Improved methods and systems for prediction of major histocompatibility complex (MHC)-peptide binding of neoepitopes

Through the recursive neural network encoder and attention weighting mechanism, the problem of difficult to adapt to variable-length peptide sequences in the prior art is solved, the accuracy and performance of MHC-peptide binding affinity prediction is improved, and the binding of peptides to MHC molecules is achieved is achieved. More accurate predictions of location and affinity.

CN119091970APending Publication Date: 2024-12-06NANTOMICS LLC
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411099365.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Priority Date
2018-11-09
Filing Date
2019-08-14
Publication Date
2024-12-06

AI Technical Summary

Technical Problem

Existing neural network models are difficult to adapt to variable-length peptide sequences when predicting MHC-peptide binding affinity, resulting in limited prediction performance and the inability to accurately determine the binding affinity between the peptide and MHC molecules.

Method used

Using recurrent neural network encoder and attention weighting mechanism, a neural network model is generated by training multiple variable-length peptide sequences, and the final hidden state and intermediate state output of the recurrent neural network are used to generate attention weighted output, and a fixed output is generated after linear combination. Output of dimensions to improve the accuracy of MHC-peptide binding affinity prediction.

Benefits of technology

Improve the accuracy and performance of the prediction of MHC-peptide binding affinity for variable-length peptide sequences, enables more accurately determining the binding position and affinity of the peptide to MHC molecules, and improves the prediction of the immune system's response to tumor antigens.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure FT_1
    Figure FT_1
  • Figure FT_2
    Figure FT_2
  • Figure FT_3
    Figure FT_3
Patent Text Reader

Abstract

Methods and systems are provided for improving major histocompatibility complex (MHC)-peptide binding prediction for neoepitopes using a recurrent neural network encoder and attention weighting. Techniques for predicting MHC-peptide binding affinity are provided. A plurality of training peptide sequences are obtained, and a neural network model is trained using the training peptide sequences to predict MHC-peptide binding affinity. Comprising a neural network model of the RNN is configured to process an input training peptide sequence to generate attention weighted outputs by applying a final hidden state of the RNN to an intermediate state output of the RNN, and to linearly combine the attention weighted outputs to generate an output encoding a fixed dimension. A fully connected layer after the encoder is configured to process the encoded fixed dimension output to generate an MHC-peptide binding affinity prediction output. The computing device is configured to predict the MHC-peptide binding affinity for the test peptide sequence using the trained neural network.
Need to check novelty before this filing date? Find Prior Art

Description

[0001] This application is a divisional application of an invention application with a filing date of August 14, 2019, a Chinese national application number of 201980054372.3, and an invention name of “Methods and systems for improving major histocompatibility complex (MHC)-peptide binding prediction for new epitopes using recurrent neural network encoders and attention weighting”. Technical Field

[0002] The present disclosure relates generally to predicting major histocompatibility complex (MHC)-peptide binding, and more particularly to neural network models that employ one or more recurrent neural networks to generate MHC-peptide binding affinity predictions. Background Art

[0003] T cells, or T lymphocytes, are a type of lymphocyte (a subtype of white blood cells) that play a central role in cell-mediated immunity. A unique feature of T cells is their ability to distinguish between healthy cells and abnormal cells in the body, such as infected or cancerous cells. Healthy cells typically express large amounts of self-derived peptide-major histocompatibility complexes (pMHC) on their cell surface, and although the T cell antigen receptor can interact with at least a portion of these self-derived pMHC, T cells typically ignore these healthy cells. However, when the same cell contains even trace amounts of pathogen-derived pMHC, the T cell becomes activated and initiates an immune response. Positively selected T cells will have affinity for the pMHC and perform useful functions in the body, including interacting with the MHC and peptide complex to achieve an immune response, while negatively selected T cells that bind too strongly to self-antigens expressed on MHC molecules are eliminated to allow self-tolerance of the immune system.

[0004] Cytotoxic T cells (also known as TC cells, CTLs, T killer cells, killer T cells) destroy virus-infected cells and tumor cells. These cells (also known as CD8 T cells because they express CD8 glycoprotein on their surface) recognize virus-infected or tumor cell targets by binding to fragments of non-self proteins (peptide antigens), which are usually 8-15 amino acids in length and presented by major histocompatibility complex (MHC) class I molecules. Peptides of a specific length are usually referred to as "N-mers". For example, a peptide sequence with a length of 9 amino acids can be referred to as a 9-mer.

[0005] Class I MHC molecules are one of the two main classes of major histocompatibility complex (MHC) molecules (the other is class II MHC) and are present on the surface of all nucleated cells in humans. Their function is to display intracellular peptide antigens to cytotoxic T cells, thereby triggering an immediate response of the immune system to the specific non-self antigen displayed.

[0006] A current challenge in immunology is to understand which peptides bind well to which class I MHC molecules, i.e. which peptides are best suited to activate a cytotoxic T cell response, especially since each allele (variant form) of an MHC compound has different properties. If such MHC-peptide binding affinities could be accurately predicted for protein fragments of various lengths, new immunotherapies could be developed, for example, based on determining which tumor antigens are most likely to trigger an immune system response.

[0007] Neural networks have been used to predict MHC-peptide binding affinity. While class I MHC molecules can bind peptides of 6-20 amino acids in length (although they are typically 8-15 amino acids in length) and class II MHC molecules can bind peptides of 10-30 amino acids in length (although they are typically 12-25 amino acids in length), one current shortcoming is that the inputs to these neural network models are typically fixed length and cannot accommodate variable peptide sequence lengths without padding (i.e., adding one or more '0' or null values ​​to the encoded peptide sequence to match the fixed input length of the neural network). While such padding has been shown to have no effect on predictive performance when neural networks are trained using single-length peptide sequences (e.g., datasets containing only 9-mer peptide sequences, only 10-mer peptide sequences, etc.), current neural network models using such padding do not reach their full predictive performance potential when trained using peptide sequences of variable length. Therefore, there remains a need for techniques to improve the performance of MHC-peptide binding affinity predictions when neural networks are trained using peptide sequences of variable length. Furthermore, it will improve the MHC-peptide binding affinity prediction performance, making it possible to determine the peptide positions of the test input sequences that are most important for predicting MHC-peptide binding affinity. Summary of the invention

[0008] Devices, systems, methods, and articles of manufacture are described herein that involve the use of neural network models to predict MHC-peptide binding affinities. Various embodiments are based on a neural network model that employs a recursive neural network encoder and attention weighting to generate MHC-peptide binding affinity predictions with improved accuracy when trained with peptide sequences of variable lengths. In this way, accurate MHC-peptide binding affinity predictions can be made for test peptide sequences that are similar to training peptide sequences for which binding affinity data is known, but of different lengths.

