A method and system for screening of a population of tcrs against hepatocellular carcinoma for broadly recognizing candidate peptides

Abstract

The present application relates to the field of bioinformatics and tumor antigen screening, and discloses a screening method and system for a hepatocellular carcinoma-oriented group TCR widely recognized candidate peptide segment. The present application aims to solve the problem that the existing method is limited to single TCR evaluation and does not consider the clonal abundance and VJ diversity, resulting in the lack of group-level recognition potential and generalization ability of the candidate peptide segment. The scheme comprises the following steps: obtaining immune sequence data to construct a prediction model comprising an encoder and a cross-attention module, pre-training the prediction model, and outputting a combination probability through two-stage supervised training; obtaining a candidate peptide segment and a TCR library containing abundance and VJ genes, constructing a combination input model to obtain a probability matrix; filtering low probability, determining weight in combination with clonal abundance, calculating strong binding, breadth and VJ diversity score, and obtaining a group TCR recognition comprehensive score; and outputting the candidate peptide segment according to the sorting. The present application is used for in vitro data processing for non-disease diagnosis and treatment purposes.

Description

Technical Field

[0001] This invention relates to the fields of bioinformatics, machine learning, immunoinformatics, and tumor antigen candidate peptide screening, and particularly to a method and system for broadly identifying candidate peptides for population TCR in hepatocellular carcinoma. Background Technology

[0002] Hepatocellular carcinoma (HCC) is one of the most common types of malignant tumors. Vaccine candidate design based on tumor-associated antigens, virus-associated antigens, or mutated neoantigens typically requires screening a large number of candidate peptides to identify those more likely to be recognized by the immune system. Whether a candidate peptide can become a valuable immune candidate usually depends on its ability to be presented by HLA molecules and recognized by T-cell receptors (TCRs).

[0003] Existing computational methods fall into two categories: one primarily predicts the binding or presentation capacity of a peptide with HLA, while the other primarily predicts the binding probability of a peptide with a specific TCRCDR3 sequence. These methods can assess interactions between specific pairings, but they still have limitations in designing candidate peptides for population-based vaccines.

[0004] First, predicting whether a single TCR binds to a single peptide can easily limit the screening to a single patient, a single TCR clone, or a few known CDR3 sequences. For scenarios where the goal is to screen for candidate peptides that can be recognized by a wider range of hepatocellular carcinoma-related TCRs, it is more necessary to assess the overall identifiability of the peptide within the population-level HCCTCR recognition space.

[0005] Second, HCC-related TCR data exhibit differences in clonal abundance, V / J gene usage, and patient origin. If only the number of TCRs above a threshold is counted, the importance of high-frequency TCR clones may be overlooked; if only the highest binding probability is considered, peptides recognized by only a few TCRs may be screened; and if V / J combination diversity is not taken into account, peptides primarily dependent on a single clonal background may be misjudged as having broad recognition potential.

[0006] Third, general peptide-HLA or peptide-TCR models may not be adequately suited to hepatocellular carcinoma (HCC) scenarios. HCC-associated antigens, HBV / HCV-associated antigens, liver immune peptides, tumor-infiltrating TCRs, and vaccine-response-related TCRs have specific distributions; directly using general models may result in insufficient generalization ability in specific HCC peptide or TCR contexts.

[0007] Therefore, a candidate peptide screening method is needed that can first construct an immune binding prediction model for hepatocellular carcinoma (HCC), then combine candidate peptides with a large number of HCC-related TCR CDR3s for prediction, and further integrate binding strength, TCR clonal abundance, recognition breadth, and VJ diversity. The goal of this method is not to determine whether a specific CDR3 in a particular patient recognizes a particular peptide, but rather to assess, with the peptide as the primary focus, whether that peptide is more likely to be recognized by multiple HCC-related CDR3s in a broad sense. Summary of the Invention

[0008] The purpose of this invention is to solve the technical problems of existing prediction methods that are limited to the evaluation of a single TCR sequence and do not fully consider TCR clonal abundance, VJ diversity and hepatocellular carcinoma scenario specificity, resulting in the lack of broad recognition potential of screened candidate peptides, susceptibility to single clonal bias and insufficient generalization ability.

[0009] To achieve the above objectives, the present invention employs the following technical means:

[0010] This invention provides a method for screening candidate peptides for broad TCR recognition in hepatocellular carcinoma populations, used for in vitro data processing for non-disease diagnosis and treatment purposes, including:

[0011] Step 1: Obtain immune sequence training data, which includes HLA sequence data, immune peptide sequence data, T cell receptor CDR3 sequence data, and hepatocellular carcinoma-associated antigen or T cell receptor data.

[0012] Step 2: Construct a hepatocellular carcinoma targeted immune binding prediction model based on the immune sequence training data. The hepatocellular carcinoma targeted immune binding prediction model includes an HLA encoder, a peptide encoder, a TCR encoder, a peptide-HLA cross-attention module, a peptide-TCR cross-attention module, and a classification output module.

[0013] The HLA encoder, peptide encoder, and TCR encoder were pre-trained using the immune sequence training data, respectively, to achieve domain-adaptive pre-training.

[0014] The pre-trained encoder is connected to the cross-attention module through a projection adaptation layer, and two-stage supervised training is performed using paired data to obtain a hepatocellular carcinoma targeted immune binding prediction model that can output the probability of peptide binding to TCRCDR3.

[0015] Step 3: Obtain the candidate peptide set and the hepatocellular carcinoma-related TCRCDR3 library. Each TCR record in the hepatocellular carcinoma-related TCRCDR3 library includes the CDR3 amino acid sequence, V gene fragment, J gene fragment, and clonal abundance information.

[0016] Step 4: Construct peptide-TCR combinations for each candidate peptide in the candidate peptide set and each TCR record in the hepatocellular carcinoma-related TCRCDR3 library, and input the peptide-TCR combinations into the hepatocellular carcinoma-directed immune binding prediction model to obtain the binding probability between each candidate peptide and each TCR record, and construct a binding probability matrix based on the binding probability.

