the peptide of (a) determined in the same biological sample as (a); and / or

[0174] (d) optionally, properties of biological sample in which (a) and if used (b) and / or (c) were determined, preferably the properties are selected from MHC-alleles present, disease status of biological sample; ploidy of biological sample, mutation status of particular genes and tissue type. The method according to any one of items 1-6 and 8-10, wherein the mathematical model is a neural network and, wherein the negative data comprises:

[0175] (i) random peptide sequences obtained from a sequence database, preferably a proteome database, having a length or range of lengths as the sequences of (a), wherein a random list of MHC-alleles is provided;

[0176] (ii) optionally, random quantitative data of proteins comprising at least one sequence of (i);

[0177] (iii) optionally, random quantitative data of nucleic acids, preferably protein coding RNA, encoding the peptide of (i); and / or

[0178] (iv) optionally, properties of biological sample as defined in (d) randomly assigned to sequences of (i); or a random combination of the features of the positive data with the proviso that the data is not the positive data The method according to any one of items 1-6 and 8-11, wherein the positive and negative data is combined in a ratio that reflects the rarity of peptides being presented by a specific MHC-allele, preferably a ratio of positive data to negative data of 1:10 to 1:100. An in vitro method for determining if a peptide query sequence is an MHC-presented peptide in a subject comprising the following steps:

[0179] (1) obtaining a biological sample from said subject;

[0180] (2) determine genomic, transcriptomic and / or proteomic data in said biological sample;

[0181] (3) obtaining a trained mathematical model according to the method of item 3;

[0182] (4) providing at least one peptide query sequence and utilizing the trained mathematical model to determine if the at least one peptide query sequence is an MHC-presented peptide;

[0183] (5) determine if the at least one peptide query sequence being an MHC-presented peptide of step (4) is an MHC-presented peptide in said subject by utilizing the data obtained in step (2). omputing device, comprising: a memory storing instructions; and at least one processor configured to execute the instructions to: collect multiallelic data that includes (i) a plurality of known MHC-presented peptide sequences and (ii) a list of MHC-alleles; train a machine learning model using the multiallelic data to generate a trained machine learning model; input the peptide query sequence into the trained machine learning model, the peptide query sequence corresponding to a peptide presented by a specific MHC-allele to output a score, the specific MHC- allele not being included in the list of MHC-alleles; based on the score being less than or equal to a predetermined value, determine that the peptide query sequence corresponds to an MHC-presented peptide; and based on the score greater less than the predetermined value, determine that the peptide query sequence does not correspond to an MHC-presented peptide.

[0184] 15. A non-transitory computer-readable medium storing instructions, the instructions comprising: one or more instructions that, when executed by one or more processors of a device, cause the one or more processors to: collect multiallelic data that includes (i) a plurality of known MHC-presented peptide sequences and (ii) a list of MHC-alleles; train a machine learning model using the multiallelic data to generate a trained machine learning model; input the peptide query sequence into the trained machine learning model, the peptide query sequence corresponding to a peptide presented by a specific MHC-allele to output a score, the specific MHC-allele not being included in the list of MHC-alleles; based on the score being less than or equal to a predetermined value, determine that the peptide query sequence corresponds to an MHC-presented peptide; and based on the score greater less than the predetermined value, determine that the peptide query sequence does not correspond to an MHC-presented peptide.

[0185] Examples

[0186] The disclosed embodiments intend to provide the means to determine if any given peptide sequence is a MHC-presented peptide (or HLA-presented peptide in the human context). The mathematical model can be trained with unconvoluted immunopeptidomic data, mass-spectrometry-related data (e.g. used antibodies), sample characteristics (like HLA typing, tumor content), expression of relevant transcripts and proteins, etc. In particular the use of unconvoluted immunopeptidomic data provides a benefit compared to methods known in the art.

[0187] Example 1: Deconvolution of immunopeptidomic data

[0188] The indicative positive data used for training consists of multi-allelic data. Each positive is found on samples with 3-6 HLA class I alleles. It is not known which of the expressed HLAs actually presented the peptide. Therefore, the positive training data features a list of possible source alleles for each peptide entry. The list of possible source alleles can for example be determined by using affinity purification with affinity binders being specific for a certain subset of HLA-alleles or by HLA-typing of each donor sample. This allows to determine more likely and less likely source alleles for HLA-presented peptides identified in a specific sample. A further requirement is that the HLA-presented peptide is detected in more than one sample. The simplified general principle underlying the assignment of a MHC-allele to a MHC-presented peptide is explained in the following example. Table 1 below lists the HLA-alleles of eight donor samples in which a specific HLA-presented peptide with the amino acid sequence YLKVLPQEL (SEQ ID NO: 1) was detected.

[0189] Table 1: HLA-types in eight donor samples in which the same HLA-presented peptide was detected.

[0190] Based on the HLA-alleles being detected in the eight donors it can be determined that the peptide with the amino acid sequence YLKVLPQEL (SEQ ID NO: 1) is presented by HLA- B*08:01, which is present in all eight donors.

