Neoantigen prediction device, neoantigen prediction method, and computer program
The neoantigen prediction device uses embedding vectors and an avidity calculation model to identify neoantigens that can bind to HLA genes, addressing the challenge of predicting immune-inducing neoantigens for cancer treatment by determining binding affinity and aiding in cancer cell recognition.
Patent Information
- Application Number
- PCT/KR2024/005713
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2023-05-10
- Filing Date
- 2024-04-26
- Publication Date
- 2024-11-14
AI Technical Summary
Current methods fail to accurately identify neoantigens that can induce an immune response against cancer cells, as they struggle to predict the binding affinity of peptide sequences to HLA genes, which is crucial for antigen presentation and immune recognition.
A neoantigen prediction device and method that generates embedding vectors for peptide and HLA sequences using embedding layers and processing layers, followed by an avidity calculation model to determine the binding affinity, enabling the identification of neoantigens capable of binding and inducing an immune response.
Effectively predicts neoantigens that can bind to HLA genes, facilitating the recognition and attack of cancer cells by immune cells, thereby aiding in cancer treatment.
Smart Images

Figure KR2024005713_14112024_PF_FP_ABST
Abstract
Description
Neoantigen prediction device, neoantigen prediction method, and computer program
[0001] The present disclosure relates to a neoantigen prediction device, a neoantigen prediction method, and a computer program, and more specifically, to a neoantigen prediction device, a neoantigen prediction method, and a computer program, which generate a first embedding vector corresponding to a peptide sequence using a first embedding layer and a first processing layer, generate a second embedding vector corresponding to an HLA sequence using a second embedding layer and a second processing layer, input the first embedding vector and the second embedding vector into a machine-learned model to predict the binding affinity between a peptide and an HLA gene, and predict a neoantigen capable of binding based on the predicted result.
[0002] Among the somatic mutations that occur in cancer, neoantigens created by non-synonymous mutations are closely related to cancer immunity.
[0003] The binding of peptides to HLA structures is essential for antigen presentation on the cell surface. When peptides bind to HLA to form peptide-HLA complexes and are presented on the cell surface, CD8 T cells recognize the antigen, triggering an immune response. Theoretically, hundreds of types of binding occur between peptides derived from cancer cell mutations and HLA genes.
[0004] However, only a small percentage of immune cells, T cells, can properly function as antigens, allowing them to recognize and attack cancer cells. Therefore, accurately identifying neoantigens that drive cancer attacks is paramount. Therefore, identifying which peptide sequences bind best to which HLA sequences is the most crucial step in identifying neoantigens.
[0005] To solve the above-mentioned problem, the present invention can determine a neoantigen that can be used for the treatment of cancer tissue by using peptide sequence data obtained from a target cancer tissue and HLA sequence data of a patient.
[0006] The present invention generates a first embedding vector corresponding to a peptide sequence through a first embedding layer and a first processing layer, and generates a second embedding vector corresponding to an HLA sequence through a second embedding layer and a second processing layer, thereby outputting a binding affinity value for peptide sequence-HLA sequence data using the first embedding vector and the second embedding vector as input.
[0007] A neoantigen prediction method according to embodiments of the present disclosure comprises: a step in which a neoantigen prediction device obtains peptide sequence data extracted from a target cancer tissue; a step in which the neoantigen prediction device inputs HLA sequence data; a step in which the neoantigen prediction device converts the peptide sequence data into a first input vector having a fixed length and reflecting amino acid positions; a step in which the neoantigen prediction device inputs the first input vector into a first processing layer and outputs a first embedding vector for the peptide sequence; a step in which the neoantigen prediction device converts the HLA sequence data into a second input vector having a fixed length and reflecting amino acid positions; a step in which the neoantigen prediction device inputs the second input vector into a second processing layer and outputs a second embedding vector for the HLA sequence; The neoantigen prediction device may include a step of inputting the first input vector, the second input vector, the first embedding vector, and the second embedding vector into a binding affinity calculation model to output a binding affinity value corresponding to the binding affinity between a peptide sequence and an HLA sequence; and a step of generating information on a neoantigen capable of binding based on the binding affinity value.
[0008] The step of converting into the above first input vector may include one-hot encoding to change each amino acid in the peptide sequence data into a predetermined integer value, and inputting the encoded vector into an embedding layer learned using an artificial neural network to convert into a first input vector having a fixed length and combining information on the position of each amino acid.
[0009] The step of converting into the second input vector may include one-hot encoding to change each amino acid in the HLA sequence data into a predetermined integer value, and inputting the encoded vector into an embedding layer learned using an artificial neural network to convert into a first input vector having a fixed length and combining information on the position of each amino acid.
[0010] The step of outputting the first embedding vector may include using the first processing layer to hiddenly transform the first input vector into a first query vector, a first key vector, and a first value vector, and outputting a first embedding vector including values indicating relevance to other amino acids in the peptide sequence based on the first query vector, the first key vector, and the first value vector.
[0011] The step of outputting the first embedding vector may output the first embedding vector using a first processing layer learned by taking input vectors for peptide sequences as input.
[0012] The step of outputting the second embedding vector may include using the second processing layer to hiddenly transform the second input vector into a second query vector, a second key vector, and a second value vector, and outputting a second embedding vector including values indicating relevance to other amino acids in the peptide sequence based on the second query vector, the second key vector, and the second value vector.