[0017] Step 5: Filter the binding probabilities in the binding probability matrix that are below a preset probability threshold, determine the TCR weight of each TCR record based on the clonal abundance information, and calculate the strong binding score, breadth score, and VJ diversity score for each candidate peptide; calculate the population TCR identification comprehensive score for each candidate peptide based on the strong binding score, the breadth score, and the VJ diversity score.

[0018] Step 6: Sort the candidate peptide set according to the population TCR identification comprehensive score and output the candidate peptide ranking results.

[0019] In the above scheme, the immune sequence training data in step 1 also includes HLA pseudo-sequence and HLA binding groove related residue sequences;

[0020] The domain-adaptive pre-training in step 2 includes:

[0021] Perform at least one of the following tasks on the HLA encoder: masking language modeling task, binding slot residue enhancement masking task, HLA allele family prediction task, and complete sequence vs. pseudo-sequence consistency comparison learning task.

[0022] Perform at least one of the following on the peptide encoder: short peptide mask language modeling, continuous segment mask reconstruction, peptide length interval prediction, and peptide source type prediction;

[0023] The TCR encoder performs at least one of the following: TCR mask language modeling, V gene fragment prediction, J gene fragment prediction, paired αβ chain contrastive learning, and tissue or disease origin-assisted prediction.

[0024] In the above scheme, the two-stage supervised training in step 2 includes:

[0025] In the first stage, input peptide-HLA pairing data, freeze the TCR encoder and its adapter layer and peptide-TCR cross-attention module, and update the peptide adapter layer, HLA adapter layer, peptide-HLA cross-attention module and peptide-HLA classification output module.

[0026] In the second stage, peptide-TCR pairing data is input, the HLA encoder and its adapter layer, peptide-HLA cross-attention module and peptide-HLA classification output module are frozen, and the peptide adapter layer, TCR adapter layer, peptide-TCR cross-attention module and peptide-TCR classification output module are updated; wherein, the peptide adapter layer is updated in both the first and second stages.

[0027] In the above scheme, the candidate peptide set in step 3 includes at least one of the following: mutant neoantigen candidate peptides generated from mutation sites of hepatocellular carcinoma samples, antigenic peptides derived from tumor-associated antigens or viral proteins, or hepatocellular carcinoma-associated HLA-presenting peptides obtained by mass spectrometry detection.

[0028] The construction of the binding probability matrix in step 4 includes: assigning rows of the binding probability matrix to candidate peptides, columns to TCRCDR3, and matrix elements to the binding probabilities between the corresponding candidate peptides and the corresponding TCRCDR3.

[0029] In the above scheme, step 5, filtering binding probabilities below a preset probability threshold in the binding probability matrix, and determining the TCR weights of each TCR record based on the clonal abundance information, includes:

[0030] Definition of the first The first TCR record and the first The binding probability between the candidate peptides is The preset probability threshold is The effective binding probability after filtering is... satisfy:

[0031]

[0032] When the TCR record includes clone abundance (cloneFraction), cloneFraction or its normalized result is used as the TCR weight of the corresponding TCR record. When the TCR records include clone abundance (cloneCount), the TCR weights are determined according to the proportion of each TCR record's cloneCount to the total cloneCount. .

[0033] In the above scheme, step 5, which calculates the strong binding score and breadth score for each candidate peptide, includes:

[0034] Regarding the first The strong binding score of the candidate peptides and breadth rating Calculate according to the following formulas respectively:

[0035]

[0036]

[0037] in, Indicates the first The TCR weight of each TCR record It is a strong binding index.

[0038] In the above scheme, step 5, calculating the VJ diversity score, and step 6, identifying the population TCR comprehensive score, include:

[0039] TCR records were divided into multiple VJ groups based on the combination of V and J gene fragments, and the number of fragments was calculated. The VJ group paired with the first Contribution ratio of each candidate peptide :

[0040]

[0041] Calculate the first VJ diversity entropy of candidate peptides :

[0042]

[0043] According to the first Number of VJ groups with non-zero contributions from candidate peptides ,right Normalization yields ,when When greater than 1:

[0044]

[0045] When Kⱼ equals 1:

[0046]

[0047] The population TCR identification comprehensive score Calculate according to the following formula:

[0048]

[0049] in, Indicates the first A VJ group, For VJ diversity weights, This is the breadth weight.

[0050] The domain-adaptive pre-training and two-stage supervised training in step 2 are optimized using a multi-task joint loss function, which includes:

[0051] The masking language modeling loss, binding slot residue enhancement masking loss, and complete sequence vs. pseudo-sequence consistency comparison learning loss are used for the HLA encoder.

[0052] For the peptide encoder, there are continuous fragment mask reconstruction loss and source type prediction loss; for the TCR encoder, there are V gene fragment classification loss, J gene fragment classification loss and paired αβ chain contrastive learning loss.

[0053] And the peptide-HLA binary classification loss, peptide-TCR binary classification loss, and virtual adversarial training loss in two-stage supervised training;

[0054] The multi-task joint loss function sums the above losses using preset weight parameters.

[0055] Regarding the various losses of the HLA encoder:

[0056]

[0057]

[0058]

[0059]

[0060] in, Represents HLA sequence, Indicates the first One real residue, For the set of occluded locations, This represents the input after masking. Represents the set of binding groove residues. For residue weights, For a set of categories, and These are the true label and the predicted probability, respectively. and These are represented as complete sequences and pseudo-sequences, respectively. For the candidate set within the batch, For similarity function, Temperature coefficient;

[0061] Regarding the various losses of the peptide encoder:

[0062]

[0063]

[0064]

[0065]

[0066] Indicates peptide sequence, Given a set of random occlusion locations, For a set of continuous occlusion segments, It is a continuous segment. For the segment length, and These are respectively the length range category and the source type category;

[0067] Regarding the various losses of the TCR encoder:

[0068]

[0069]

[0070]

[0071]

[0072]

[0073] Indicates TCR sequence, For the set of occlusion locations, and These are sets of V gene and J gene fragment categories, respectively. and Let them represent the α chain and the β chain, respectively. Category of origin for organization or disease;

[0074] Losses during the two-phase supervised training:

[0075]

[0076] Among them, virtual adversarial perturbation The calculation formula is:

[0077]

[0078]

[0079] and These are peptide-HLA and peptide-TCR training sets, respectively. and For sample weights, For binary categories, To predict probabilities, The total number of samples, Let KL divergence be the KL divergence. To predict the distribution for the model, For continuous embedding, The disturbance radius is... For gradient, For the anchor sample set, For the positive sample set, For the candidate sample set, To supervise and compare the learning temperature coefficient.