[0191] Example 2: Using an iterative Bayesian approach as an unbiased HLA deconvolution tool

[0192] The method described in this example aims to assign each peptide in the dataset to one HLA allele. This is done using an iterative Bayesian approach. It uses the frequency of the peptide across all HLA alleles. Meaning the more often a peptide is found on a donor with a specific HLA allele, the more likely the presentation is. In addition, specific antibodies are used in the HLA immunoprecipitation, allowing to narrow down the number of possible HLA alleles. For example, the antibody BB7.2 is specific for HLA-A*02 and will only immunoprecipitate peptides from this allele. In the immunoprecipitation, multiple antibodies are combined. For instance, the pan-HLA class I antibody W6 / 32 is subsequently used after the BB7.2, which results in a depletion of HLA-A*02 peptides in the eluate of the W6 / 32 antibody. Using this information, the Bayesian approach can calculate the probability that a ligand is originating from a specific HLA allele.

[0193] Bayesian approach:

[0194] Estimate HIX.D) = best HLA for peptide X on donor D

[0195] • Maximum a posteriori (MAP) estimate H'(X,D)=argmax_H P(H | D)

[0196] • H = peptide X is ligand of HLA allele H

[0197] • D = peptide X is identified at 5% EDR on donor D

[0198] Bayes’ theorem:

[0199] • Posterior = Likelihood * Prior

[0200] Likelihood

[0201] • Does typing match HLA H? A*03 ligand on A*03+ donor more likely than on A*02+ donor

[0202] • Does antibody capture HLA H? A*02 on BB7.2 more likely than on A*02- depleted w6 / 32, more likely than on Bl.23.3

[0203] Prior

[0204] • Is peptide a ligand for HLA H?

[0205] • Estimation

[0206] • Initialize with uniform distribution.

[0207] • Update with probability across donors from which sample was obtained.

[0208] • Update with probability across peptides.

[0209] Example 3: Training of Position Specific Scoring Matrix (PSSM) as predictor

[0210] The Position Specific Scoring Matrix (PSSM) is intended as a prediction tool and can be used to calculate the ligand probability. In the first iteration of the algorithm, a uniform probability for each HLA-peptide ligand probability is assumed. After the first assignment of the peptides to their HLA alleles (see example 2 above) these lists of peptides for each allele are used to compute a PSSM. And again, these PSSMs are then used in the next generation to predict the ligand probability. In total, 15 iterations are performed, in which the PSSMs are more and more optimized. The PSSMs from the last 15th iteration are the PSSMs used in the prediction.

[0211] Preprocessing (for PSSM):

[0212] • Shortening (9-12mers) of Peptides to 8mers o Find best shortening position by iterate over each position:

[0213] □ remove consecutive as many positions as necessary to achieve length of 8

[0214] □ Compute PSSM matrix for the resulting peptides (including also 9mers)

[0215] • Using p_ij* log] _2 (p_ij / pJ )

[0216] - With p_ij= amino acid j probability at position i

[0217] - And p _j= background probability for amino acid j

[0218] □ Sum up all scores of PSSM reports the Information content. o Compare the scores of all positions and pick the one with the highest score

[0219] PSSM creation:

[0220] 1. Creating position frequency matrix (PFM) by counting the occurrences of amino acid at each position.

[0221] 2. From the PFM, a position probability matrix (PPM) is computed by dividing that former amino acid count at each position by the number of sequences, thereby normalizing the values.

[0222] 3. Normalize the amino acid frequency by the corresponding amino acid frequency in the proteome.

[0223] 4. Calculate log likelihoods to compute PSSM [Final formula: log2 (pij / pj)

[0224] • With pq = amino acid j probability at position i

[0225] • And pj= background probability for amino acid j]

[0226] 5. Transform the PSSM scores to a rank based on 100.000 random Peptides.

[0227] 6. Repeat to optimize PSSM in order to predict ligand probability for particular peptide. Exemplary workflow for prediction of HLA-peptid.es using a PSSM:

[0228] • Immunoprecipitation of peptides with antibody specific for HLA allele(s)

[0229] • Assign each peptide in the database to one HLA allele (see example 2 above by using Bayesian approach) (frequency of peptides across all HLA alleles) Calculate probability that a ligand is originating from a specific HLA allele

[0230] • Algorithm assumes uniform probability for each HLA peptide ligand probability in the first iteration;

[0231] • Assignment of peptides to the corresponding HLA allele;

[0232] • Compute PSSM (see above) with peptides assigned to the corresponding HLA allele for each HLA allele.

[0233] • Repeat 15 times to optimize PSSM in order to predict ligand probability for particular peptide.

[0234] Example 4: Training of neural network as predictor

[0235] The HLA ligand prediction tool is intended to predict if a specific peptide is presented by an HLA allele on the cell surface of a given tissue donor (herein referred to as ligand probability). The input for the training of the prediction model consists of multidimensional data for human tissue samples. This data covers multiple dimensions of data for a given tissue donor including unconvoluted immunopeptidome data, mass-spectrometry-related data (e.g. used antibodies), sample characteristics (like HLA typing, tumor content) and expression of relevant transcripts and proteins. The method that is used for training of the predictor is described below.