[0013] The step of outputting the second embedding vector may output the second embedding vector using a second processing layer learned by inputting input vectors for HLA sequences as input.
[0014] The above binding force calculation model may take as input the first input vector and the first embedding vector for the peptide sequence data and the second input vector and the second embedding vector for the HLA sequence data, and calculate attention scores indicating a relationship between amino acids of the peptide sequence and amino acids of the HLA sequence, and output a binding force value based on the attention scores.
[0015] The target cancer tissue may include tumor cells obtained from or derived from multiple patients.
[0016] The target cancer tissue may include fresh or frozen tumor cells obtained from multiple patients.
[0017] The above target cancer tissue may be a tumor organoid cell obtained from multiple patients or cultured with tumor cells derived therefrom.
[0018] According to embodiments of the present disclosure, there is provided a peptide data input unit that obtains peptide sequence data extracted from a target cancer tissue and converts the peptide sequence data into a first input vector having a fixed length and reflecting the positions of amino acids; a peptide data processing unit that inputs the first input vector to a first processing layer and outputs a first embedding vector for the peptide sequence; an HLA data input unit that inputs HLA sequence data and converts the HLA sequence data into a second input vector having a fixed length and reflecting the positions of amino acids; an HLA data processing unit that inputs the second input vector to a second processing layer and outputs a second embedding vector for the HLA sequence; an avidity prediction unit that inputs the first input vector, the second input vector, the first embedding vector, and the second embedding vector into a binding affinity calculation model and outputs a binding affinity value corresponding to the binding affinity between the peptide sequence and the HLA sequence; A neoantigen prediction device is disclosed, which includes a neoantigen determination unit that generates information on a neoantigen capable of binding based on the binding affinity value.
[0019] A computer program according to an embodiment of the present invention can be stored in a medium to execute any one of the methods according to an embodiment of the present invention using a computer.
[0020] In addition, other methods for implementing the present invention, other systems, and computer-readable recording media recording a computer program for executing the above methods are further provided.
[0021] Other aspects, features and advantages other than those described above will become apparent from the following drawings, claims and detailed description of the invention.
[0022] According to embodiments of the present invention, a neoantigen prediction device and a neoantigen prediction method can be provided that determine neoantigens that can be used for the treatment of cancer tissue by using peptide sequence data obtained from a target cancer tissue and HLA sequence data of a patient.
[0023] According to embodiments of the present invention, a neoantigen prediction device and a neoantigen prediction method can be provided, which generate a first embedding vector corresponding to a peptide sequence through a first embedding layer and a first processing layer, and generate a second embedding vector corresponding to an HLA sequence through a second embedding layer and a second processing layer, and output a binding affinity value for peptide sequence-HLA sequence data using the first embedding vector and the second embedding vector as input.
[0024] FIG. 1 is a block diagram of a neoantigen prediction device according to embodiments of the present disclosure.
[0025] FIG. 2A is an exemplary diagram of an input vector for a peptide sequence according to embodiments of the present disclosure. FIG. 2B is an exemplary diagram of an embedding vector for a peptide sequence according to embodiments of the present disclosure.
[0026] FIG. 3 is an exemplary diagram of an input vector for an HLA sequence according to embodiments of the present disclosure.
[0027] FIG. 4 is an exemplary drawing of an embedding vector for a peptide sequence and an HLA sequence according to embodiments of the present disclosure.
[0028] Figure 5 is a flowchart of a neoantigen prediction method according to embodiments of the present disclosure.
[0029] FIG. 6 is an exemplary diagram of detailed components of an embedding layer and a processing layer for peptide sequence data according to embodiments of the present disclosure.
[0030] FIG. 7 is an exemplary diagram of detailed components of an embedding layer and a processing layer for HLA sequence data according to embodiments of the present disclosure.
[0031] FIG. 8 is an exemplary drawing of detailed components of a bonding force calculation model according to embodiments of the present disclosure.
[0032] Figure 9 is a table comparing the binding affinity values between peptides and HLA with predicted values.
[0033] The configuration and operation of the present invention will be described in detail with reference to embodiments of the present invention illustrated in the attached drawings below.
[0034] The present invention is capable of various modifications and embodiments. Specific embodiments are illustrated in the drawings and described in detail in the detailed description. The effects and features of the present invention, as well as the methods for achieving them, will become clearer with reference to the embodiments described in detail below, along with the drawings. However, the present invention is not limited to the embodiments disclosed below and can be implemented in various forms.
[0035] Hereinafter, embodiments of the present invention will be described in detail with reference to the attached drawings. When describing with reference to the drawings, identical or corresponding components are given the same drawing reference numerals, and redundant descriptions thereof will be omitted.
[0036] Hereinafter, the term “upper” or “upper” may include not only things that are directly above in contact, but also things that are above in a non-contact manner.
[0037] In the following examples, the terms first, second, etc. are not used in a limiting sense, but are used for the purpose of distinguishing one component from another.
[0038] In the examples below, singular expressions include plural expressions unless the context clearly indicates otherwise.
[0039] In the following examples, terms such as “include” or “have” mean that a feature or component described in the specification is present, and do not preclude the possibility that one or more other features or components may be added.