[0080] The present invention also provides a population-wide TCR broad-recognition candidate peptide screening system for hepatocellular carcinoma that implements the method, comprising:

[0081] The training data acquisition module is used to acquire immune sequence training data;

[0082] The model building and training module is used to build a targeted immune binding prediction model for hepatocellular carcinoma and to perform domain-adaptive pre-training and two-stage supervised training on the encoder.

[0083] The candidate data acquisition and combination module is used to acquire a set of candidate peptides and a TCRCDR3 library related to hepatocellular carcinoma, and to construct peptide-TCR combinations.

[0084] The probability prediction and matrix construction module is used to output the binding probability and construct the binding probability matrix; the scoring and ranking module is used to filter the binding probability, determine the TCR weight, calculate the strong binding score, breadth score, VJ diversity score and population TCR identification comprehensive score, and output the candidate peptide ranking results.

[0085] The present invention also provides an electronic device, including a processor, a memory, and a communication interface, wherein the memory stores a computer program, and the computer program, when executed by the processor, implements the method described herein.

[0086] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described thereon.

[0087] Because the present invention employs the above-mentioned technical means, it has the following beneficial effects:

[0088] 1. This invention performs domain-adaptive pre-training on the HLA encoder, peptide encoder, and TCR encoder in step S103, and retains the token-level hidden states of HLA, peptide, and TCR in step S104 for computation by the cross-attention module. This solves the technical problem that existing candidate peptide screening methods cannot simultaneously characterize the peptide-HLA presentation relationship and peptide-TCR recognition relationship at the residue level, thus improving the ability of the immune binding prediction model to characterize local interaction features in the hepatocellular carcinoma scenario.

[0089] 2. This invention obtains a candidate peptide set and an HCC-related TCRCDR3 library through step S107, and performs full combination prediction and peptide-level scoring on the candidate peptides with a large number of HCC-related TCRCDR3s through steps S108 to S110. This solves the technical problem that existing methods are prone to screening only around a single TCR, a single patient's CDR3, or a small number of paired samples, which makes it difficult to reflect the identifiability of the population TCR. It achieves the effect of converting the pair-level binding probability into a peptide-level population identification score and improving the population applicability of candidate peptides.

[0090] 3. By retaining cloneCount or cloneFraction in step S107 and calculating TCR weights based on TCR clonal abundance in step S110, this invention solves the technical problem that TCR clones with different amplification degrees are simply treated equally in the scoring and cannot reflect the antigen-driven clonal amplification signal. This achieves the effect of enabling high-frequency HCC-related TCR clones to make a more reasonable contribution to the candidate peptide scoring and reducing the influence of low abundance noise.

[0091] 4. This invention filters the binding probability by low probability in step S110 and combines strong binding score and breadth score, which solves the technical problems that using only the highest binding probability is easily affected by a single abnormally high allocation pair and using only the coverage quantity is easily affected by a large amount of low probability noise. It achieves the effect that the candidate peptide ranking reflects both binding strength and the TCR coverage range.

[0092] 5. This invention obtains TCRV gene fragment and J gene fragment information in step S107, and calculates VJ diversity score based on V / J gene combination entropy in step S110. This solves the technical problem that candidate peptide scores may be dominated by background TCRs of a single VJ clonal type, making it difficult to distinguish between broad recognition and local clonal amplification. It achieves the effect of rewarding candidate peptides that can be recognized by background TCRs of different VJ clonal types and improving the screening ability of candidate vaccine peptides with broad recognition potential.

[0093] 6. This invention, by dividing data into patient-level, TCR-level, peptide-level, HLA allele-level, or data source-level categories and combining them with candidate peptide-TCR full combination prediction, TCR clonal abundance weighting, strong binding score, breadth score, and VJ diversity score, produces a consistent constraint synergistic effect from model training evaluation to peptide-level ranking output. It solves the technical problems that easily occur when using paired probability prediction, single clonal abundance indicators, or single diversity indicators alone, such as data leakage, single patient or single clonal type bias, and unstable candidate peptide ranking. It achieves the effect that the final candidate peptide ranking simultaneously reflects HLA presentation capability, population TCR coverage, clonal amplification signal, and VJ clonal type diversity. Attached Figure Description

[0094] Figure 1 This is a simplified flowchart of the present invention;

[0095] Figure 2 This is a schematic diagram of the overall process of the candidate peptide screening method provided in the embodiments of the present invention.

[0096] Figure 3 This is a schematic diagram of the structure of the hepatocellular carcinoma targeted immune binding prediction model provided in an embodiment of the present invention.

[0097] Figure 4 This is a schematic diagram of the probability matrix and scoring process for the binding of candidate peptides to HCC-related TCRCDR3, provided in an embodiment of the present invention.

[0098] Figure 5 This is a schematic diagram of the candidate peptide screening system module structure provided in an embodiment of the present invention.

[0099] Figure 6 This is a schematic diagram illustrating the data partitioning, model training, and evaluation process provided in an embodiment of the present invention.