[0236] The predictor has the following properties:

[0237] 1. Patient- specific: Predicting “if a given peptide sequence is presented by a patient with the given data” rather than predicting “if a peptide is presented by a given HLA allele in general”

[0238] 2. Population- scale prediction: Cross-allele network layout to cover the majority of the population of alleles

[0239] 3. Pan-length prediction: Smart padding of input sequences to cover relevant epitope lengths without loss of speed and accuracy

[0240] 4. Deconvolution-free prediction: Network-inherent deconvolution of ligand mixture 5. Mass-Spec-aware training: Incorporation of immunoprecipitation information (like antibody, retention time)

[0241] 6. Multidimensional feature set: Incorporation of expression information for relevant genes and / or proteins (e.g. source gene and antigen processing pathway information like HLA, TAP, Proteasome)

[0242] In these methods, the predictor is trained based on immunopeptidomic data (typically mass spectrometry) and sequencing data (typically RNAseq) from human tumor and healthy tissue. In addition, the HLA typing of the subject, tumor content of the subject, ploidy of the subject, and mutation status of particular genes are optionally provided within the training methods.

[0243] The complete training method covers the following steps:

[0244] 1. Obtain sample from human subject (e.g. biospecimen, such as tissue sample)

[0245] 2. Divide human sample into multiple aliquots for further analysis

[0246] 3. Isolation of peptides from first aliquot with HLA- specific antibody generating a mixture of ligands from different HLA alleles (the peptide eluate comprises ligands from several HLAs since the antibody is not specific to one particular HLA type) to provide peptidomics data

[0247] 4. Optional isolation of tryptic peptides from first aliquot generating tryptic lysate that are not enriched by HLA-specific antibody to provide proteomics data

[0248] 5. Identify, and optionally quantitate, peptides by mass spectrometry (MS) generating (quantitative) whole peptidomics data, and optionally proteomics data

[0249] 6. Optional RNA sequencing (transcriptome) of second aliquot generating whole transcriptome data in order to obtain expression levels of all genes comprised in sample, i.e. source gene, i.e. the gene encoding the protein from which the peptide identified in MS is derived, and all other genes.

[0250] 7. Combine these datasets (peptidome, transcriptome, proteome) into multidimensional feature set, wherein multidimensional feature set incorporates

[0251] (i) the peptidomics data obtained in 5,

[0252] (ii) the RNAseq data obtained in 6, and

[0253] (iii) proteomics data obtained in 5 (if available)

[0254] 8. Analysis of characteristics of the subject from third aliquot, like HLA typing of the subject, tumor content of the subject, ploidy of the subject, mutation status of particular genes, such as tumor markers (wherein this step does not comprise an assessment of neoantigens of the peptides identified in the MS analysis) Train a neural network for ligand prediction:

[0255] • Input features:

[0256] □ Peptide sequence, padded to configurable max. length, encoded as a vector (for example one-hot encoding)

[0257] □ Mass spectrometry related information (e.g. used antibody, uRT)

[0258] □ Expression levels for relevant genes / proteins (e.g. source gene, genes encoding antigen processing pathway proteins, like HLA, TAP, Proteasome)

[0259] □ Characteristics of the subject (e.g. mutational load)

[0260] • Output:

[0261] □ Predicted probability of peptide being presented by HLA

[0262] □ One prediction per supported HLA allele

[0263] □ List of supported alleles is the superset of alleles in the HLA typings of the human subjects in 1

[0264] • Training dataset:

[0265] □ Positive training examples taken from multidimensional feature set as defined in 7, combined with characteristics of the subject as defined in 8

[0266] □ Negative training examples generated by combining entries from the multidimensional feature set with characteristics of other subjects

[0267] □ Negative training examples generated by combining peptides from the multidimensional feature set with original characteristics of the subject but artificial (low) expression levels

[0268] □ Negative training examples generated by combining randomly sampled peptides (from the human proteome) with artificial expression levels, characteristics, etc.

[0269] □ Combined in a ratio that reflects the relative rarity of ligands (e.g. 1:100) Example 5: Training of mathematical model

[0270] The training dataset consists of positive peptides (= ligands) acquired via an immuno- peptidomics mass spectrometry (MS) pipeline, as well as negative peptides (= non-ligands) randomly sampled from the reference proteome (e.g. EMBL-EBI Ensembl).

[0271] Positive Peptides

[0272] Immunopeptidomic data were selected using the following criteria:

[0273] • Source has 4-digit HLA typing

[0274] • Antibody used in immuno-precipitation matches at least one of the source’s MHC class I alleles

[0275] • Peptides identified via mass spectrometry

[0276] Peptides resulting from selected immunopeptidomic data were selected using the following criteria:

[0277] • Length between 8 and 11

[0278] • Containing only standard amino acids (i.e. A, R, N, D, C, Q, E, G, H, I, L, K, M, F, P, S, T, W, Y, V)

[0279] • Occurs more than once (among all runs matching the conditions listed above)

[0280] Negative Peptides

[0281] Negative 8- to 11-mers were sampled from a reference proteome. The ratio between randomly sampled sequences (negatives) and sequences from immunopeptidomic s (i.e. MassSpec data) (positives) was set to 3:1. The sampling was done by creating peptide lists for each length, containing all possible unique peptides of that length from all transcripts of the reference.

[0282] HLA Typing

[0283] The positive training dataset consists of multi-allelic data. Each positive is found on samples with 3-6 HLA class I alleles. It is not known which of the expressed HLAs actually presented the peptide. Therefore, the training data features a list of possible source alleles for each peptide entry.

[0284] For the negative peptides, random HLA typings from the set of immunopeptidomic tissue samples were assigned. MHC Probabilities from affinity purification

[0285] Based on the antibodies used in the immuno-precipitation step (before MS), it is possible to determine more likely and less likely source alleles (among the samples expressed alleles) for the peptides identified in the MS step.