[0040] For convenience of explanation, the sizes of components in the drawings may be exaggerated or reduced. For example, the sizes and thicknesses of each component shown in the drawings are arbitrarily indicated for convenience of explanation, and thus the present invention is not necessarily limited to what is shown.
[0041] Additionally, terms such as “…part”, “…area”, etc. described in this specification may mean a unit that processes at least one function or operation.
[0042] Here, the target cancer tissue refers to the tissue that is the subject of the experiment. For example, the target cancer tissue may be a cancer tissue for which an antigen capable of inducing an immune response is to be detected. Preferably, the target cancer tissue may be a tumor cell, a tumor organoid, or a collection of normal cells.
[0043] Here, mutation refers to all phenomena in which the arrangement of the base sequence A (adenine), T (thymine), G (guanine), and C (cytosine) of the genes that contain the genetic information of each organism is changed to differ from the original genetic information of the species. These mutations cause structural changes on a small or large scale. Small mutations include point mutations that occur when a single base sequence is changed, and mutations in which base sequences are additionally inserted or deleted may also exist. Mutations that occur on a large scale and affect the structure include gene duplication, gene deletion, chromosomal inversion, interstitial deletion, chromosomal translocation, and loss of heterozygosity.
[0044] Mutations are broadly categorized into germline mutations and somatic mutations, depending on the type of cell in which they occur. Somatic mutations are genetic mutations that occur in somatic cells, also known as somatic mutations or cell mutations. They can be caused by genetic mutations or chromosomal abnormalities.
[0045] The occurrence of these mutations can alter the function of the protein produced by the corresponding gene, potentially leading to the loss of a specific function or the activation of a different one. Because these changes in protein function can trigger or accelerate the development of cancer, these mutations may be directly or indirectly linked to the development and progression of cancer.
[0046] The base sequence in DNA, which contains the genetic information of a living organism, is made up of A, T, G, and C. When three of these base sequences are combined in a row, they form a code that forms a specific amino acid, and when multiple of these codes are combined, they can be converted into a protein. Amino acids are made up of alanine (Ala), cysteine (Cys), aspartic acid (Asp), glutamic acid (Glu), phenylalanine (Phe), glycine (Gly), histidine (His), isoleucine (Ile), lysine (Lys), leucine (Leu), methionine (Met), asparagine (Asn), parolysine (Ply), proline (Pro), glutamine (Gln), arginine (Arg), serine (Ser), threonine (Thr), selenocysteine (Sec), valine (Val), tryptophan (Trp), and tyrosine (Tyr).
[0047] Peptide can refer to a peptide or polypeptide formed by amino acid sequences. Living organisms have an immune system to remove foreign substances that are not derived from the genetic information of each species, and in particular, among the external peptides, there are immunogenic peptides that can induce an immune response. Mutations that occur during the process of cancer development that differ from the original genetic information also produce such immunogenic peptides, and these immunogenic peptides can be capable of binding to HLA class I or II proteins after going through a series of processes within the immune system. Furthermore, immunogenic peptides can have a mutant amino acid sequence, and their amino acid length can be 25 or less, but is not limited thereto, and can be of various lengths.
[0048] Here, machine learning refers to using a model composed of multiple parameters and optimizing those parameters with given data. Machine learning can be categorized into supervised learning, unsupervised learning, and reinforcement learning, depending on the type of learning problem. Supervised learning learns the mapping between input and output and is applicable when input-output pairs are given as data.
[0049] In this specification, a neoantigen may be a peptide that elicits an immune response, i.e., an immunogenic peptide. Neoantigens may be induced by specific mutations in tumor cells and may be represented as epitopes on tumor cells. For clarity, immunogenic peptides will be referred to as neoantigens in the following description.
[0050] In this specification, an artificial neural network is a statistical learning algorithm inspired by biological neural networks in machine learning and cognitive science, and refers to a model in which artificial neurons formed by combining synapses form a network and change the binding strength of synapses through learning to have problem-solving capabilities.
[0051] In this specification, a vector means data in one dimension, and can be expressed as a matrix in multiple dimensions.
[0052] According to an embodiment of the present disclosure, a neoantigen prediction device may receive information on the location and expression profile of important mutations for each gene based on exome sequencing data obtained through Next Generation Sequencing (NGS) analysis of patient tissue and RNA expression level data obtained through RNA sequencing. This data may be input separately for peptide sequences and HLA sequences.
[0053] The neoantigen prediction device can predict the binding affinity between peptides and HLA genes.
[0054] FIG. 1 is a block diagram of a neoantigen prediction device according to embodiments of the present disclosure.
[0055] The neoantigen prediction device (100) can predict neoantigens that can be used for the treatment of target cancer tissue based on genetic data obtained from target cancer tissue and genetic data obtained from the patient's general tissue.
[0056] It may include a peptide data input unit (110), a peptide data processing unit (120), an HLA data input unit (130), an HLA data processing unit (140), and / or a binding affinity prediction unit (150), a neoantigen determination unit (160).
[0057] The peptide data input unit (110) can convert peptide sequence data into a first input vector using a predetermined encoding method. The peptide data input unit (110) can encode peptide sequence data using a predetermined encoding method and input it into an embedding layer to generate a first input vector.