[0100] Explanation of reference numerals in the attached figures:

[0101] 10 Candidate peptide screening system; 11 Training data acquisition module; 12 Encoder pre-training module; 13 Model building module; 14 Supervised training module; 15 Candidate data acquisition module; 16 Combinatorial construction module; 17 Binding probability prediction module; 18 Scoring module; 19 Ranking output module; 20 Data storage module. Detailed Implementation

[0102] The embodiments of the present invention will be described in detail below. Although the present invention will be described and illustrated in conjunction with some specific embodiments, it should be noted that the present invention is not limited to these embodiments. On the contrary, any modifications or equivalent substitutions made to the present invention should be covered within the scope of the claims of the present invention.

[0103] Furthermore, to better illustrate the present invention, numerous specific details are set forth in the following detailed embodiments. Those skilled in the art will understand that the present invention can be practiced without these specific details.

[0104] This invention aims to solve at least the following technical problems:

[0105] 1. How to construct an immune binding prediction model for hepatocellular carcinoma that utilizes residue-level features of HLA sequences, peptide sequences, and TCRCDR3 sequences simultaneously.

[0106] 2. How to assess the overall identifiability of candidate peptides in a population-level HCCTCR library without relying on specific CDR3 sequences of individual patients as screening targets.

[0107] 3. How to convert peptide-TCR pairing-level prediction results into peptide-level ranking results so that the ranking can simultaneously reflect strong binding ability, TCR coverage breadth, clonal abundance weight, and VJ clonal diversity.

[0108] 4. How to reduce the bias of high probability of a single TCR, high abundance of a single patient clone, or a single VJ combination on the screening results of candidate peptides.

[0109] To address the aforementioned technical problems, this invention provides a method for broadly identifying candidate peptides for population-based TCR in hepatocellular carcinoma. This method includes a model construction phase and a candidate peptide screening phase.

[0110] In the model building phase, HLA sequences, immune peptide sequences, TCRCDR3 sequences, and HCC-related antigen or TCR data were acquired; HLA encoders, peptide encoders, and TCR encoders were constructed respectively; domain-adaptive pre-training was performed on the three encoders; the three encoders were connected to a unified cross-attention structure through a projection adaptation layer; two-stage supervised training was performed using peptide-HLA data and peptide-TCR data; and then targeted fine-tuning was performed using HCC-specific strong supervision, weak supervision, or ranking data to obtain a hepatocellular carcinoma targeted immune binding prediction model.

[0111] In the candidate peptide screening stage, a set of candidate peptides and an HCC-related TCRCDR3 library are obtained; combinations of all candidate peptides with TCRCDR3 are constructed; the binding probability of each combination is predicted using a hepatocellular carcinoma-directed immune binding prediction model; a candidate peptide × TCRCDR3 binding probability matrix is ​​constructed; low-probability predictions are filtered out; TCR weights are calculated based on cloneFraction or cloneCount; for each candidate peptide, a strong binding score Fj, a breadth score Bj, a VJ diversity score Hj, and a comprehensive score Scorej are calculated; and candidate peptides are ranked from high to low according to Scorej.

[0112] Example 1:

[0113] This embodiment provides a method for screening candidate peptides for broad TCR recognition in hepatocellular carcinoma populations, including the following steps.

[0114] S101, Obtain immune sequence training data. Immune sequence training data includes HLA class I sequences, HLA pseudo-sequences, immune peptide sequences, HLA ligand data, TCRCDR3 sequences, TCRV / D / J gene fragment information, HCC immune peptide data, HCC-related TCR data, tumor-associated antigen peptide data, and virus-associated antigen peptide data.

[0115] S102, construct the HLA encoder, peptide encoder, and TCR encoder respectively. The HLA encoder and peptide encoder can be initialized using a protein language model, while the TCR encoder can be initialized using a TCR-specific language model. In a specific embodiment, the HLA encoder and peptide encoder can be initialized using ESM-C, ESM-2, ProtT5, or other protein language models, and the TCR encoder can be initialized using TCR-BERT, SCEPTR, tcrLM, or other TCR language models.

[0116] S103, perform domain-adaptive pre-training on the three encoders respectively to obtain the HLA-DAPT encoder, Peptide-DAPT encoder and TCR-DAPT encoder.

[0117] S104, the three encoders are connected to a unified immune binding prediction model through a projection adaptation layer. The immune binding prediction model includes a peptide-HLA cross-attention module and a peptide-TCR cross-attention module. The HLA token-level features output by the HLA encoder and the peptide token-level features output by the peptide encoder enter the peptide-HLA cross-attention module; the TCR token-level features output by the TCR encoder and the peptide token-level features output by the peptide encoder enter the peptide-TCR cross-attention module.

[0118] S105 uses peptide-HLA pairing data and peptide-TCR pairing data for two-stage supervised training. The first stage is peptide-HLA training, and the second stage is peptide-TCR training. In both stages, the peptide adaptor layer is updated so that the peptide representation can simultaneously absorb HLA presentation information and TCR recognition information.

[0119] S106 utilizes HCC-related data for targeted fine-tuning. HCC-related data includes HCC / HBV immunopeptidome data, HCC-related vaccine response TCR data, HCC tumor invasion TCR data, HCC mutant neoantigen candidate peptides, HCC tumor-associated antigen peptides, and peptide-TCR samples with HLA restriction information. Ordinary tumor invasion TCRs are not directly used as true positive markers for antigen binding, but can be used as TCR characterization adaptation or weak supervision data.

[0120] S107, Obtain the candidate peptide set and the HCC-related TCRCDR3 library. The candidate peptide set can be derived from mutated neoantigens, tumor-associated antigens, virus-associated antigens, or HLA-presenting peptides detected by mass spectrometry. The HCC-related TCRCDR3 library consists of TCR records from multiple HCC samples, public databases, or experimental data. Each record includes at least one of the following: CDR3 amino acid sequence, V gene fragment, D gene fragment, J gene fragment, cloneCount, and cloneFraction.

[0121] S108 constructs combinations of all candidate peptides with all HCC-related TCRCDR3, inputs them into the hepatocellular carcinoma-directed immune binding prediction model, and outputs the binding probability pᵢⱼ.