[0286] In addition to a “pan” antibody (e.g. W6 / 32) that binds all class I HLA alleles, the following antibodies are used for depletion, narrowing down the set of possible source alleles for peptides identified in that prep:

[0287] BB7.2 binds all A*02 and A*96 alleles

[0288] GAP- A3 binds all A*03 alleles

[0289] B.1.23.2 binds all B and C alleles, as well as some groups of A alleles o A*23, A*25, A*26, A*29, A*30, A*31, A*32, A*33, A*34, A*43, A*66, A*74 o A*24 and A* 80 weakly

[0290] A common scenario is to use one of the more specific antibodies first (most probably one that matches an expressed HLA-A allele), and then capture all remaining class I ligands with a pan antibody.

[0291] Based on the antibodies binding profiles, an allele probability distribution can be derived for each preparation. For “depleted” alleles, i.e. those whose ligands have already been captured in a previous preparation, a probability greater zero is still assumed, since the antibody capturing process is never exhaustive. One possibility is to evenly distribute a probability mass of 10% evenly over all depleted alleles (if any) and then distribute the remaining probability mass evenly over the non-depleted alleles. Alleles which are not captured by the antibody are assigned a probability of 0%.

[0292] Validation Data

[0293] 5% of the positive peptides (=positive indicative data) was held out for validation. The data was split so that there were no sequence overlap between the training and the validation peptides. For each positive peptide, 100 negative peptides were randomly sampled from the reference proteome. Model Architecture

[0294] The model is a neural network (Multilayer Perceptron) that predicts per-allele ligand probabilities for a given peptide sequence (see also figure 2).

[0295] The model in this example is designed specifically for MHC class I alleles.

[0296] The main input of the model is the peptide query sequence. The minimum and maximum lengths of the input sequence are limited to the peptide lengths in the training dataset, i.e. 8-11 amino acids.

[0297] Before being fed into the trained part of the model, the amino acid sequence is transformed into a representation more suitable for the subsequent layers. First, the sequence is padded to the maximum supported input length using “X”s (indicating “no amino acid”). This is done independently once from the N-terminus and once from the C-terminus so that the model can learn positional motifs from both ends of the strand. Both padded versions are then concatenated and the resulting sequence of 22 amino acids is encoded using a one-hot encoding (each residue represented as a vector of length 20, with the position corresponding to the amino acid being 1, all other positions being 0). One-hot encoding is a method to associate a vector with a token where each amino acid is represented by a unit binary vector of length n, containing a single one and n-1 zeros (e.g., [1,0,0, ..., 0] for one amino acid and [0,1,0, ..., 0] for another amino acid). This solution treats all amino acids equally without using any prior knowledge.

[0298] The trained part of the network (the two boxes labeled “Shared submotif layer” and “Allele score layer” in figure 2) consists of a single fully connected hidden layer of 8192 ReLU units (without bias), followed by a fully connected allele score layer with one sigmoid unit (with bias) for each of the alleles expressed in the training samples.

[0299] The output of the allele score layer is a vector of per-allele ligand probabilities (0.0 - 1.0). By taking a (weighted) average of the scores of the alleles expressed in a sample, a samplelevel ligand probability can be computed.

[0300] During training, the antibody-derived allele probability distribution (see section “Antibody Probabilities” above) is used to aggregate the output vector into a single output value. The probabilities for non-expressed alleles are all set to zero. That single output value is used to compute the loss (i.e. divergence from the expected output) for backpropagation.

[0301] Model Training

[0302] The neural network model was trained using the Python package “TensorFlow”, using only standard building blocks and methods. Binary cross-entropy was used as loss function and Adam as optimizer. The default back-propagation training loop of the training framework was used, providing it with a batch generator to create training batches on the fly.

[0303] The generator generates batches of training examples with 64 negatives for each positive. For each training example, the batches contain the peptide sequence, as well as a vector with a probability (0.0- 1.0) for each of the alleles in the training dataset. These vectors represent the antibody probabilities described above. For all alleles not expressed in the associated tissue sample, the vector contains zeros.

[0304] Positive examples are chosen randomly from the list of positives (described above). Negative examples are chosen randomly from a list of (previously) randomly selected peptides from the reference proteome. This pre-generated list is constructed so that it contains an equal amount of transcriptomic peptides for each supported peptide length. This leads to an approximately uniform length distribution among the negative peptides seen by the model during training.

[0305] This stochastic sampling approach does not guarantee that all positive examples are seen equally often. Some examples might get chosen more often than others, and some might not get chosen at all. However, due to the size of the dataset, even if examples were presented one after another, the training would not reach the end of the list before beginning to overfit on the training data. Therefore, surprisingly sampling with or without replacement makes little difference in practice, as long as each training example has an equal chance of being chosen. Keeping track of already seen peptide would however come with a substantial performance overhead, which is avoided by the present method.

[0306] During training we periodically saved snapshots of the model weights and computed validation performance on a held-out set of validation data (described above). In contrast to the training batches, the validation batches were pre-computed once before training, to guarantee comparable performance read-outs. After finishing training, we selected the model snapshot that achieved the best validation loss.