[0058] More specifically, the peptide data input unit (110) can convert each amino acid included in the peptide sequence data into a predetermined integer value. The peptide data input unit (110) can one-hot encode each amino acid included in the peptide sequence data into a respective integer. For example, if the peptide sequence includes 20 amino acids, vectors for the 20 amino acids can be generated and the vectors can be converted into a single matrix. The vectors for the amino acids can include a vector set to 1 (presence) or 0 (absence) for the presence or absence of the first amino acid. 20 vectors can be generated for the 20 amino acids. The 20-dimensional vectors generated in this way can be converted into a single matrix.
[0059] The peptide data input unit (110) can convert a one-hot encoded vector into a vector of a fixed length through a first embedding layer. Here, the first embedding layer refers to a layer trained to input a vector expressing amino acids included in peptide sequence data and output an input vector of a fixed length that combines information on the position of each amino acid. The first embedding layer may refer to a layer trained using an artificial neural network. It can convert peptide sequence data of different lengths into an input vector of a fixed length. Through the first embedding layer, the input vector can have a dimension (n+1) that adds a dimension (1) expressing position information to the dimension (n) according to the number of amino acids.
[0060] Through the above process, the peptide data input unit (110) can generate a first input vector for peptide sequence data.
[0061] The peptide data processing unit (120) can input a first input vector into a first processing layer and output a first embedding vector to which the relevance (attention) between amino acids in the peptide sequence is applied. The peptide data processing unit (120) can normalize the values of the first embedding vector to values between 0 and 1.
[0062] The first processing layer can be trained with amino acids included in the first input vector and output a vector for values according to the correlation between amino acids. The first processing layer can convert the first input vector into a first query vector, a first key vector, and a first value vector, input the first query vector, the first key vector, and the first value vector to calculate an attention score, and output a first embedding vector reflecting the attention score. The first processing layer can convert the first query vector, the first key vector, and the first value vector, which express the correlation between amino acids included in the first input vector. The first query vector, the first key vector, and the first value vector are vectors that implicitly include information about amino acids included in a peptide sequence, and may be data of the same dimension as the input vector.
[0063] The first query vector refers to an information vector about amino acids in the first input vector. The first key vector refers to a vector generated for comparison with the first query vector. The first value vector may refer to a vector reflecting the relevance resulting from the comparison of the first query vector and the first key vector. The first query vector, the first key vector, and the first value vector may be vectors generated from a hidden state of the first processing layer. The first processing layer may output an attention score matrix for the first query vector, the first key vector, and the first value vector generated from the hidden state. The first processing layer may be trained by relevance values (attention values) between predetermined, pre-trained amino acids. The first processing layer may be generated using an artificial neural network that mimics human attention. The artificial neural network that mimics human attention is trained by a machine learning technique that refers back to an important input part, and may be trained by a dot product attention technique or a multi-head attention technique. The attention score matrix may include values representing relationships between amino acids included in the peptide sequence data.
[0064] The attention score matrix can be produced through operations between the first query vector, the first key vector, and the first value vector, for example, through operations such as dot product and square root. The attention score matrix can be converted into a probability value.
[0065] The first processing layer can classify the amino acids contained in the first input vector into multiple groups using an attention score matrix and output a first embedding vector containing correlation values between the values for amino acids belonging to each group. The first embedding vector can be output using a multi-head technique.
[0066] The peptide data processing unit (120) can generate an embedding vector for a peptide sequence using a layer that has learned the interactions and relationships between amino acids at different positions within a peptide. The embedding vector can include values for the relationships between different amino acids. As illustrated in FIG. 2B, the embedding vector has columns according to amino acid type and rows according to position, and can include values proportional to the relationship with other amino acids. For example, among the amino acids at position 1, the amino acid in column N can be output as having the highest relationship value of 0.9 with other amino acids.
[0067] The HLA data input unit (130) can convert HLA sequence data into an embedding vector using a predetermined encoding method. The HLA data input unit (130) can encode HLA sequence data using a predetermined encoding method and input it into a second embedding layer to generate a second input vector.
[0068] More specifically, the HLA data input unit (130) can convert each amino acid included in the HLA sequence data into a predetermined integer value. The HLA data input unit (130) can one-hot encode each amino acid included in the HLA sequence data into a respective integer. For example, if the HLA sequence includes 20 amino acids, vectors for the 20 amino acids can be generated and the vectors can be converted into a single matrix. The vectors for the amino acids can include a vector set to 1 (presence) or 0 (absence) for the presence or absence of the first amino acid. 20 vectors can be generated for the 20 amino acids. The 20-dimensional vectors generated in this way can be converted into a single matrix.
[0069] The HLA data input unit (130) can convert a one-hot encoded vector into a vector of a fixed length through a first embedding layer. Here, the first embedding layer refers to a layer trained to input a vector expressing amino acids included in HLA sequence data and output an input vector of a fixed length that combines information on the position of each amino acid. The first embedding layer may refer to a layer trained using an artificial neural network. It can convert peptide sequence data of different lengths into an input vector of a fixed length. Through the first embedding layer, the input vector can have a dimension (n+1) that adds a dimension (1) expressing position information to the dimension (n) according to the number of amino acids.