[0009] In one embodiment, a plurality of training peptide sequences are obtained, and a neural network model is configured to be trained using these training peptide sequences to predict MHC-peptide binding affinity. An encoder of a neural network model comprising a recurrent neural network (RNN) is configured to process an input training peptide sequence to generate an attention weighted output by applying a final hidden state of the RNN to an intermediate state output of the RNN, and linearly combining the attention weighted outputs to generate an output encoding a fixed dimension. Each attention weighted output can be a single value and corresponds to an amino acid position of the input training peptide sequence. The neural network model is trained using multiple batches of training peptide sequences, and a computing device is configured to use the trained neural network model to predict MHC-peptide binding affinity for a test peptide sequence.

[0010] In some embodiments, the RNN may include a long short-term memory (LSTM) RNN or a gated recurrent unit (GRU) RNN, or any variant thereof.

[0011] In some embodiments, the RNN may include a bidirectional RNN, and the output encoding the fixed dimension may be determined by concatenating the outputs of the bidirectional RNNs.

[0012] In some embodiments, applying the final hidden state to the intermediate state output of the RNN to generate the attention-weighted output may include taking a dot product of the final hidden state and the intermediate state output.

[0013] In some embodiments, weights learned from training the neural network model may be applied to the final hidden state before the final hidden state is applied to the intermediate state outputs of the RNN to generate an attention-weighted output.

[0014] In some embodiments, the final hidden state can be concatenated with a final hidden state of an encoder of a second neural network model before being applied to the intermediate state outputs of the RNN to generate an attention-weighted output. The second neural network model can be configured to predict MHC-peptide binding affinity for MHC allele inputs.

[0015] In some embodiments, the training peptide sequence may comprise a plurality of sequence lengths between 6-20 or 10-30 amino acids in length, and may be one of a one-hot encoding, a BLOSUM encoding, a PAM encoding, or a learned embedded encoding sequence. Each training peptide sequence may be a positive MHC-peptide binding example.

[0016] In some embodiments, the sequence length of the test peptide sequence can be 6-20 or 10-30 amino acids in length. The test peptide sequence can have a sequence length different from the sequence length of at least one of the training peptide sequences, and can be one of a one-hot encoding, a BLOSUM encoding, a PAM encoding, or a learned embedded encoding sequence.

[0017] In some embodiments, each MHC-peptide binding prediction output can be a single prediction, and the MHC-peptide binding affinity predictions for the test peptide sequences can be correlated with activation of T cell responses to tumors.

[0018] In some embodiments, at least one fully connected layer (e.g., two fully connected layers) after the encoder can be configured to process the output of the encoding fixed dimensionality to generate an MHC-peptide binding affinity prediction output. At least one fully connected layer can include one of a deep convolutional neural network, a residual neural network, a densely connected convolutional neural network, a fully convolutional neural network, or a RNN.

[0019] In some embodiments, predicting MHC-peptide binding affinity for a test peptide sequence may include processing the test training peptide sequence using an encoder of a trained neural network model to generate an attention-weighted output by applying a final hidden state of the RNN to an intermediate state output of the RNN, and linearly combining the attention-weighted outputs to generate an output encoding a fixed dimension, and processing the output encoding a fixed dimension using at least one fully connected layer of the trained neural network model to generate an MHC-peptide binding affinity prediction output.

[0020] Various objects, features, aspects, and advantages of the present subject matter will become more apparent from the following description and accompanying drawings, in which like numerals represent like parts. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 Shown is a visual representation of an MHC molecule binding to a peptide on the surface of a nucleated cell according to one embodiment.

[0022] Figure 2 An example of a one-hot encoded peptide sequence according to one embodiment is shown.

[0023] Figure 3 An overall flow chart of an exemplary operation for predicting MHC-peptide binding affinity according to one embodiment is shown.

[0024] Figure 4 A block diagram of a system for predicting MHC-peptide binding affinity according to one embodiment is shown.

[0025] Figure 5 A schematic diagram of a recurrent neural network that can be used to encode an input peptide sequence is shown according to one embodiment.

[0026] Figure 6 A schematic diagram of a bidirectional recurrent neural network that can be used to encode an input peptide sequence is shown according to one embodiment.

[0027] Fig. 7A Shown is a visualization of attention weights determined for peptide positions of an input peptide sequence according to one embodiment.

[0028] Figure 7B Shown is a visualization of attention weights determined for peptide positions of an input peptide sequence according to one embodiment.

[0029] Figure 7C Shown is a visualization of an attention map determined for peptide positions of an input peptide sequence according to one embodiment.

[0030] Figure 8 A flow chart illustrating exemplary operations for training a neural network model to predict MHC-peptide binding affinity using training peptide sequences of variable lengths, according to one embodiment.

[0031] Fig. 9 A flow chart illustrating exemplary operations for predicting MHC-peptide binding affinity for test peptide sequences using a trained neural network model, according to one embodiment.

[0032] 10 shows a graphical representation of the performance of neural network validation on variable length peptide sequences using a neural network model according to one embodiment relative to alternative approaches.

[0033] Fig.11 A block diagram illustrating an exemplary client-server relationship that may be used to implement one or more aspects of the various embodiments; and

[0034] Fig.12 A block diagram of a distributed computer system that may be used to implement one or more aspects of the various embodiments is shown.

[0035] Although the present invention is described with reference to the above drawings, the drawings are intended to be illustrative and other embodiments are consistent with the spirit of the invention and are within the scope of the invention.

[0036] manual

[0037] Various embodiments will now be described more fully below with reference to the accompanying drawings, which form a part thereof and illustrate specific examples of practical embodiments in a graphical manner. However, this specification may be embodied in many different forms and should not be construed as being limited to the embodiments set forth herein; on the contrary, these embodiments are provided so that this specification will be exhaustive and complete, and will fully convey the scope of the invention to those skilled in the art. Among other things, this specification may be embodied as a method or apparatus. Therefore, any of the various embodiments herein may take the following forms: an all-hardware embodiment, an all-software embodiment, or an embodiment combining software and hardware aspects. Therefore, the following description should not be considered restrictive.

[0038] Throughout the specification and claims, unless the context clearly dictates otherwise, the following terms have the meanings explicitly associated herein:

[0039] The phrase "in one embodiment" as used herein does not necessarily refer to the same embodiment, although it may. Thus, the various embodiments of the invention may be readily combined as described below without departing from the scope or spirit of the invention.

[0040] As used herein, the term "or" is an inclusive "or" operator, and is equivalent to the term "and / or," unless the context clearly dictates otherwise.

[0041] Unless the context clearly indicates otherwise, the term "based on" is not exclusive and allows for being based on additional factors not described.