[0122] S109, construct the candidate peptide × TCRCDR3 binding probability matrix. Each row of the matrix corresponds to a candidate peptide, each column corresponds to a TCRCDR3, and the elements in the matrix are the binding probabilities of the corresponding combinations.

[0123] S110 performs low-probability filtering, TCR weight calculation, strong binding score calculation, breadth score calculation, VJ diversity score calculation, and comprehensive score calculation on a complete row of probability vectors for each candidate peptide.

[0124] S111: Candidate peptides are sorted in descending order according to their comprehensive scores, and a ranking table of candidate peptides is output. This ranking table can be used for subsequent in vitro experimental validation, immunogenicity validation, or as a reference for candidate vaccine peptide design, but it is not directly used as a clinical diagnosis or treatment plan.

[0125] Example 2: Domain Adaptive Pre-training of HLA Encoders

[0126] In one embodiment, the HLA encoder is initialized from a protein language model. The input can be a complete HLA class heavy chain sequence, an HLA pseudo-sequence, or peptide-binding groove-related residues extracted after encoding a complete HLA sequence.

[0127] HLA preprocessing includes:

[0128] 1. Standardize the HLA allele naming format, for example, HLA-A*02:01.

[0129] 2. Remove records containing non-standard amino acids or incomplete sequences.

[0130] 3. Classify according to HLA-A, HLA-B, HLA-C and allele families.

[0131] 4. Construct a mapping relationship between complete HLA sequences and HLA pseudo-sequences.

[0132] HLA domain adaptive pre-training tasks include:

[0133] 1. Masking language modeling task: Randomly mask HLA amino acid residues and predict the masked residues.

[0134] 2. Enhanced masking task based on dip residues: Set mask probabilities higher than normal positions for dip residues related to the dip-binding groove.

[0135] 3. Allele family prediction task: Predict HLA-A, HLA-B, HLA-C or more fine-grained allele groups based on HLA representation.

[0136] 4. Learning task comparing complete sequence and pseudo-sequence consistency: The complete sequence representation and pseudo-sequence representation of the same allele are brought closer together, while the representations of different alleles are pushed apart.

[0137] The HLA-DAPT loss function can be expressed as:

[0138]

[0139] in, For masked language modeling loss, To enhance mask loss by combining slots, Predicting loss for allele families, The loss is learned by comparing the consistency between the complete sequence and the pseudo-sequence, with λ1, λ2 and λ3 being the weight parameters.

[0140] In a preferred embodiment, the weight parameters of the HLA-DAPT loss function are set as follows: =0.5, =0.2, =0.2. This setting maintains a high weight for the binding groove residue enhancement masking task, while using allele family prediction and complete sequence / pseudo-sequence consistency constraints as auxiliary regularization terms.

[0141] The specific loss items mentioned above can be expressed as follows:

[0142]

[0143]

[0144]

[0145]

[0146] in, This indicates an HLA class I heavy chain sequence, an HLA pseudo-sequence, or a binding groove residue sequence. This represents the actual HLA amino acid residue at position t. The set of randomly masked HLA residue locations. This indicates the masked HLA input; This refers to the set of peptide-binding groove residues located in the HLA class I α1 / α2 domain that are in contact with peptide anchor sites or the peptide backbone. To determine the weight of residues in the binding groove, residues near the anchor point can be made higher than ordinary residues; This represents a set of allelegroup categories, such as HLA-A, HLA-B, HLA-C, or more granular categories. and These are the true allele category labels and predicted probabilities, respectively. and , respectively, represent the complete sequence representation and pseudo-sequence representation of the same HLA allele, B is the set of candidate pseudo-sequences within the training batch, sim is the cosine similarity or dot product similarity, and τ is the contrastive learning temperature coefficient.

[0147] Example 3: Domain-Adaptive Pre-training of Peptide Encoders

[0148] In one embodiment, the peptide encoder is initialized from a protein language model and pre-trained on domain-adaptive peptides, HLAligand, viral peptides, tumor-associated antigen peptides, and HCC / HBV-related peptides.

[0149] Peptide pretreatment includes:

[0150] 1. Merge immune epitope database, HLAligand database, immune peptide database and HCC / HBV-related immune peptide data.

[0151] 2. Standardize peptide sequences to uppercase amino acid sequences.

[0152] 3. Remove peptides containing non-standard amino acids.

[0153] 4. Retain the main training peptide segment with a length of 8 to 15 amino acids, and can retain the extended training peptide segment with a length of 7 to 30 amino acids.

[0154] 5. Label the peptide source type, including HLA ligand, Tcellepitope, viral peptide, tumor-associated peptide, neoantigen candidate, or HCC / HBV-related peptide.

[0155] 6. Remove duplicates of the same peptide while retaining metadata from multiple sources.

[0156] Peptide domain adaptive pre-training tasks include:

[0157] 1. Short peptide mask language modeling task.

[0158] 2. Spanmasking task involving 2 to 4 consecutive amino acid residues.

[0159] 3. Peptide length range prediction task, such as predicting length ranges of 8, 9, 10, 11, 12 to 15 amino acids.

[0160] 4. Peptide origin type prediction task.

[0161] The Peptide-DAPT loss function can be expressed as:

[0162]

[0163] in, For the loss of reconstructing continuous segment masks, For predicting loss over a length interval, Predict loss based on source type.

[0164] In a preferred embodiment, the weight parameters of the Peptide-DAPT loss function are set as follows: =0.5, =0.1, =0.2. This setting prioritizes the learning of local motifs of short peptides in continuous fragment mask reconstruction, while leveraging length range and source type labels with lower weights to enhance the immunological priors of peptide characterization.