[0070] Through the above process, the HLA data input unit (130) can generate a second input vector for HLA sequence data.
[0071] The HLA data processing unit (140) can input a second input vector into a second processing layer and output a second embedding vector to which the relevance (attention) between amino acids for the HLA sequence is applied. The HLA data processing unit (140) can normalize the values of the second embedding vector to values between 0 and 1.
[0072] The second processing layer may refer to a layer trained with amino acids included in the second input vector. The second processing layer may convert the second input vector into a second query vector, a second key vector, and a second value vector, input the second query vector, the second key vector, and the second value vector to calculate an attention score, and output a second embedding vector reflecting the attention score. The second processing layer may convert the second query vector, the second key vector, and the second value vector that express the relationship between amino acids included in the third embedding vector. The second query vector, the second key vector, and the second value vector are vectors that implicitly include information about amino acids included in the HLA sequence.
[0073] The second query vector refers to an information vector about amino acids in the second input vector. The second key vector refers to a vector generated for comparison with the second query vector. The second value vector may refer to a vector reflecting the relevance resulting from the comparison of the second query vector and the second key vector. The second query vector, the second key vector, and the second value vector may be vectors generated from a hidden state of the second processing layer. The second processing layer may output an attention score matrix for the second query vector, the second key vector, and the second value vector generated from the hidden state. The second processing layer may be trained by relevance values (attention values) between amino acids in a predetermined, pre-trained HLA sequence. The second processing layer may be generated using an artificial neural network that mimics human attention. The artificial neural network that mimics human attention is trained by a machine learning technique that refers back to an important input part, and may be trained by a dot product attention technique or a multi-head attention technique. The attention score matrix may include values representing relationships between amino acids included in the HLA sequence data.
[0074] The attention score matrix can be calculated through operations between the second query vector, the second key vector, and the second value vector. For example, it can be calculated through operations such as dot product and square root, but is not limited thereto and can be calculated in various ways. The attention score matrix can be converted into a probability value.
[0075] The second processing layer can classify the amino acids contained in the second input vector into multiple groups using the attention score matrix and output a second embedding vector containing correlation values between the values of amino acids belonging to each group. The second embedding vector can be output using a multi-head technique.
[0076] The HLA data processing unit (140) can generate an embedding vector for an HLA sequence using a layer that has learned the interactions and relationships between amino acids at different positions among HLAs. The embedding vector can include values for the relationships between amino acids.
[0077] The binding affinity prediction unit (150) can input a first input vector and a first embedding vector for peptide sequence data and a second input vector and a second embedding vector for HLA sequence data into a binding affinity calculation model to calculate a binding affinity value between the second peptide sequence and the HLA sequence. The binding affinity value can be a value between 0 and 1.
[0078] The binding affinity calculation model refers to a model trained using as input a first input vector and a first embedding vector for peptide sequence data, and a second input vector and a second embedding vector for HLA sequence data. The binding affinity calculation model can calculate the binding affinity based on the first embedding vector and the second embedding vector in which a relevance value is reflected for the peptide sequence and the HLA sequence. In addition, the binding affinity calculation model can calculate the binding affinity based on the first input vector and the second input vector in which a relevance value is not reflected. The binding affinity calculation model can calculate attention scores indicating a relationship between amino acids of the peptide sequence and amino acids of the HLA sequence by using as input the first input vector and the first embedding vector for peptide sequence data, and the second input vector and the second embedding vector for HLA sequence data, and can output a binding affinity value between amino acids of the peptide sequence and amino acids of the HLA sequence based on the attention scores.
[0079] Calculating attention scores involves combining the first embedding vector for peptide sequence data and the second embedding vector for HLA sequence data to produce three vectors, and calculating based on the three vectors. Calculating attention scores can be done by applying a predetermined function to any two of the three vectors, and can be output through Softmax, for example.
[0080] The three vectors are a query vector, a key vector, and a value vector, which are produced for each amino acid included in the peptide sequence data and each amino acid included in the HLA sequence data, using the first embedding vector for the peptide sequence data and the second embedding vector for the HLA sequence data.
[0081] The binding force calculation model can output a third query vector, a third key vector, and a third value vector by multiplying a binding vector including a first embedding vector for peptide sequence data and a second embedding vector for HLA sequence data by a predetermined matrix.
[0082] The avidity calculation model can output whether there is a binding affinity between amino acids contained in a peptide sequence and amino acids contained in an HLA sequence. The avidity matrix can be set to a value of 1 at the position corresponding to the presence or absence of a binding affinity between the third amino acid of the peptide sequence and the fourth amino acid of the HLA sequence, if there is a binding affinity between the third amino acid of the peptide sequence and the fourth amino acid of the HLA sequence.
[0083] The neoantigen determination unit (160) can generate neoantigen information between the peptide sequence and the HLA sequence based on the calculated binding affinity value input into the binding affinity calculation model.
[0084] If the avidity value is 1, the peptide can act as a neoantigen for HLA.
[0085] FIG. 2A is an exemplary diagram of an input vector for a peptide sequence according to embodiments of the present disclosure. FIG. 2B is an exemplary diagram of an embedding vector for a peptide sequence according to embodiments of the present disclosure.