[0307] We repeated this process eight times with different random seeds to obtain eight separate set of model weights. The seed affects the set and order of presented training examples, thus leading to different model weights on each repetition. The final trained model is an ensemble model that averages the outputs of the eight individual models. On average, the predictions of the ensemble can be expected to better than the predictions of each individual model. Motif Deconvolution

[0308] The neural network learns allele- specific motifs from multi- allelic training examples by leveraging per-peptide allele frequency statistics. For example, assume a peptide that was detected on tissue samples from three donors with following HLA genotypes (reduced to 2- digit typings for simplicity):

[0309] During model training, all three identifications will be presented as positive training examples to the network to learn from. For each instance, only the output neurons associated with the alleles expressed in the donor will be updated, pushing the weights towards a more positive output signal for the given peptide query sequence (see figure 3). Even though weights for all of the alleles in the table will be updated, the B*08 output neuron will overall receive the strongest update signal, and thus converge most towards the motif displayed by the presented peptide sequence. After training on millions of positive examples (and billions of negatives to provide a background distribution), weights for all alleles converge to recognize shared patterns within the subset of peptide sequences identified for each of the covered alleles. The model thus learns motifs from the allele frequency statistics of a vast dataset of multi-allelic peptide data.

[0310] Example 6: Performance (i.e. benchmark) of trained model

[0311] The mono-allelic peptide dataset published in Sarkizova el al (Nature Biotechnology volume 38, pages 199-209 (2020)) was used to measure the prediction performance of the methods disclosed herein using the PSSM (referred to herein as PSSM) or the neural network (referred to herein as Neural Network) and compare them to MHCflurry, a state-of-the-art competitor.

[0312] Dataset

[0313] The published dataset and the used training datasets share the following 80 common alleles:

[0314] A*01 : 01JA*02 : 01j A*02 : 02j A*02 : 03JA*02 : 05, A*02 : 06, A*02 : 07, A*02 : 11JA*03 : 01j A*ll : 01j A*11 : 02JA*23 : 01JA*24 : 02, A*24 : 07, A*25 : 01JA*26 : 01JA*29 : 02, A*30:01, A*30:02j A*31:01, A*32:01, A*33:01, A*33:03, A*66:01, A*68:01, A*68:02, A*74:01j 6*07:02, B*08:01, B*13:01, B*13:02, B*14:02, B*15:01, B*15:03, B*15:17, B*18:01, B*27:05, B*35:01, B*35:03, B*37:01, B*38:01, B*38:02, B*40:01, B*40:02, B*40:06, B*44:02, B*44:03, B*45:01, B*46:01, B*49:01, B*50:01, B*51:01, B*52:01, B*53:01, B*54:01, B*55:01, B*55:02, B*56:01, B*57:01, B*57:03, B*58:01, C*01:02, C*02:02, C*03:02, C*03:03, C*03:04, C*04:01, C*05:01, C*06:02, C*07:01, C*07:02, C*07:04, C*08:01, C*08:02, C*12:02, C*12:03 C*14:02, C*15:02, C*16:01, C*17:01

[0315] For those alleles, positive published peptides were combined with randomly sampled negative peptides to form a benchmark dataset. For each allele, the negative list was augmented with the published binders of all other alleles (i.e. published HLA-A*24:02 binders were used as negatives for HLA-A*02:01). This enables the benchmark to measure how well the predictors can differentiate between the motifs of the different alleles, not just whether they can differentiate between ligands and non-ligands in general. The ratio between positive and negatives was set to 1:1000, to simulate the natural sparsity of true ligands.

[0316] The prediction task was to separate both sets of peptides (positives and negatives) as cleanly as possible. The (area under the) ROC curve was used as a measure for this. Because of the strong imbalance of the two classes, and the more severe consequences a false positive prediction could have in a clinical scenario, the Precision-Recall-Curve was considered as well, which plots the true positive rate against precision instead of the true negative rate. This shifts the attention more towards the divergence between the positive subset of the benchmark data and the set of peptides predicted to be ligands.

[0317] PSSM and Neural Network are compared against MHCflurry 2.0, a widely used open- source ligand prediction algorithm. Figure 4A shows the average areas under the curve per HLA gene. The Average Precision metric is used to compute the area under the Precision-Recall- Curve (see figure 4B). This variant of the ROC metric works better for jagged, non- monotonous curves. The PPV @ 40% Sensitivity metric (see figure 4C) showcases average precision values at a specific recall threshold (PPV = precision, sensitivity = recall).

[0318] It can be seen in figure and 5 A-C that the herein provided Neural Network outperforms MHCflurry consistently. PSSM can achieve competitive precision for some alleles, in some cases even outperforming the other two (see figure 6). Example 7: Training of mathematical model

[0319] The training dataset consists of positive peptides (= ligands) acquired via an immuno- peptidomics mass spectrometry (MS) pipeline, as well as negative peptides (= non-ligands) sampled from the remaining space of theoretical human peptides.