[0042] As used herein, and unless the context indicates otherwise, the term "coupled to" is intended to include direct coupling (where two elements coupled to each other are in contact with each other) and indirect coupling (where at least one additional element is located between the two elements). Therefore, the terms "coupled to" and "coupled with..." are used synonymously. In the context of a network environment where two or more components or devices are able to exchange data, the terms "coupled to" and "coupled with..." are also used to mean "communicatively coupled with...", possibly via one or more intermediate devices.

[0043] In addition, throughout the specification, the meanings of “a / an” and “the” include plural references, and the meaning of “in” includes “in” and “on”.

[0044] Although some of the various embodiments presented herein constitute a single combination of the present invention's elements, it should be understood that the present invention's subject matter is considered to include all possible combinations of the disclosed elements. Therefore, if one embodiment includes elements A, B, and C, and another embodiment includes elements B and D, the present invention's subject matter is also considered to include other remaining combinations of A, B, C, or D, even if not explicitly discussed herein. In addition, the transitional term "includes" means to have as a part or member, or those parts or members. As used herein, the transitional term "includes" is inclusive or open, and does not exclude otherwise unstated elements or method steps.

[0045] Throughout the following discussion, there will be a lot of references to servers, servers, interfaces, engines, modules, clients, peers, ports, platforms, or other systems formed by computing devices. It should be understood that the use of such terms should be considered to represent one or more computing devices having at least one processor (e.g., ASIC, FPGA, DSP, x86, ARM, ColdFire, GPU, multi-core processor, etc.), which is configured to execute software instructions stored on a computer-readable tangible non-transitory medium (e.g., hard drive, solid-state drive, RAM, flash memory, ROM, etc.). For example, a server may include one or more computers that operate as a network server, database server, or other type of computer server in a manner that satisfies the described roles, responsibilities, or functions. It should be further understood that the disclosed computer-based algorithms, processes, methods, or other types of instruction sets may be embodied as a computer program product, which includes a non-transitory tangible computer-readable medium storing instructions that cause the processor to perform the disclosed steps. Various servers, systems, databases, or interfaces can use standardized protocols or algorithms to exchange data, possibly based on HTTP, HTTPS, AES, public-private key exchange, network service APIs, known financial transaction protocols, or other electronic information exchange methods. Data exchange may occur over a packet-switched network, a circuit-switched network, the Internet, a LAN, a WAN, a VPN, or other types of networks.

[0046] As used in the description herein and throughout the following claims, when a system, engine, server, device, module, or other computing element is described as being configured to perform or implement functions on data in a memory, the meaning of "configured to" or "programmed to" is defined as one or more processors or cores of the computing element being programmed by a set of software instructions stored in the memory of the computing element to perform the set of functions on target data or data objects stored in the memory.

[0047] It should be noted that any language for computers should be read to include any suitable combination of computing devices, including servers, interfaces, systems, databases, agents, peers, engines, controllers, modules, or other types of computing devices that operate alone or together. It should be understood that the computing device includes a processor configured to execute software instructions stored on a tangible non-transitory computer-readable storage medium (e.g., a hard drive, FPGA, PLA, solid-state drive, RAM, flash memory, ROM, etc.). The software instructions configure or program the computing device to provide roles, responsibilities, or other functions as discussed below with respect to the disclosed device. In addition, the disclosed technology can be embodied as a computer program product, which includes a non-transitory computer-readable medium storing software instructions, which enables the processor to perform the disclosed steps associated with the implementation of a computer-based algorithm, process, method, or other instruction. In some embodiments, various servers, systems, databases, or interfaces use standardized protocols or algorithms to exchange data based on HTTP, HTTPS, AES, public key-private key exchange, network service API, known financial transaction protocols, or other electronic information exchange methods. Data exchange between devices may be performed via: a packet-switched network, ie, the Internet, a LAN, a WAN, a VPN, or other types of packet-switched networks; a circuit-switched network; a cell-switched network; or other types of networks.

[0048] The disclosed subject matter focuses on enabling the construction or configuration of a computing device to operate on large amounts of digital data beyond human capabilities for purposes including predicting MHC-peptide binding affinities of variable length peptide sequences.

[0049] It should be appreciated that the disclosed technology provides a number of advantageous technical effects, including the use of neural network models to improve the scope, accuracy, compactness, efficiency and speed of MHC-peptide binding affinity predictions for variable length peptide sequences. It should also be appreciated that the following description is not intended as an extensive overview, and therefore, concepts may be simplified for the sake of clarity and brevity.

[0050] Predicting MHC-peptide binding affinity of variable-length peptide sequences using a recurrent neural network encoder and attention weighting force

[0051] In current neural network-based MHC-peptide binding affinity prediction models, the neural network input is typically fixed-length and cannot accommodate variable-length peptide sequences without padding (i.e., adding one or more '0' or null values ​​to the encoded peptide sequence to match the fixed-length neural network input). While such padding has been shown to have no performance impact on neural networks trained using single-length peptide sequences (e.g., datasets containing only 9-mer peptide sequences, only 10-mer peptide sequences, etc.), each of the current prediction models has been shown to still have room for improved prediction performance when trained using variable-length peptide sequences in conjunction with the use of a single padding approach.

[0052] However, the performance limitations of the MHC-peptide binding affinity prediction model can be improved by a neural network model including a recurrent neural network encoder that is configured to use attention weighting for peptide positions of an input peptide sequence. Once trained, such a neural network model can determine attention weights for peptide positions of a test input sequence and generate MHC-peptide binding affinity predictions with increased accuracy based on the attention weights.

[0053] Figure 1 A visual representation of MHC molecules binding to peptides on the surface of a nucleated cell according to one embodiment is shown. Illustration 100 shows a class II MHC molecule 102, which presents a stably bound peptide 104 that is necessary for overall immune function. Class II MHC molecules 102 interact primarily with immune cells such as helper (CD4) T cells 106. For example, peptides 104 (e.g., antigens) can regulate how CD4 T cells 106 respond to infections. Typically, stable peptide binding is necessary to prevent the separation and degradation of peptides, which may occur without being firmly attached to class II MHC molecules 102. This separation and degradation will prevent T cells from recognizing antigens, T cell recruitment, and appropriate immune responses. CD4 T cells are so called because they express CD4 glycoproteins on their surfaces, which can be used for antigenic activation of CD8 T cells (e.g., CD8 T cells 108). Therefore, the activation of CD4 T cells can be beneficial to the action of CD8 T cells.