[0086] An embedding vector for a peptide sequence may represent the type and position of each amino acid as two-dimensional coordinates, and may include values for relevance (attention) according to rows and columns. As illustrated in Figure 2b, the embedding vector is expressed by converting it into a color corresponding to the value, with a darker color indicating a larger value. A larger value may indicate a higher relevance to other amino acids. Here, a high relevance to other amino acids may indicate an amino acid with high binding affinity. Figure 2b represents an embedding vector for a peptide sequence, but an embedding vector for an HLA sequence can also be generated in a similar manner.
[0087] FIG. 3 is an exemplary diagram of an input vector for an HLA sequence according to embodiments of the present disclosure.
[0088] FIG. 4 is an exemplary drawing of an embedding vector for a peptide sequence and an HLA sequence according to embodiments of the present disclosure.
[0089] Figure 5 is a flowchart of a neoantigen prediction method according to embodiments of the present disclosure.
[0090] The neoantigen prediction device (100) can input peptide sequence data extracted from target cancer tissue and HLA sequence data, which is an antigen of the patient.
[0091] In S110, the neoantigen prediction device (100) can convert peptide sequence data into a first input vector having a fixed length and reflecting the position of an amino acid. The peptide data input unit (110) can convert each amino acid included in the peptide sequence data into a predetermined integer value. The peptide data input unit (110) can one-hot encode each amino acid included in the peptide sequence data into a respective integer. The peptide data input unit (110) can convert the one-hot encoded vector into a first input vector that is a vector of a fixed length and combines information on the position of each amino acid through a first embedding layer. The first input vector is as illustrated in FIG. 2.
[0092] In S120, the neoantigen prediction device (100) can input a first input vector into a first processing layer and output a first embedding vector for a peptide sequence. The neoantigen prediction device (100) can learn with amino acids included in the first input vector using the first processing layer and output a vector for values according to the correlation between amino acids. At this time, the hidden layer of the first processing layer can generate a first query vector, a first key vector, and a first value vector for the first input vector, and output a first embedding vector that reflects an attention score for the first input vector based on the first query vector, the first key vector, and the first value vector. The first processing layer can output the first embedding vector using a multi-head attention technique. The first embedding vector is as illustrated in FIG. 2B.
[0093] In S130, the neoantigen prediction device (100) can convert HLA sequence data into a second input vector having a fixed length and reflecting the positions of amino acids. The neoantigen prediction device (100) can one-hot encode each amino acid included in the HLA sequence data into a respective integer. The neoantigen prediction device (100) can convert the one-hot encoded vector into an input vector having a fixed length and combining information on the positions of amino acids through a first embedding layer. The second input vector is as illustrated in FIG. 3.
[0094] In S140, the neoantigen prediction device (100) can input a second input vector into a second processing layer and output a second embedding vector to which the relevance (attention) between amino acids for the HLA sequence is applied. The neoantigen prediction device (100) can convert the second input vector into a second query vector, a second key vector, and a second value vector using the second processing layer, input the second query vector, the second key vector, and the second value vector to calculate an attention score, and output a second embedding vector that reflects the attention score.
[0095] In S150, the neoantigen prediction device (100) can input the first input vector, the second input vector, the first embedding vector, and the second embedding vector into the binding affinity calculation model to output a binding affinity value corresponding to the binding affinity between the peptide sequence and the HLA sequence. At this time, the binding affinity calculation model can input a vector in which the first embedding vector and the second embedding vector are combined and a vector in which the first input vector and the second input vector are combined. The vector in which the first embedding vector and the second embedding vector are combined is as illustrated in FIG. 4. The binding affinity calculation model can output whether there is a binding affinity between amino acids included in the peptide sequence and amino acids included in the HLA sequence. The binding affinity matrix can be set to a value of 1 at a position for whether there is a binding affinity between the third amino acid and the fourth amino acid when there is a binding affinity between the third amino acid of the peptide sequence and the fourth amino acid of the HLA sequence.
[0096] In S160, the neoantigen prediction device (100) can calculate whether a peptide sequence corresponds to a neoantigen based on the binding affinity value.
[0097] FIG. 6 is an exemplary diagram of detailed components of an embedding layer and a processing layer for peptide sequence data according to embodiments of the present disclosure.
[0098] As illustrated in Figure 6, the neoantigen prediction device can process peptide sequence data by embedding it and encoding it by position.
[0099] Peptide sequence data can be converted into a vector in which each amino acid is one-hot encoded with a fixed integer value. This converted vector can be converted into a fixed-length vector through an embedding layer (Sequence Embedding). This fixed-length vector can then be converted into an input vector through positional encoding. The embedding layer (S11) can be implemented by including sequence embedding, which converts it into a fixed-length vector, and positional encoding, which combines positional information for amino acids. The positional encoding value can be calculated using the mathematical formula below.
[0100]
[0101]
[0102] pos and i are position values in the one-hot encoded vector, and d_model represents the dimension of the one-hot encoded vector. pos is a value set from 1, such as 1, 2, 3, 4, etc., and i can be a value set from 0, such as 0, 1, 2, 3, etc. That is, if pos is 2, i can be set to pos-1=1.
[0103] 2i means that the place value is even and can be calculated by PE(pos, 2i), and 2i+1 means that the place value is odd and can be calculated by PE(pos, 2i+1).