[0320] Positive Peptides

[0321] Immunopeptidomic data were selected using the following criteria:

[0322] • 4-digit HLA typing is available for the tissue donor

[0323] • Antibody used in immuno-precipitation matches at least one of the source’s MHC class I alleles

[0324] • Peptides identified via mass spectrometry

[0325] Peptides resulting from selected immunopeptidomic data were selected using the following criteria:

[0326] • Length between 8 and 11

[0327] • Containing only standard amino acids (i.e. A, R, N, D, C, Q, E, G, H, I, L, K, M, F, P, S, T, W, Y, V)

[0328] • Occurs more than once (among all runs matching the conditions listed above)

[0329] Negative Peptides

[0330] For each donor HEA typing in the positive dataset, a list of negative peptides was generated by concatenating randomly sampled sequences from a reference genome (e.g. EMBE-EBI Ensembl) with random sequences from immunopeptidomic s that were never identified on any of the donors’ expressed allele. Random sequences were restricted to the same length classes as the positive peptides (8-11) and sampled with a uniform length distribution. Overall, the ratio between random sequences and immunopeptidomic s sequences was set to 4:1.

[0331] MHC Probabilities from affinity purification

[0332] Based on the antibodies used in the immuno-precipitation step (before MS), it is possible to determine more likely and less likely source alleles (among the samples expressed alleles) for the peptides identified in the MS step. In addition to a “pan” antibody (e.g. W6 / 32) that binds all class I HLA alleles, the following antibodies are used for depletion, narrowing down the set of possible source alleles for peptides identified in that prep:

[0333] BB7.2 binds all A*02 and A*96 alleles

[0334] GAP- A3 binds all A*03 alleles

[0335] B.1.23.2 binds all B and C alleles, as well as some groups of A alleles o A*23, A*25, A*26, A*29, A*30, A*31, A*32, A*33, A*34, A*43, A*66, A*74 o A*24 and A* 80 weakly

[0336] A common scenario is to use one of the more specific antibodies first (most probably one that matches an expressed HLA-A allele), and then capture all remaining class I ligands with a pan antibody.

[0337] Based on the antibodies binding profiles, an allele probability distribution can be derived for each preparation. For “depleted” alleles, i.e. those whose ligands have already been captured in a previous preparation, a probability greater zero is still assumed, since the antibody capturing process is never exhaustive. One possibility is to evenly distribute a probability mass of 10% evenly over all depleted alleles (if any) and then distribute the remaining probability mass evenly over the non-depleted alleles. Alleles which are not captured by the antibody are assigned a probability of 0%.

[0338] Validation Data

[0339] 5% of the positive peptides (=positive indicative data) was held out for validation. For each positive peptide, 100 negative peptides were randomly sampled from the reference proteome.

[0340] Model Architecture

[0341] The model is a neural network (Multilayer Perceptron) that predicts per-allele ligand probabilities for a given peptide sequence (see also figure 2).

[0342] The model in this example is designed specifically for MHC class I alleles.

[0343] The main input of the model is the peptide query sequence. The minimum and maximum lengths of the input sequence are limited to the peptide lengths in the training dataset, i.e. 8-11 amino acids.

[0344] Before being fed into the trained part of the model, the amino acid sequence is transformed into a representation more suitable for the subsequent layers. First, the sequence is padded to the maximum supported input length using “X”s (indicating “no amino acid”). This is done independently once from the N-terminus and once from the C-terminus so that the model can learn positional motifs from both ends of the strand. Both padded versions are then concatenated and the resulting sequence of 22 amino acids is encoded using a one-hot encoding (each residue represented as a vector of length 20, with the position corresponding to the amino acid being 1, all other positions being 0). One-hot encoding is a method to associate a vector with a token where each amino acid is represented by a unit binary vector of length n, containing a single one and n-1 zeros (e.g., [1,0,0, ..., 0] for one amino acid and [0,1,0, ..., 0] for another amino acid). This solution treats all amino acids equally without using any prior knowledge.

[0345] The trained part of the network (the two boxes labeled “Shared submotif layer” and “Allele score layer” in figure 2) consists of a single fully connected hidden layer of 8192 ReLU units (without bias), followed by a fully connected allele score layer with one sigmoid unit (with bias) for each of the alleles expressed in the training samples.

[0346] The output of the allele score layer is a vector of per-allele ligand probabilities (0.0 - 1.0). By taking a (weighted) average of the scores of the alleles expressed in a sample, a samplelevel ligand probability can be computed.

[0347] During training, the antibody-derived allele probability distribution (see section “Antibody Probabilities” above) is used to aggregate the output vector into a single output value. The probabilities for non-expressed alleles are all set to zero. That single output value is used to compute the loss (i.e. divergence from the expected output) for backpropagation.

[0348] Model Training

[0349] The neural network model was trained using the Python package “PyTorch”, using only standard building blocks and methods. Binary cross-entropy was used as loss function and AdamW as optimizer.

[0350] The default back-propagation training loop of the training framework was used, providing it with a batch generator to create training batches on the fly.

[0351] The generator generates batches of training examples with 64 negatives for each positive. For each training example, the batches contain the peptide sequence, as well as a vector with a probability (0.0- 1.0) for each of the alleles in the training dataset. These vectors represent the antibody probabilities described above. For all alleles not found in the donor HLA typing the vector contains zeros.

[0352] Positive and negative examples are chosen randomly from the lists of positive and negative peptides (described above). This stochastic sampling approach does not guarantee that all positive examples are seen equally often. Some examples might get chosen more often than others, and some might not get chosen at all. However, due to the size of the dataset, even if examples were presented one after another, the training would not reach the end of the list before beginning to overfit on the training data. Therefore, sampling with or without replacement makes little difference in practice, as long as each training example has an equal chance of being chosen. Keeping track of already seen peptide would however come with a substantial performance overhead, which is avoided with the chosen approach.