[0054] CD8 T cells 108 are cytotoxic T cells expressing CD8 glycoprotein on their surface. Cytotoxic T cells (also referred to as TC cells, CTL, T killer cells, killer T cells) destroy virus-infected cells and tumor cells. These cells recognize virus-infected targets or tumor cell targets by binding to fragments of non-self proteins (peptide antigens), and the fragments of these non-self proteins are 6-20 amino acids in length (although they are usually 8-15 amino acids in length) and presented by major histocompatibility complex (MHC) class I molecules (e.g., class I MHC molecules 110). Class I MHC molecules are present on the surface of all human nucleated cells. Their function is to display intracellular peptide antigens (e.g., peptide 112) to cytotoxic T cells, thereby triggering the immediate response of the immune system to the displayed peptide antigens. Understanding which peptide is well combined with which class I MHC molecule (i.e., which peptide is best suited for activating cytotoxic T cell responses) is crucial for current immunological research, especially because each allele (variant form) of MHC compounds has different characteristics. The embodiments herein improve the operation of a neural network-based MHC-peptide binding affinity prediction model by using combined variable-length training peptide sequences to generate more accurate predictions.

[0055] Figure 2 An example of a one-hot encoded peptide sequence according to one embodiment is shown. In an exemplary embodiment, for the techniques described herein, the training peptide sequence can be a one-hot encoded sequence of any length. For example, the one-hot encoding matrix 200 represents the one-hot encoding of the 9-mer protein / peptide sequence "ALATFTVNI", where single letter codes are used to represent the 20 naturally occurring amino acids. In the one-hot encoding matrix 200, the legal combinations of values ​​are only those with a single high ("1") bit, while the other values ​​are low ("0"). Although in Figure 2 9-mer protein / peptide sequences are shown in FIG, but the training peptide sequences described herein may include multiple sequence lengths of 6-20 or 10-30 amino acids in length. Figure 2 As an alternative to the one-hot encoding shown in , the training peptide sequences can be encoded for the techniques described herein as a BLOcks substitution matrix (BLOSUM) of the type commonly used for protein sequence alignment, a acceptable point mutation (PAM) matrix (in which each column and row represents the 20 standard amino acids), or a matrix for learning embedding encodings.

[0056] Figure 3An overall flow chart of an exemplary operation for predicting MHC-peptide binding affinity according to one embodiment is shown. In the flow chart 300, variable length training peptide sequences 1 to N 302, 304, and 306 are used to train a neural network model 308 to predict MHC-peptide binding affinity. According to an embodiment of the present invention, the neural network model 308 is configured to be trained using multiple (e.g., multiple batches) of training peptide sequences 1 to N 302, 304, and 306 to predict MHC-peptide binding affinity. In one embodiment, an encoder 310 of the neural network model 308 includes a recurrent neural network (RNN) configured to process an input training peptide sequence to generate an attention weighted output by applying the final hidden state of the RNN to the intermediate state output of the RNN, and linearly combining the attention weighted output to generate an output of an encoding fixed dimension. At least one fully connected layer 312 after the encoder 310 can be configured to process the output of the encoding fixed dimension to generate an MHC-peptide binding affinity prediction output 314 for the input training peptide sequence. Once training is complete, the trained neural network 316 including the trained encoder 318 and at least one trained fully connected layer 320 can be configured to receive a test peptide sequence 322. According to one embodiment, the test peptide sequence 322 is also processed by the trained encoder 318 to generate an output of the encoding fixed dimension by applying the final hidden state of the RNN to the intermediate state output of the RNN to generate an attention weighted output, and linearly combining the attention weighted outputs. At least one trained fully connected layer 320 after the trained encoder 318 can be configured to process the output of the encoding fixed dimension to generate an MHC-peptide binding affinity prediction output 324 for the input training peptide sequence 322.

[0057] Figure 4A block diagram of a system for predicting MHC-peptide binding affinity according to one embodiment is shown. In block diagram 400, elements for predicting MHC-peptide binding affinity in a test peptide sequence include a training engine 410, a prediction engine 420, a persistent storage device 430, and a main memory device 440. In one embodiment, the training engine 410 can be configured to obtain training peptide sequences 1 to N 302, 304, and 306 from either or both of the persistent storage device 430 and the main memory device 440. The training engine 410 can then use the training peptide sequences 1 to N 302, 304, 306 as training inputs to configure and train a neural network model 308, which can be stored in either or both of the persistent storage device 430 and the main memory device 440. For example, the training engine 410 can configure the encoder 310 of the neural network model 308 including a recurrent neural network (RNN) to process the input training peptide sequence to generate an output of the encoding fixed dimension by applying the final hidden state of the RNN to the intermediate state output of the RNN to generate an attention weighted output, and linearly combining the attention weighted output. At least one fully connected layer 312 after the encoder 310 can be configured by the training engine 410 to process the output of the encoding fixed dimension to generate an MHC-peptide binding affinity prediction output 314 for the input training peptide sequence. The training engine 410 can also configure the prediction engine 420 to use the trained neural network model 316 to predict the MHC-peptide binding affinity in the genomic sample input including the test peptide sequence 322. For example, the prediction engine 420 can obtain a test peptide sequence 322 and predict the MHC-peptide binding affinity by processing the test peptide sequence 322 via the trained encoder 318 to generate an output of the encoding fixed dimension by applying the final hidden state of the RNN to the intermediate state output of the RNN to generate an attention weighted output, and linearly combining the attention weighted outputs. Then, at least one trained fully connected layer 320 after the trained encoder 318 can process the output of the encoding fixed dimension to generate an MHC-peptide binding affinity prediction output 324 for the input training peptide sequence 322.

[0058] It should be noted that although it is exemplary, it is described in this way only for the purpose of easy understanding. Figure 4410, prediction engine 420, persistent storage device 430, and main memory device 440 should be read to include any suitable combination of computing devices, including servers, interfaces, systems, databases, agents, peers, engines, controllers, modules, or other types of computing devices that operate alone or in conjunction to perform the functions attributed to the various elements. In addition, those skilled in the art will understand that the functions attributed to the various elements may be performed by any of the other elements and / or by elements (not shown) configured to perform a combination of the various functions. Therefore, it should be noted that any language for training engine 410, prediction engine 420, persistent storage device 430, and main memory device 440 should be read to include any suitable combination of computing devices, including servers, interfaces, systems, databases, agents, peers, engines, controllers, modules, or other types of computing devices that operate alone or in conjunction to perform the functions attributed to the various elements. In addition, those skilled in the art will understand that the functions attributed to the various elements described herein are not limited to the functions of the various elements. Figure 4 One or more functions of the system may be performed in the context of a client-server relationship, such as by one or more servers, one or more client devices (e.g., one or more user devices), and / or by a combination of one or more servers and client devices.