[0104] The input vector output through the embedding layer (S11) can be input to S12 and output as an embedding vector.
[0105] S12 can take an input vector as input and output an embedding vector containing the correlation values of each amino acid in the peptide sequence with other amino acids.
[0106] In S12, a query vector (Q), a key vector (K), and a value vector (V) can be generated by processing the input vector with a predetermined function. The processing layer (Masked Multi-head Peptide Self-Attention) can use the generated Q, K, and V to calculate an attention score, which is a correlation between the types and positions of each amino acid. The attention score can be calculated using the mathematical formula below.
[0107]
[0108] d_k represents the dimension of the key vector. The dimension is determined by the number of values contained in the vector. If the number of values contained in the vector is 10, d_k is 10.
[0109] The processing layer (Masked Multi-head Peptide Self-Attention) can calculate the attention score by taking the inner product of the transposes of the Q and K vectors and dividing this value by the square root. Softmax can be applied as described above to calculate the attention score.
[0110] The vector containing the attention scores can undergo further normalization.
[0111] FIG. 7 is an exemplary diagram of detailed components of an embedding layer and a processing layer for HLA sequence data according to embodiments of the present disclosure.
[0112] HLA sequence data can be converted into a vector in which each amino acid is one-hot encoded with a fixed integer value. The vector thus converted can be converted into a fixed-length vector through a second embedding layer (Sequence Embedding). The fixed-length vector can be converted into an input vector through positional encoding. The second embedding layer (S21) can be implemented by including sequence embedding, which converts into a fixed-length vector, and positional encoding, which combines positional information for amino acids. The positional encoding can be calculated using the following mathematical formula.
[0113]
[0114]
[0115] pos and i are position values in the one-hot encoded vector, and d_model represents the dimension of the one-hot encoded vector. pos is a value set from 1, such as 1, 2, 3, 4, etc., and i can be a value set from 0, such as 0, 1, 2, 3, etc. That is, if pos is 2, i can be set to pos-1=1.
[0116] 2i means that the place value is even and can be calculated by PE(pos, 2i), and 2i+1 means that the place value is odd and can be calculated by PE(pos, 2i+1).
[0117] The input vector output through the embedding layer (S21) can be input to S22 and output as an embedding vector.
[0118] S22 can take an input vector as input and output an embedding vector containing the correlation values of each amino acid in the HLA sequence with other amino acids.
[0119] In S22, a query vector (Q), a key vector (K), and a value vector (V) can be generated by processing the input vector with a predetermined function. The processing layer (Masked Multi-head Peptide Self-Attention) can use the generated Q, K, and V to calculate an attention score, which is a correlation between the types and positions of each amino acid. The attention score can be calculated using the mathematical formula below.
[0120]
[0121] d_k represents the dimension of the key vector. The dimension is determined by the number of values contained in the vector. If the number of values contained in the vector is 10, d_k is 10.
[0122] The processing layer (Masked Multi-head Peptide Self-Attention) can calculate the attention score by taking the inner product of the transposes of the Q and K vectors and dividing this value by the square root. Softmax can be applied as described above to calculate the attention score.
[0123] The vector containing the attention scores can undergo further normalization.
[0124] FIG. 8 is an exemplary drawing of detailed components of a bonding force calculation model according to embodiments of the present disclosure.
[0125] As shown in Fig. 8, the binding force between the peptide sequence and the HLA sequence can be predicted by inputting the first embedding vector applying the relevance values for amino acids from the peptide sequence and the second embedding vector applying the relevance values for amino acids from the HLA sequence into S31.
[0126] The binding force calculation model can be implemented by including multiple layers such as Masked Multi-head peptide-HLA attention, Dense Layer (128), Dense Layer (64), Dense Layer (32), Sigmoid or Softmax.
[0127] Masked Multi-head peptide-HLA attention can produce a query vector, a key vector, and a value vector based on the first embedding vector and the second embedding vector, and can produce a matrix for the correlation values (attention) between amino acids between peptides and HLA using the query vector, key vector, and value vector.
[0128] In Fig. 8, Dense Layer (128) is a layer that has 128 neurons and uses ReLU as an activation function, Dense Layer (64) is a layer that has 64 neurons and uses ReLU as an activation function, Dense Layer (32) is a layer that has 32 neurons and uses ReLU as an activation function, and Sigmoid or Softmax can be a layer that has 1 neuron and uses Sigmoid or Softmax as an activation function.
[0129] By passing through Dense Layer (128), Dense Layer (64), Dense Layer (32), Dense Layer (1), Sigmoid or Softmax, the predictive power probability value for the binding strength between amino acids can be converted.
[0130] Figure 9 is a table comparing the binding affinity values between peptides and HLA with predicted values.
[0131] Predicted binding affinity score refers to a binding affinity value output according to an embodiment of the present disclosure. The binding affinity score value may be calculated by another method, for example, a value calculated by a method such as NetMHCpan, DeepHLApan, or MHCnuggets.
[0132] In addition, although the preferred embodiments of the present invention have been illustrated and described above, the present invention is not limited to the specific embodiments described above, and various modifications can be made by a person having ordinary skill in the art to which the present invention pertains without departing from the gist of the present invention claimed in the claims, and such modifications should not be understood individually from the technical idea or prospect of the present invention.