[0353] During training we periodically saved snapshots of the model weights and computed validation performance on a held-out set of validation data (described above). In contrast to the training batches, the validation batches were pre-computed once before training, to guarantee comparable performance read-outs. After finishing training, we selected the model snapshot that achieved the best validation loss.

[0354] We repeated this process eight times with different random seeds to obtain eight separate sets of model weights. The seed affects the set and order of presented training examples, thus leading to different model weights on each repetition. The final trained model is an ensemble model that averages the outputs of the eight individual models.

[0355] Example 8: Performance (i.e. benchmark) of trained models

[0356] The mono-allelic peptide dataset published in Sarkizova et al (Nature Biotechnology volume 38, pages 199-209 (2020)) was used to measure the prediction performance of the methods disclosed herein using the PSSM (referred to herein as PSSM) or the neural network (referred to herein as Neural Network) and compare them to MHCflurry, a state-of-the-art competitor.

[0357] Dataset

[0358] The published dataset and the used training datasets share the following 80 common alleles:

[0359] A*01:01. A*02:01. A*02:02. A*02:03. A*02:05. A*02:06. A*02:07. A*02:ll.

[0360] A*03:01, A*ll:01, A*ll:02, A*23:01, A*24:02, A*24:07, A*25:01, A*26:01, A*29:02,

[0361] A*30:01, A*30:02, A*31:01, A*32:01, A*33:01, A*33:03, A*66:01, A*68:01, A*68:02,

[0362] A*74:01, B*07:02, B*08:01, B*13:01, B*13:02, B*14:02, B*15:01, B*15:03, B*15:17,

[0363] B*18:01, B*27:05, B*35:01, B*35:03, B*37:01, B*38:01, B*38:02, B*40:01, B*40:02,

[0364] B*40:06, B*44:02, B*44:03, B*45:01, B*46:01, B*49:01, B*50:01, B*51:01, B*52:01,

[0365] 6*53:01, B*54:01, B*55:01, B*55:02, B*56:01, B*57:01, B*57:03, B*58:01, C*01:02, (2*02 : 02, (2*03 : 02, (2*03 : 03, (2*03 : 04, (2*04 : 01, C*05 : 01, (2*06 : 02, (2*07 : 01, (2*07 : 02, (2*07 : 04, (2*08 : 01, (2*08 : 02, (2*12 : 02, (2*12 : 03, (2*14 : 02, (2*15 : 02, C*16 : 01, (2*17 : 01

[0366] For those alleles, positive published peptides were combined with randomly sampled negative peptides to form a benchmark dataset. For each allele, the negative list was augmented with the published binders of all other alleles (i.e. published HLA-A*24:02 binders were used as negatives for HLA-A*02:01). This enables the benchmark to measure how well the predictors can differentiate between the motifs of the different alleles, not just whether they can differentiate between ligands and non-ligands in general. The ratio between positive and negatives was set to 1:1000, to simulate the natural sparsity of true ligands.

[0367] The prediction task was to separate both sets of peptides (positives and negatives) as cleanly as possible. The (area under the) ROC curve was used as a measure for this. Because of the strong imbalance of the two classes, and the more severe consequences a false positive prediction could have in a clinical scenario, the Precision-Recall-Curve was considered as well, which plots the true positive rate against precision instead of the true negative rate. This shifts the attention more towards the divergence between the positive subset of the benchmark data and the set of peptides predicted to be ligands.

[0368] PSSM and Neural Network are compared against MHCflurry 2.2, a widely used open- source ligand prediction algorithm. Figure 7 A shows the average areas under the curve per HLA gene. The Average Precision metric is used to compute the area under the Precision-Recall- Curve (see figure 7B). This variant of the ROC metric works better for jagged, non- monotonous curves. The PPV @ 40% Sensitivity metric (see figure 7C) showcases average precision values at a specific recall threshold (PPV = precision, sensitivity = recall).

[0369] It can be seen in figure 7 and 8 A-C that the herein provided Neural Network outperforms MHCflurry consistently. PSSM can achieve competitive precision for some alleles, in some cases even outperforming MHCflurry (see figure 9, e.g. C* 17:01).

Claims

Claims1. A method for determining that a peptide query sequence is a MHC-presented peptide, wherein the method comprises the following steps:(I) obtaining a trained mathematical model, wherein the data used to train the mathematical model comprises positive data, which is indicative of MHC- presented peptides, wherein a list of MHC-alleles that may present the MHC- presented peptide is provided for each MHC-presented peptide, wherein the specific MHC-allele presenting the peptide is not provided and wherein the positive data indicative of MHC-presented peptides that is used to train the mathematical model is multiallelic data, and wherein the positive data indicative of MHC-presented peptides contains only MHC-presented peptides identified in more than one sample;(II) providing at least one peptide query sequence and utilizing the trained mathematical model to determine for each peptide query sequence a score indicating that the at least one peptide query sequence is an MHC-presented peptide.

2. The method of claim 1 further comprising the steps:(III) providing a list of MHC-alleles present in a specific sample;(IV) combining the scores of step (II) with the list of MHC-alleles present in the specific sample to determine if the at least one peptide query sequence is an MHC-presented peptide in the specific sample.