[0059] Figure 5 5 shows a schematic diagram of a recurrent neural network that can be used to encode an input peptide sequence according to one embodiment. In general, a recurrent neural network (RNN) 500 processes an input sequence x using an internal state o to x t 502, 504, 506 and 508, the input sequence is an input peptide sequence of any length used in the embodiments of this document (e.g., the input peptide sequence "ALATFTVNI" 200 represented by a one-hot code). In RNN 500, the node RNN 0 To RNN t The connections between 510, 512, 514, and 516 form a directed graph (i.e., layer) along the sequence. In particular, RNN 500 may include nodes that are input nodes that receive data from outside the RNN, output nodes that produce results, or hidden nodes that modify data on the way from input to output along the sequence. o to x t During the processing of 502, 504, 506 and 508, each node RNN 0 To RNN t 510, 512, 514, and 516 generate intermediate state outputs, as shown by intermediate state outputs h0 through h2 518, 520, and 522. The final output along the sequence is the final hidden state output 524.

[0060] In one embodiment, the final hidden state output 524 is applied to each of the intermediate state outputs h0 to h2 518, 520, and 522 of the RNN 500 to generate an attention-weighted output. For example, the attention-weighted output can be generated by taking the dot product of the final hidden state output and the intermediate state output of each node. In some embodiments, before the final hidden state is applied to the intermediate state output to generate the attention-weighted output, the weights learned by training the neural network model can be applied to the final hidden state. In addition, the RNN 500 can generate the attention-weighted output by applying the final hidden state output 524 of the RNN to the intermediate state outputs h 0 to h 2 518, 520 and 522 are used to generate attention weighted outputs, and linearly combine the attention weighted outputs to generate an output encoding a fixed dimension.

[0061] A fully connected layer including a plurality of hidden neurons may follow an encoder (e.g., encoder 310 or 318) including RNN 500 to perform classification on the attention-weighted output. In one embodiment, a fully connected layer (e.g., fully connected layer 312 or 320) is configured to receive the encoded attention-weighted output from an encoder (e.g., encoder 310 or 318) including RNN 500 and generate an output value (e.g., output 314 or 324) representing an MHC-peptide binding affinity prediction.

[0062] As described above, a neural network model including an RNN 500 can be trained using multiple batches of training peptide sequences 1 to N 302, 304, and 306 to predict MHC-peptide binding affinity in the following manner: processing the input training peptide sequence to generate an output encoding a fixed dimension, so that the final hidden state of the RNN is applied to the intermediate state output of the RNN to generate an attention-weighted output, and the attention-weighted outputs are linearly combined to generate an output encoding a fixed dimension.

[0063] Figure 6 A schematic diagram of a bidirectional recurrent neural network that can be used to encode an input peptide sequence according to one embodiment is shown. In an alternative embodiment, a bidirectional recurrent neural network (RNN) 600 including a forward RNN 602 and a reverse RNN 604 can be used to process an input peptide sequence. Typically, a bidirectional RNN uses a finite sequence to predict or label each element of an input sequence based on the past and future context of the element. This is accomplished by processing an input sequence 606 (e.g., an input peptide sequence) of arbitrary length from left to right using a forward RNN 602 and from right to left using a reverse RNN 604, then cascading the outputs 608 and 610 of the two RNNs, wherein the combined output is used to determine the output of the encoding fixed dimension.

[0064] During the processing of the input peptide sequence 606, each node of the forward and reverse RNNs 602 and 604 generates an intermediate state output. In one embodiment, the cascade outputs 608 and 610 represent the final hidden state output of the bidirectional RNN 600, which is applied to each intermediate state output of the forward and reverse RNNs 602 and 604 to generate an attention-weighted output. For example, the attention-weighted output can be generated by taking the dot product of the final hidden state output and the intermediate state output of each node of the forward and reverse RNNs 602 and 604. The attention-weighted outputs can then be linearly combined to generate an output that encodes a fixed dimension. In some embodiments, before the final hidden state is applied to each intermediate state output to generate an attention-weighted output, the weights learned by training the neural network model can be applied to the final hidden state.

[0065] Although Figure 5 and Figure 6 The recursive neural network shown in is an example for implementing the embodiments herein, but those skilled in the art will appreciate that other recursive neural network architectures employing, for example, long short-term memory units (LSTM) and / or gated recursive units (GRU) may be utilized. Thus, RNN 500 should not be construed as being strictly limited to the embodiments described herein.

[0066] Fig. 7A The visualization of the attention weight determined for the peptide position of the input peptide sequence according to one embodiment is shown. The attention distribution of the peptide position for the associative and non-associative HLA-A-11019 polymers determined using the bidirectional recurrent neural network encoder described herein is shown in small figures 700 and 710 respectively. In small figures 700 and 710, the model based on the RNN encoder is trained for all lengths, but only for visualization, 9 polymer peptides are selected separately to take the average value of their position attention scores. In particular, small figure 700 shows that among the peptide positions of the HLA-A-11019 polymer binder, the attention weight of peptide position 8702 is the highest (average distribution is usually between 0.6 and 0.9), and peptide position 7704 has the second highest average distribution (usually between 0.05 and 0.23). Similarly, inset 710 shows that among the peptide positions of the HLA-A-11019-mer non-binder, the attention weight of peptide position 8712 is the highest (the average distribution is generally between 0.0 and 0.4).

[0067] Figure 7BA visualization of attention weights determined for peptide positions of an input peptide sequence according to one embodiment is shown. Inset 720 shows an attention map showing the attention weights of specific amino acids at different peptide positions of binding (722 and 724) and non-binding (726 and 728) HLA-A-11019 polymers. For example, attention map 722 (and Figure 724, which is a filtered version of Figure 722) shows a relatively high attention weight for the amino acid R in peptide position 8 of the HLA-A-11019 polymer binder. Similarly, attention map 726 (and Figure 728, which is a filtered version of Figure 726) shows, for example, relatively high attention weights for amino acids W, K, and R in peptide position 8 of the HLA-A-11019 polymer non-binder. In addition, Figure 7C Inset 740 is shown showing a visualization of an example attention map for HLA-A-1101 binders of variable length determined using the bidirectional recurrent neural network encoder described herein.

[0068] Figure 8 A flow chart of exemplary operations for training a neural network model to predict MHC-peptide binding affinity using training peptide sequences of variable lengths is shown according to one embodiment. In flowchart 800, at step 802, a plurality of training peptide sequences are obtained, for example, by training engine 410. In one embodiment, the plurality of training peptide sequences may include variable sequence lengths, such as sequence lengths of 6-20 amino acids or even 10-30 amino acids (e.g., for predicting class II MHC-peptide binding affinity). At step 804, an encoder (e.g., encoder 310) comprising a recurrent neural network (RNN) is configured to process the input training peptide sequence to generate an output of encoding fixed dimensionality by applying the final hidden state of the RNN to the intermediate state outputs of the RNN to generate an attention weighted output, and linearly combining the attention weighted outputs. In one embodiment, the final hidden state (h t ) may have additional information attached to it that the attention decision wishes to take into account. For example, in the case where one neural network model including an RNN encoder is processing an input peptide sequence and a second neural network model including an RNN encoder is processing an MHC allele sequence, the final hidden states of the two encoders [h1_N, h2_M] may be concatenated, or only the final hidden state from the other MHC allele sequence encoder may be used, such that the attention weight for the peptide encoder may be determined as:

[0069] f([h1_N, h2_M], h1_0), f([h1_N, h2_M], h1_1)...f([h1_N, h2_M], h1_N)

[0070] or

[0071] f(h2_M, h1_0), f(h2_M, h1_1)...f(h2_M, h1_N).