[0133] Therefore, the spirit of the present invention should not be limited to the embodiments described above, and not only the scope of the patent claims described below, but also all scopes equivalent to or equivalently modified from the scope of the patent claims are considered to fall within the scope of the spirit of the present invention.
Claims
1. A step in which a neoantigen prediction device acquires peptide sequence data extracted from a target cancer tissue; The above neoantigen prediction device comprises a step of inputting HLA sequence data; The above neoantigen prediction device converts the peptide sequence data into a first input vector having a fixed length and reflecting the position of an amino acid; The above neoantigen prediction device inputs the first input vector into a first processing layer and outputs a first embedding vector for the peptide sequence; A step of the neoantigen prediction device converting the HLA sequence data into a second input vector having a fixed length and reflecting the position of amino acids; A step in which the neoantigen prediction device inputs the second input vector into a second processing layer and outputs a second embedding vector for the HLA sequence; The neoantigen prediction device inputs the first input vector, the second input vector, the first embedding vector, and the second embedding vector into a binding force calculation model to output a binding force value corresponding to the binding force between the peptide sequence and the HLA sequence; and A method for predicting a neoantigen, comprising: a step of generating information on a neoantigen capable of binding based on the binding affinity value by the neoantigen prediction device.
2. In paragraph 1, The step of converting to the above first input vector is, A method for predicting neoantigens, which performs one-hot encoding to change each amino acid in the above peptide sequence data into a fixed integer value, and inputs the encoded vector into an embedding layer learned using an artificial neural network to convert it into a first input vector having a fixed length and combining information about the position of each amino acid.
3. In paragraph 1, The step of converting to the second input vector is as follows: A neoantigen prediction method, which performs one-hot encoding to change each amino acid in the above HLA sequence data to a fixed integer value, and inputs the encoded vector into an embedding layer learned using an artificial neural network to convert it into a first input vector having a fixed length and combining information on the position of each amino acid.
4. In paragraph 1, The step of outputting the above first embedding vector is A method for predicting neoantigens, wherein the first processing layer is used to hiddenly transform the first input vector into a first query vector, a first key vector, and a first value vector, and outputs a first embedding vector including values indicating relevance to other amino acids in the peptide sequence based on the first query vector, the first key vector, and the first value vector.
5. In paragraph 4, The step of outputting the above first embedding vector is A method for predicting neoantigens, which outputs a first embedding vector using a first processing layer learned by inputting input vectors for peptide sequences.
6. In paragraph 1, The step of outputting the second embedding vector is A method for predicting neoantigens, wherein the second input vector is hiddenly converted into a second query vector, a second key vector, and a second value vector using the second processing layer, and a second embedding vector is output including values indicating relevance to other amino acids in the peptide sequence based on the second query vector, the second key vector, and the second value vector.
7. In paragraph 6, The step of outputting the second embedding vector is A method for predicting neoantigens, which outputs a second embedding vector using a second processing layer learned by inputting input vectors for HLA sequences.
8. In paragraph 1, The above bonding force calculation model is A neoantigen prediction method, wherein the first input vector and the first embedding vector for the peptide sequence data and the second input vector and the second embedding vector for the HLA sequence data are input, attention scores indicating the relationship between amino acids of the peptide sequence and amino acids of the HLA sequence are calculated, and an avidity value is output based on the attention scores.
9. In paragraph 1, The above target cancer tissue is A method for predicting neoantigens, comprising tumor cells obtained from or derived from multiple patients.
10. In paragraph 1, The above target cancer tissue is A method for predicting neoantigens, comprising fresh or frozen tumor cells obtained from multiple patients.
11. In paragraph 1, The above target cancer tissue is A method for predicting neoantigens, wherein the tumor organoid cells are cultured in a culture medium obtained from or derived from multiple patients.
12. A peptide data input unit that obtains peptide sequence data extracted from a target cancer tissue and converts the peptide sequence data into a first input vector having a fixed length and reflecting the position of an amino acid; A peptide data processing unit that inputs the first input vector to a first processing layer and outputs a first embedding vector for the peptide sequence; An HLA data input unit that inputs HLA sequence data and converts the HLA sequence data into a second input vector having a fixed length and reflecting the position of an amino acid; An HLA data processing unit that inputs the second input vector to a second processing layer and outputs a second embedding vector for the HLA sequence; An avidity prediction unit that inputs the first input vector, the second input vector, the first embedding vector, and the second embedding vector into a binding avidity calculation model and outputs a binding avidity value corresponding to the binding avidity between a peptide sequence and an HLA sequence; and A neoantigen prediction device, comprising a neoantigen determination unit that generates information on a neoantigen capable of binding based on the above binding affinity value.
13. A computer program stored in a computer-readable storage medium to execute the method of paragraph 1 using a computer.
Citation Information
Patent Citations
Method for predicting neoantigen using a peptide sequence and HLA allele sequence and computer program
KR102159921B1
Method and computer program to predict neoantigen for pan-HLA class ii allele sequences and pan-length of peptides
KR102330099B1
Method and computer program to predict neoantigen by standardizing the length of peptides of various lengths by folding
KR102516826B1