3. A method for generating a mathematical model for determining that a peptide query sequence is a MHC-presented peptide, wherein the method comprises training a mathematical model with data that comprises positive data, which is indicative of MHC- presented peptides, wherein a list of MHC-alleles that may present the MHC-presented peptide is provided for each MHC-presented peptide, wherein the specific MHC-allele presenting the peptide is not provided and wherein the positive data indicative of MHC- presented peptides, that is used to train the mathematical model is multiallelic data and wherein the positive data indicative of MHC-presented peptides contains only MHC- presented peptides identified in more than one sample; wherein the trained mathematical model can be used to determine if the peptide query sequence is a MHC- presented peptide.

4. A method for preparing an immunogenic composition comprising at least one peptide determined to be a MHC-presented peptide, wherein the method comprises the following steps:(A) obtaining a trained mathematical model, wherein the data used to train the mathematical model comprises positive data, which is indicative of MHC- presented peptides, wherein a list of MHC-alleles that may present the MHC- presented peptide is provided for each MHC-presented peptide, wherein the specific MHC-allele presenting the peptide is not provided and wherein the positive data indicative of MHC-presented peptides, that is used to train the mathematical model is multiallelic data and wherein the positive data indicative of MHC-presented peptides contains only MHC-presented peptides identified in more than one sample;(B) providing at least one peptide query sequence and utilizing the trained mathematical model to determine if the at least one peptide query sequence is an MHC-presented peptide;(C) including at least one MHC-presented peptide identified in step (B) into the immunogenic composition; optionally further including a pharmaceutically acceptable carrier.

5. The method according to any one of claims 1 to 4, wherein the mathematical model is selected from a neural network, a position- specific- scoring matrix (PSSM), machine learning methods (preferably support- vector machines or random forests).

6. The method according to any one of claims 1 to 4, wherein the mathematical model is a neural network, trained with multi-allelic positive and negative data, wherein the positive data is indicative of MHC-presented peptides and optionally further comprises positive data associated with MHC-presented peptides; and wherein the multi-allelic negative data is not indicative of or associated with the MHC- presented peptides.

7. The method according to any one of claims 2 and 5-6, wherein step (IV) comprises the use of a weighted average based on the list of MHC-alleles present in the specificsample, wherein the weighted average is applied to the score of step (II) to determine if the peptide query sequence is a peptide presented by a specific MHC-allele.

8. The method according to any one of claims 1 to 5, wherein the mathematical model is a position- specific- scoring matrix (PSSM), preferably wherein an iterative Bayesian method is used for training, wherein the PSSM resulting from a first training is used for a further training, preferably wherein up to 10, up to 15, up to 20 iterative rounds of training are applied to obtain a trained PSSM.

9. The method according to any one of the preceding claims, wherein the positive data comprises a list of MHC-alleles that may present the MHC-presented peptide of the positive data, wherein the list is derived from MHC-typing of the donors of the samples and optionally by MHC-affinity purification of the samples.

10. The method according to claim 9, wherein the MHC-affinity purification is performed with binders, preferably antibodies or antigen binding fragments thereof, specific for certain MHC-alleles or sub-classes of MHC-alleles.

11. The method according to any one of the preceding claims, wherein the positive data comprises:(a) sequence of MHC-presented peptides;(b) optionally, quantitative data of the MHC-presented peptides of (a) and / or proteins comprising at least one sequence of (a) determined in the same biological sample as (a);(c) optionally, quantitative data of nucleic acids, preferably protein coding RNA, encoding the peptide of (a) determined in the same biological sample as (a); and / or(d) optionally, properties of biological sample in which (a) and if used (b) and / or (c) were determined, preferably the properties are selected from MHC-alleles present, disease status of biological sample; ploidy of biological sample, mutation status of particular genes and tissue type.

12. The method according to any one of claims 1-7 and 9-11, wherein the mathematical model is a neural network and, wherein the negative data comprises:(i) random peptide sequences obtained from a sequence database, preferably a proteome database, having a length or range of lengths as the sequences of (a), wherein a random list of MHC-alleles is provided;(ii) optionally, random quantitative data of proteins comprising at least one sequence of (i);(iii) optionally, random quantitative data of nucleic acids, preferably protein coding RNA, encoding the peptide of (i); and / or(iv) optionally, properties of biological sample as defined in (d) randomly assigned to sequences of (i); or a random combination of the features of the positive data with the proviso that the data is not the positive data13. The method according to any one of the preceding claims, wherein the MHC-allele is a MHC class I allele.

14. The method according to any one of claims 1-7 and 9-13, wherein the positive and negative data is combined in a ratio that reflects the rarity of peptides being presented by a specific MHC-allele, preferably a ratio of positive data to negative data of 1:10 to 1:100.

15. An in vitro method for determining if a peptide query sequence is an MHC-presented peptide in a subject comprising the following steps:(1) obtaining a biological sample from said subject;(2) determine genomic, transcriptomic and / or proteomic data in said biological sample;(3) obtaining a trained mathematical model according to the method of claim 3;(4) providing at least one peptide query sequence and utilizing the trained mathematical model to determine if the at least one peptide query sequence is an MHC-presented peptide;(5) determine if the at least one peptide query sequence being an MHC-presented peptide of step (4) is an MHC-presented peptide in said subject by utilizing the data obtained in step (2).

Images