[0072] In the case of dual encoders, it should be noted that the MHC allele sequence encoder and the peptide encoder may have similar or different architectures and may share some but not all components. For example, the amino acid embedding layer may be shared, but the sequence processing architecture may be different.

[0073] At step 806, at least one fully connected layer (eg, fully connected layer 312) is configured to process the output encoding the fixed dimensionality to generate an MHC-peptide binding affinity prediction output. For example, the at least one fully connected layer may include a plurality of fully connected layers.

[0074] At step 808, a neural network is trained using a plurality of training peptide sequences. For example, each output value may be compared to a known label value (e.g., a known MHC-peptide binding affinity value corresponding to the input encoded peptide sequence) to determine a loss or error factor that may be used to determine parameter updates within the fully connected layer. For example, a stochastic gradient descent algorithm or variants thereof (e.g., Adagrad, RMSprop, Adam, etc.) may be used to determine parameter updates.

[0075] At step 810, a computing device (e.g., prediction engine 420) is configured to generate an MHC-peptide binding affinity prediction for a test peptide sequence using a trained neural network, wherein generating the MHC-peptide binding affinity prediction may include processing the test peptide sequence via a trained encoder to generate an output of an encoding fixed dimension by applying a final hidden state of the RNN to an intermediate state output of the RNN to generate an attention-weighted output, and linearly combining the attention-weighted outputs. Then, at least one trained fully connected layer following the trained encoder may process the output of the encoding fixed dimension to generate an MHC-peptide binding affinity prediction output for the input test peptide sequence.

[0076] Fig. 9 A flowchart is shown of exemplary operations for predicting MHC-peptide binding affinity for a test peptide sequence using a trained neural network model according to one embodiment. In flowchart 900 , at step 902 , a test peptide sequence is obtained, for example, by prediction engine 420 .

[0077] At step 904 , a test input sequence is input into a trained neural network model, such as trained neural network model 316 .

[0078] At step 906, the test training peptide sequence is processed using the encoder of the trained neural network model to generate an output of fixed dimension by applying the final hidden state of the RNN to the intermediate state output of the RNN to generate an attention weighted output, and linearly combining the attention weighted outputs.

[0079] At step 908, the output encoding the fixed dimensionality is processed using the fully connected layers of the trained neural network model to generate an MHC-peptide binding affinity prediction output. For example, the MHC-peptide binding affinity prediction for the test peptide sequence can be correlated with the activation of T cells in response to tumors.

[0080] FIG10 shows a graphical representation of the performance of neural network validation for variable length peptide sequences using a neural network model according to one embodiment relative to an alternative approach. In the chart 1000, columns A to D 1002 include 9-mer and 10-mer peptide data sets organized for each allele (e.g., HLA-A-0101, 9 and HLA-A-0101, 10), and columns E and F 1004 show the results of two independent neural networks for each allele trained for 9-mer and 10-mer peptides, respectively, without filling. For example, two neural networks trained with HLA-A-0101 9-mer and HLA-A-0101 10-mer, respectively, without filling, obtained receiver operating characteristic curve area (ROC AUC) = 0.951, precision-recall curve area (PRAUC) = 0.812, and ROC AUC = 0.766, PR AUC = 0.514.

[0081] Columns G to J 1006 show the performance of a single neural network model trained for each allele using a single fill method on data from 9-mers and 10-mers. The single fill method places the peptide in the center and fills at the beginning and end to reach a fixed length of 13. For example, when the single fill method is used for models trained on 9-mers and 10-mers of HLA-A-0101, the overall performance is ROC AUC=0.933, PR AUC=0.735, and the performance measured by peptide length alone is ROC AUC=0.953, PR AUC=0.810 (for 9-mers), and ROC AUC=0.811, PRAUC=0.522 (for 10-mers).

[0082] Columns K to N 1008 show the results of a single neural network trained for each allele using the extended fill-in technique on data from 9-mers and 10-mers. When trained on 9-mers and 10-mers of HLA-A-0101, the model achieved an overall ROCAUC = 0.933, PR AUC = 0.771, and when measured separately by peptide length, ROCAUC = 0.943, PRAUC = 0.794 (for 9-mers), and ROC AUC = 0.865, PR AUC = 0.682 (for 10-mers).

[0083] Columns O to T 1010 show the results of a single neural network trained for each allele on data from 9-mers and 10-mers using the bidirectional recurrent neural network encoder and attention weights described herein. When trained on 9-mers and 10-mers of HLA-A-0101, the model achieved an overall ROC AUC = 0.946, PR AUC = 0.812, and when measured separately by peptide length, ROC AUC = 0.960, PR AUC = 0.841 (for 9-mers), and ROC AUC = 0.859, PRAUC = 0.699 (for 10-mers). Those skilled in the art will note that PRAUC is a more reliable metric for distinguishing between methods because it has been shown to be less sensitive to the imbalance of the number of positive and negative instances in the data (which can lead to high ROC AUC values).

[0084] Thus, graph 1000 demonstrates that techniques for predicting MHC-peptide binding affinities for variable length peptide sequences using a recurrent neural network encoder and attention weighting compare favorably to results obtained from neural networks trained separately for each peptide length. In addition, neural networks trained using the techniques described herein can provide useful and improved affinity predictions for peptide sequences of other lengths, including those for which little or no affinity prediction data is available.

[0085] The systems, devices, and methods described herein can be implemented using digital circuits or using one or more computers using well-known computer processors, memory units, storage devices, computer software, and other components. Typically, a computer includes a processor for executing instructions and one or more memories for storing instructions and data. The computer may also include or be coupled to one or more mass storage devices, such as one or more magnetic disks, internal hard disks and removable disks, magneto-optical disks, optical disks, and the like.

[0086] The systems, devices and methods described herein can be implemented using computers operating in a client-server relationship. Typically, in such a system, the client computer is remote from the server computer and interacts via a network. The client-server relationship can be defined and controlled by computer programs running on the respective client and server computers.