[0165] The specific loss terms of the Peptide-DAPT mentioned above can be expressed as follows:

[0166]

[0167]

[0168]

[0169]

[0170] in, This indicates an immune peptide, HLA ligand, viral peptide, tumor-associated antigen peptide, neoantigen candidate peptide, or HCC / HBV-related peptide sequence. This represents the actual amino acid residue at position t of the peptide. Given a set of random occlusion locations, Let be a set of continuous masked fragments, where s is one of the continuous fragments, and |s| is the number of amino acid residues contained in that fragment. This indicates the input of the masked peptide; Indicates peptide length range categories, such as 8, 9, 10, 11, 12 to 15 amino acids or extended length ranges; The peptide origin type category is represented by y, which includes HLA ligand, Tcellepitope, viral peptide, tumor-associated peptide, neoantigen candidate, and HCC / HBV-related peptide; y and ŷ represent the true class label and the model predicted class probability, respectively.

[0171] Example 4: TCR Encoder Domain Adaptive Pre-training

[0172] In one embodiment, the TCR encoder is initialized using a TCR-specific language model. The input can be a CDR3β sequence or a paired CDR3α / β sequence. When using paired α / β chains, the following format can be used:

[0173] [CLS]CDR3α[SEP]CDR3β[SEP]

[0174] When only β chain data is available, the following format can be used:

[0175] [CLS]CDR3β[SEP]

[0176] TCR pretreatment includes:

[0177] 1. Standardize the CDR3 amino acid sequence format.

[0178] 2. Remove non-productive TCR sequences, abnormally long sequences, and sequences containing non-standard amino acids.

[0179] 3. Standardize the nomenclature of V gene fragments, D gene fragments, and J gene fragments.

[0180] 4. Retain metadata such as cloneCount, cloneFraction, sample origin, tissue origin, and disease origin.

[0181] Adaptive pre-training tasks in the TCR domain include TCR mask language modeling, V gene fragment prediction, J gene fragment prediction, paired αβ chain contrastive learning, and tissue or disease origin-assisted prediction.

[0182] The TCR-DAPT loss function can be expressed as:

[0183]

[0184] in, Modeling loss for CDR3 masked language, The cross-entropy loss for classifying V gene fragments. For the classification of J gene fragments, cross-entropy loss, The consistency comparison learning loss is used to pair the α-chain and β-chain representations, and L_tissue-disease is the auxiliary classification loss for tissue origin or disease origin. , , and These are the weight parameters.

[0185] In a preferred embodiment, the weight parameters of the TCR-DAPT loss function are set as follows: =0.3, =0.2, =0.2, =0.1. This setting uses V / J gene fragment information and αβ chain pairing consistency as a moderate-strength auxiliary constraint, and tissue or disease origin classification as a low-weight domain adaptation constraint.

[0186] The above TCR-related loss terms can be expressed as follows:

[0187]

[0188]

[0189]

[0190]

[0191]

[0192] in, This indicates a TCRCDR3β sequence or a paired CDR3α / β sequence. This represents the actual CDR3 amino acid residue at position t. The set of occluded CDR3 locations, marked with a tilde. This indicates the TCR input after masking; and These represent the V gene fragment category set and the J gene fragment category set after standardization according to IMGT or equivalent nomenclature, such as TRBV, TRBJ, TRAV, or TRAJ gene fragments; , , and These are the actual V / J gene tags and predicted probabilities, respectively. and The TCR α and β chains are represented by the same T cell clone, respectively. Represents other β chains within the batch; Indicates the type of tissue or disease source, such as tumor tissue, peripheral blood, liver tissue, HCC, HBV infection, or healthy control; τ is the contrast learning temperature coefficient, and B is the training batch.

[0193] Example 5: Structure of a Predictive Model for Targeted Immunization in Hepatocellular Carcinoma

[0194] The hepatocellular carcinoma targeted immune binding prediction model in this embodiment includes three coding branches and two cross-attention branches.

[0195] The HLA branch includes the HLA-DAPT encoder and the HLA adapter layer. The HLA-DAPT encoder takes as input the complete HLA sequence, HLA pseudo-sequence, or HLA binding groove residue sequence, and outputs the HLA token-level hidden state.

[0196] The peptide branch includes a Peptide-DAPT encoder and a peptide adaptation layer. The Peptide-DAPT encoder takes candidate peptide sequences as input and outputs peptide token-level hidden states.

[0197] The TCR branch includes the TCR-DAPT encoder and the TCR adaptation layer. The TCR-DAPT encoder takes a TCRCDR3β sequence or a CDR3α / β sequence as input and outputs a TCRtoken-level hidden state.

[0198] Each adapter layer may include:

[0199] 1. LayerNorm;

[0200] 2.Linear(d_encoder→d_model);

[0201] 3. GELU;

[0202] 4. Dropout;

[0203] 5.Linear(d_model→d_model);

[0204] 6. LayerNorm.

[0205] Where d_encoder is the hidden dimension of the corresponding pre-trained encoder, and d_model is the dimension of the unified cross-attention model. d_model can be set to 512, or it can be set to 128, 256, 768, or other dimensions depending on computing resources.

[0206] The peptide-HLA cross-attention module outputs pHLA binding scores based on peptide token features and HLA token features. The peptide-TCR cross-attention module outputs pTCR binding scores based on peptide token features and TCR token features. For triplet data containing HLA, peptide, and TCR, a combined triplet output header can also be set to output a triplet immunogenicity score.

[0207] Example 6: Supervised Training and HCC Oriented Fine-Tuning

[0208] In one embodiment, the HLA-DAPT encoder, Peptide-DAPT encoder, and TCR-DAPT encoder are frozen, and the HLA adapter layer, peptide adapter layer, TCR adapter layer, cross-attention module, and classification output module are trained first.

[0209] The two-phase training includes:

[0210] The first stage is pHLA training. Inputting peptides and HLA, updating the peptide adaptation layer, HLA adaptation layer, peptide-HLA cross-attention module, and pHLA classification output module, and freezing TCR-related modules.

[0211] The second stage is pTCR training. Input peptides and TCRCDR3, update the peptide adaptation layer, TCR adaptation layer, peptide-TCR cross-attention module, and pTCR classification output module, and freeze the HLA-related modules.