[0087] Fig.11 A high-level block diagram of an exemplary client-server relationship that can be used to implement the systems, devices, and methods described herein is shown. The client-server relationship 1100 includes a client 1110 communicating with a server 1120 via a network 1130, and shows one possible division of the MHC-peptide binding affinity prediction task between the client 1110 and the server 1120. For example, according to the various embodiments described above, the client 1110 can obtain a test peptide sequence; access a neural network model trained using a training peptide sequence, such as via the server 1120; and generate an MHC-peptide binding affinity prediction using the trained neural network model. Server 1120 can then obtain multiple training peptide sequences; configure an encoder of a neural network model including a recurrent neural network (RNN) to process the input training peptide sequence to generate an attention-weighted output by applying the final hidden state of the RNN to the intermediate state output of the RNN, and linearly combine the attention-weighted output to generate an output of encoding fixed dimension; configure one or more fully connected layers of the neural network model to process the output of encoding fixed dimension to generate an MHC-peptide binding affinity prediction output; train the neural network model using multiple batches of training peptide sequences; and configure a computing device to use the trained neural network model to generate an MHC-peptide binding affinity prediction for a test peptide sequence.

[0088] Those skilled in the art will understand that Fig.11 The exemplary client-server relationship shown is only one of many client-server relationships that can be used to implement the systems, devices, and methods described herein. Fig.11 The client-server relationship shown should in no way be construed as limiting.Examples of client device 1110 may include a cell phone, kiosk, personal data assistant, tablet computer, robot, vehicle, web camera, or other type of computing device.

[0089] The systems, devices, and methods described herein may be implemented using a computer program product tangibly embodied in an information carrier, such as a non-transitory machine-readable storage device, for execution by a programmable processor; and the method steps described herein (including Figure 8 and Fig. 9One or more of the steps of the method may be implemented using one or more computer programs executable by such a processor. A computer program is a set of computer program instructions that can be used directly or indirectly in a computer to perform a certain activity or produce a certain result. A computer program may be written in any form of programming language, including compiled or interpreted languages, and it may be used in any form, including as a stand-alone program or a module, component, subroutine, or other unit suitable for a computing environment.

[0090] Fig.12 A high-level block diagram of an exemplary device that can be used to implement the systems, devices, and methods described herein is shown. Device 1200 includes a processor 1210 operably coupled to a persistent storage device 1220 and a main memory device 1230. Processor 1210 controls the overall operation of device 1200 by executing computer program instructions that define such operations. The computer program instructions can be stored in persistent storage device 1220 or other computer-readable media and loaded into main memory device 1230 when execution of the computer program instructions is desired. For example, training engine 410 and prediction engine 420 may include one or more components of computer 1200. Thus, Figure 8 and Fig. 9 The method steps may be defined by computer program instructions stored in the main memory device 1230 and / or the persistent storage device 1220 and controlled by the processor 1210 that executes the computer program instructions. For example, the computer program instructions may be implemented as a program that is programmed by a person skilled in the art to execute the method. Figure 8 and Fig. 9 Thus, by executing the computer program instructions, the processor 1210 performs the Figure 8 and Fig. 9 The device 1200 also includes one or more network interfaces 1280 for communicating with other devices via a network. The device 1200 may also include one or more input / output devices 1290 (e.g., a display, keyboard, mouse, speaker, button, etc.) for enabling user interaction with the device 1200.

[0091] Processor 1210 may include both general and special purpose microprocessors, and may be the only processor or one of multiple processors of device 1200. Processor 1210 may include one or more central processing units (CPUs) and one or more graphics processing units (GPUs), which may, for example, work separately from and / or multitask with one or more CPUs to accelerate processing, for example, for various deep learning and analysis applications described herein. Processor 1210, persistent storage device 1220, and / or main memory device 1230 may include, be supplemented by, or be incorporated in, one or more application specific integrated circuits (ASICs) and / or one or more field programmable gate arrays (FPGAs).

[0092] The persistent storage device 1220 and the main memory device 1230 each include a tangible non-transitory computer-readable storage medium. The persistent storage device 1220 and the main memory device 1230 may each include a high-speed random access memory, such as a dynamic random access memory (DRAM), a static random access memory (SRAM), a double data rate synchronous dynamic random access memory (DDR RAM), or other random access solid-state memory devices, and may include non-volatile memory, such as one or more magnetic disk storage devices (such as internal hard disks and removable disks), magneto-optical disk storage devices, optical disk storage devices, flash memory devices, semiconductor memory devices, such as erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), compact disk read-only memory (CD-ROM), digital versatile disk read-only memory (DVD-ROM) disks or other non-volatile solid-state storage devices.

[0093] The input / output device 1290 may include peripheral devices such as a printer, a scanner, a display screen, etc. For example, the input / output device 1290 may include a display device such as a cathode ray tube (CRT), a plasma or liquid crystal display (LCD) monitor for displaying information (e.g., DNA accessibility prediction results) to a user, a keyboard, and a pointing device such as a mouse or trackball by which a user can provide input to the device 1200.

[0094] Any or all of the systems and devices discussed herein, including training engine 410 and prediction engine 420 , may be executed by and / or incorporated in a device, such as device 1200 .

[0095] Those skilled in the art will recognize that actual computer or computer system implementations may have other structures and may include other components, and Fig.12 is a high level representation of some of the components of such a computer for purposes of illustration.

[0096] The foregoing description should be understood to be illustrative and exemplary in every respect, rather than restrictive, and the scope of the invention disclosed herein should not be determined according to the description, but rather determined by the claims interpreted according to the full breadth permitted by patent law. It should be understood that the embodiments shown and described herein are merely illustrative of the principles of the invention, and that various modifications may be implemented by those skilled in the art without departing from the scope and spirit of the invention. Various other feature combinations may be implemented by those skilled in the art without departing from the scope and spirit of the invention.

Claims

1. A method for predicting major histocompatibility complex (MHC)-peptide binding affinity implemented by a computing system, the method comprising: generating a plurality of variable length training peptide sequences, wherein at least some of the training peptide sequences are padded to correspond to fixed length peptide sequences; The neural network is trained by one or more computing devices to predict MHC-peptide binding affinity for MHC allele sequences, wherein for each training peptide sequence of a plurality of training peptide sequences, training the neural network comprises repeatedly: The training peptide sequences were input into the neural network; generating an output encoding a fixed dimension by processing the training peptide sequence, wherein the processing comprises applying a final hidden state of the neural network to an intermediate state of the neural network with attention weighted to one or more positions of the training peptide sequence; Generate training prediction output values ​​by processing the output of the encoded fixed dimension using a neural network; Depletion factors were determined by comparing the attention weights to known MHC-peptide binding affinity values ​​corresponding to the training peptide sequences; as well as updating at least one parameter of a set of parameters of the neural network based on the loss factor; inputting the test peptide sequence into the trained neural network; as well as A test prediction output value is generated by processing the test peptide sequence with the trained neural network, wherein the test prediction output value corresponds to an MHC-peptide binding affinity prediction of the test peptide sequence with respect to the MHC allele sequence.

2. The method of claim 1 , wherein applying the final hidden state to an intermediate state of a neural network comprises taking a dot product, a weighted product, or other function of the final hidden state and the intermediate state.

3. The method of claim 1 , further comprising applying weights learned by training a neural network to the final hidden state before applying the final hidden state to an intermediate state of the neural network.

4. The method of claim 1 , further comprising concatenating the final hidden state with a final hidden state of an encoder of a second neural network model before applying the final hidden state to an intermediate state of the neural network.

5. The method of claim 4, wherein the second neural network model is configured to predict MHC-peptide binding affinity for MHC allele input based on a parameter set of a trained neural network.

6. The method of claim 1, wherein the output of the encoded fixed dimension generated by the neural network comprises one or more positions each corresponding to an amino acid position of a training peptide sequence input into the neural network.

7. The method of claim 6, wherein each of the one or more positions of the output encoding the fixed dimensionality is a single value.

8. The method of claim 1, wherein the plurality of variable length training peptide sequences comprises two or more sequence lengths.

9. The method of claim 1, wherein the plurality of training peptide sequences are one of one-hot encoded, BLOSUM encoded, PAM encoded, or learned embedded encoded sequences.

10. The method according to claim 1, wherein each of the plurality of training peptide sequences has a length of 6-20 amino acids.

11. The method according to claim 1, wherein each of the plurality of training peptide sequences has a length of 10-30 amino acids.

12. The method of claim 1, wherein at least some of the training peptide sequences are filled using a single "0" or null value filling method.

13. The method of claim 1, wherein each training peptide sequence in the plurality of training peptide sequences is a positive MHC-peptide binding example.

14. The method of claim 1, wherein the test peptide sequence is 6-20 amino acids in length.

15. The method of claim 1, wherein the test peptide sequence has a length of 10-30 amino acids.

16. The method of claim 1, wherein the test peptide sequence has a sequence length that is different from a sequence length of at least one of the plurality of training peptide sequences.

17. The method of claim 1, wherein the test peptide sequence is one of a one-hot encoded, BLOSUM encoded, PAM encoded, or a learned embedded encoded sequence.

18. The method of claim 1, wherein the test prediction output value is a single prediction value.

19. The method of claim 18, wherein the single predictive value is associated with the likelihood of activating a T cell response to a tumor.

20. The method of claim 1, wherein generating a test prediction output value by a trained neural network comprises: generating an output encoding a fixed dimension by processing a test peptide sequence, wherein the processing comprises applying a final hidden state of a trained neural network to an intermediate state of a trained neural network with attention weighted to one or more positions of the test peptide sequence; and The test prediction output values ​​are generated by processing the output of the encoded fixed dimension using a trained neural network.

21. A computer program product embedded in a non-transitory computer readable medium, comprising instructions executable by a computer processor to predict major histocompatibility complex (MHC)-peptide binding affinity, the instructions, when executed by the processor, causing the processor to perform one or more steps comprising: generating a plurality of variable length training peptide sequences, wherein at least some of the training peptide sequences are padded to correspond to fixed length peptide sequences; The neural network is trained by one or more computing devices to predict MHC-peptide binding affinity for MHC allele sequences, wherein for each training peptide sequence of a plurality of training peptide sequences, training the neural network comprises repeatedly: The training peptide sequences were input into the neural network; generating an output encoding a fixed dimension by processing the training peptide sequence, wherein the processing comprises applying a final hidden state of the neural network to an intermediate state of the neural network with attention weighted to one or more positions of the training peptide sequence; Generate training prediction output values ​​by processing the output of the encoded fixed dimension using a neural network; Depletion factors were determined by comparing the attention weights to known MHC-peptide binding affinity values ​​corresponding to the training peptide sequences; as well as updating at least one parameter of a set of parameters of the neural network based on the loss factor; inputting the test peptide sequence into the trained neural network; as well as A test prediction output value is generated by processing the test peptide sequence with the trained neural network, wherein the test prediction output value corresponds to an MHC-peptide binding affinity prediction of the test peptide sequence with respect to the MHC allele sequence.

22. A computational system for predicting major histocompatibility complex (MHC)-peptide binding affinity, comprising: processor; a main memory device; Persistent storage device; a training engine executable on the processor in accordance with software instructions stored in one of the main memory device and the persistent storage device, whereby the training engine is operable to: A neural network is trained to predict MHC-peptide binding affinity of MHC allele sequences, wherein for each of a plurality of training peptide sequences of variable length, training the neural network comprises repeatedly performing the following operations: The training peptide sequences were input into the neural network; generating an output encoding a fixed dimension by processing the training peptide sequence, wherein the processing comprises applying a final hidden state of the neural network to an intermediate state of the neural network with attention weighted to one or more positions of the training peptide sequence; Generate training prediction output values ​​by processing the output of the encoded fixed dimension using a neural network; Depletion factors were determined by comparing the attention weights to known MHC-peptide binding affinity values ​​corresponding to the training peptide sequences; as well as updating at least one parameter of a set of parameters of the neural network based on the loss factor; a prediction engine in communication with the training engine, whereby the prediction engine is operable to: obtaining a test peptide sequence; inputting the test peptide sequence into the trained neural network; as well as A test prediction output value is generated by processing the test peptide sequence with the trained neural network, wherein the test prediction output value corresponds to an MHC-peptide binding affinity prediction of the test peptide sequence with respect to the MHC allele sequence.

23. A computing device comprising: processor; a main memory device; Persistent storage device; a prediction engine executable on the processor in accordance with software instructions stored in one of the main memory device and the persistent storage device, whereby the prediction engine is operable to: obtaining a test peptide sequence; The test peptide sequence is input into a neural network that is trained to predict the MHC-peptide binding affinity of an MHC allele sequence by repeatedly performing the following operations for each of a plurality of training peptide sequences of variable length: The training peptide sequences were input into the neural network; generating an output encoding a fixed dimension by processing the training peptide sequence, wherein the processing comprises applying a final hidden state of the neural network to an intermediate state of the neural network with attention weighted to one or more positions of the training peptide sequence; Generate training prediction output values ​​by processing the output of the encoded fixed dimension using a neural network; Depletion factors were determined by comparing the attention weights to known MHC-peptide binding affinity values ​​corresponding to the training peptide sequences; as well as updating at least one parameter of a set of parameters of the neural network based on the loss factor; as well as By processing the test peptide sequence using a neural network, a test prediction output value is generated, where The test prediction output value corresponds to the MHC-peptide binding affinity prediction of the test peptide sequence with respect to the MHC allele sequence.