[0212] In both phases, the peptide adaptor layer is updated, enabling the peptide representation to simultaneously learn HLA-presentation-related features and TCR-recognition-related features.

[0213] The supervised training loss function can be expressed as:

[0214]

[0215] in, For peptide-HLA binary classification loss, For peptide-TCR binary classification loss, For virtual adversarial training loss, To supervise the contrastive learning loss, virtual adversarial perturbations are added to the projected contiguous embedding space, rather than to the amino acid sequence itself.

[0216] In a preferred embodiment, the weight parameters of the supervised training loss function are set as follows: =0.7, =0.1, =0.2. This setting prioritizes peptide-HLA presentation / binding prediction as the primary task, while assigning a higher but lower weight to the peptide-TCR recognition branch than to the pHLA branch. Virtual adversarial training and supervised contrastive learning are used to enhance the stability of the embedding space.

[0217] The loss terms of the above supervised training can be expressed as follows:

[0218]

[0219]

[0220]

[0221]

[0222]

[0223] in, This is a peptide-HLA pairing training set, where sample n contains peptide sequences. and HLA alleles ; It is a binary tag for whether a peptide is presented or bound to this HLA. Positive samples can come from HLAligand mass spectrometry, binding experiments or validated immunoepitaxes, while negative samples can come from decoy peptides, unobserved pairings or experimental negative pairings. This represents the binding or presentation probability of the peptide-HLA branch. Use sample confidence or class balance weights. This is a peptide-TCR training set, where sample n contains peptides. and TCRCDR3 or α / β chain sequence ; This indicates whether the TCR recognizes the corresponding peptide or pMHC complex. This represents the recognition probability output by the peptide-TCR branch. These are the sample weights. For continuous embedding after the projection adaptation layer, Let ε be the virtual adversarial perturbation, KL be the perturbation radius, and I be the Kullback-Leibler divergence between the predicted distributions. I represents the supervised contrastive learning anchor sample set, P(i) is the set of positive samples that have the same antigenic peptide, the same HLA restriction, or the same TCR recognition tag as anchor sample i, and A(i) is the set of candidate samples other than i in the same batch. To supervise and compare the learning temperature coefficient.

[0224] During the HCC targeted fine-tuning phase, HCCpHLA samples, HCCpTCR samples, and HCCHLA-peptide-TCR triplet samples can be used. The weight of strongly supervised samples can be set to 1.0, the weight of weakly supervised samples can be set to 0.2 to 0.5, and the weight of pseudo-labeled samples can be set to 0.05 to 0.2.

[0225] The HCC fine-tuning loss function can be expressed as:

[0226]

[0227] in, The criteria used to improve the scores of vaccine reactive peptides compared to non-reactive candidate peptides, HLA-presenting peptides detected by mass spectrometry compared to undetected decoy peptides, and amplified TCR pairings compared to non-amplified mismatched pairs.

[0228] In a preferred embodiment, the weight parameters of the HCC fine-tuning loss function are set as follows: =0.8, =0.6, =0.3, =0.05. This setting improves the impact of HCC-related TCR identification and HLA-peptide-TCR triplet samples on the final screening ranking, making the ranking loss directly constrain the priority of candidate peptides, and using virtual adversarial training as a low-weight stability regularization term.

[0229] The loss terms of the above HCC fine-tuning can be expressed as follows:

[0230]

[0231]

[0232]

[0233]

[0234] in, This indicates a collection of HCC-derived peptides and HLA samples. The peptides may be derived from AFP, GPC3, TERT, the MAGE family, HBV / HCV proteins, DNAJB1-PRKACA fusion protein, neoantigens recognized by WES / WGS, or HCC immunopeptide mass spectrometry peptides. This indicates whether the peptide is presented by the corresponding HLA allele or is predicted to be presented with high confidence. This represents the pHLA probability in the HCC scenario. This represents a collection of HCC-related peptide-TCR samples, from which TCRs may be derived from tumor-infiltrating T cells, peripheral blood expanded clones, publicly available HCC-related TCR libraries, or vaccine-responsive TCRs. This indicates whether the corresponding TCR recognizes the peptide or pMHC. Let pTCR be the probability. This represents the HLA-peptide-TCR triad set. This indicates whether the peptide can be recognized by the TCR under a specific HLA background. The weights are determined by the evidence weights of samples. Samples from strongly supervised experiments are given higher weights, while samples from weakly supervised, amplified clones, or pseudo-labeled experiments are given lower weights. For the ranking constraint set, u represents the positive example or high-confidence sample that should be ranked first, and v represents the decoy, experimental negative, undetected peptide, or mismatched TCR sample. and The overall reactivity score output by the model; The definition of the supervised training phase is adopted to improve the embedding stability within the HCC sample domain.

[0235] After HCC-oriented fine-tuning stabilizes, the top layers of the HLA encoder, peptide encoder, and TCR encoder can be unfrozen, or low-rank adaptation parameters can be added to the top layers for fine-tuning with a small learning rate. The bottom layers of the encoder remain frozen to reduce catastrophic forgetting and training resource overhead.

[0236] Example 7: Construction of Candidate Peptide Sets

[0237] The candidate peptide set can be constructed in the following ways.

[0238] One approach involves constructing candidate peptides for mutated neoantigens. MissenseSNV, in-frameindel, and frameshiftindel are extracted from WES, WGS, MAF, or VCF data of HCC samples. Candidate peptides of 8 to 11 amino acids or 8 to 15 amino acids in length are generated around the mutation site. RNA-seq expression levels are used for filtering, such as TPM greater than 1 or TPM greater than 5. Candidate peptide-HLA pairs are constructed based on HLA-typing results. Presentation capability is then scored using a hepatocellular carcinoma-directed immune binding prediction model or other pHLA prediction modules.

[0239] Another approach involves constructing HCC tumor-associated antigen peptides. These peptides can be derived from AFP, GPC3, TERT, the MAGE family, CTAG1B, HBV protein, HCV protein, DNAJB1-PRKACA fusion protein, or other hepatocellular carcinoma-associated antigens.

[0240] Another approach involves mass spectrometry detection of HLA ligand peptides for construction. HCC / HBV immunopeptidome, liver tissue HLA ligand data, or tumor immunopeptidome data can be used to identify HLA-presenting peptides as candidate peptides or positive sample sources.

[0241] The candidate peptide set can also be a combined result from the above sources. For the same peptide, metadata from multiple sources can be retained, including source protein, mutation information, HLA presentation evidence, expression level, mass spectrometry evidence, and literature evidence.

[0242] Example 8: Construction of HCC-related TCRCDR3 library

[0243] HCC-related TCRCDR3 libraries can be constructed from public databases, single-cell TCR sequencing, bulk TCR sequencing, vaccine response cohorts, tumor-infiltrating T cell data, or proprietary experimental data. Each TCR record may include the following fields:

[0244] 1. AASeq: CDR3 amino acid sequence;

[0245] 2. chain: α chain or β chain;

[0246] 3. Vregion: V gene fragment;

[0247] 4. D region: D gene fragment, mainly used for the β chain;

[0248] 5. J region: J gene fragment;

[0249] 6. cloneCount: The number of reads or occurrences of this TCR clone;

[0250] 7. cloneFraction: The frequency of this clone in the corresponding TCR library;

[0251] 8. patientId: Identifier of the patient or sample source;

[0252] 9. Tissue: Peripheral blood, tumor tissue, adjacent tissue, or other sources;

[0253] 10. Disease State: HCC, HBV-related HCC, HCV-related HCC, or other conditions.

[0254] When constructing the TCRCDR3 library, nonproductive TCR sequences, abnormally long CDR3s, low-quality sequencing records, and records containing non-standard amino acids can be removed. For duplicate CDR3s, their clone counts can be merged or records from different patient sources can be retained, depending on the application scenario.

[0255] To achieve population-level identification and assessment, the HCC-related TCRCDR3 database is preferably composed of multiple HCC patients or multiple HCC-related samples, rather than just a single patient sample. When patient origin identifiers are available, intra-patient normalization and inter-patient balanced weighting can be performed to avoid a single patient or a single highly amplified clone dominating the score.

[0256] Example 9: Constructing by combining probability matrices

[0257] Suppose the candidate peptide set includes M candidate peptides:

[0258] peptide1, peptide2, ..., peptide M .

[0259] Suppose that the HCC-related TCRCDR3 database contains N TCR records:

[0260] TCR1, TCR2, …, TCR N .

[0261] For each candidate peptide (peptidej) and each TCRi, construct a peptide-TCR combinatorial structure (peptide-TCR combinatorial structure). j TCR i Input the peptide-TCR branch of the hepatocellular carcinoma targeted immune binding prediction model, and output the binding probability p. ij , where p ij It is located between 0 and 1.

[0262] After all combination predictions are completed, the candidate peptide × TCRCDR3 binding probability matrix P is obtained:

[0263]

[0264] Where, p ij This represents the binding probability between the i-th TCR record and the j-th candidate peptide. For the j-th candidate peptide, the corresponding probability vector represents the recognition probability distribution of the candidate peptide in the HCC-related TCRCDR3 library.

[0265] In practical calculations, M×N combinations can be input into the model in batches to control GPU memory and computation time. Batch processing parameters may include the number of candidate peptides, the number of TCRs, the batch size, and the computing device.

[0266] Example 10: TCR Weighting and Low Probability Filtering

[0267] To reduce the impact of low-probability noise on the scoring, a threshold filter is applied to the combination probability pᵢⱼ. Let the preset probability threshold be... ,but:

[0268]

[0269] In one embodiment, Set to 0.1. In other embodiments, You can set it to 0.05, 0.2 or other values ​​based on the validation set results.

[0270] The weight wᵢ of each TCR record can be determined by clonal abundance. When cloneFraction has been normalized, we can let:

[0271]

[0272] When only cloneCount is available, we can set:

[0273]

[0274] When the TCR database comes from multiple patients and includes patientId, cloneFraction or cloneCount normalization can be performed within each patient first, and then each patient can be assigned the same or a preset patient weight. Let TCR... i If patient s belongs to a patient and the number of patients is S, then we can let:

[0275]

[0276] This approach makes the scoring closer to population-level identification assessment rather than clonal abundance assessment within a single patient.

[0277] Example 11: Strong Binding Score, Breadth Score, and VJ Diversity Score

[0278] For the j-th candidate peptide, strong binding score Defined as:

[0279]

[0280] Where α is the strong binding index. In one embodiment, α is set to 2. This index is used to increase the contribution of high-confidence binding probability to the score. For example, =0.81, =0.25, =0.04, therefore the strong binding score can suppress the artificially high scores caused by a large number of low probability predictions.

[0281] Breadth rating Defined as:

[0282]

[0283] when When normalized, It is located between 0 and 1. This represents the average degree to which candidate peptides are weighted and recognized by the HCC-related TCR library. The higher the value, the more likely the candidate peptide is to be strongly recognized by a few TCRs, and the more likely it is to have high overall coverage of a wider range of TCRs.

[0284] The VJ diversity score is used to assess whether the TCRs identifying a candidate peptide originate from multiple V / J combinations. First, TCR records are divided into VJ groups according to the combination of V and J regions:

[0285] .

[0286] For the j-th candidate peptide, the contribution ratio of the k-th VJ group Defined as:

[0287]

[0288] when When = 0, it indicates that the j-th candidate peptide did not obtain a valid TCR recognition contribution after low-probability filtering. In this case, , and Set all to 0.

[0289] Then calculate the Shannon entropy:

[0290]

[0291] Let Kⱼ be the number of VJ groups that make a non-zero contribution to the j-th candidate peptide, then the normalized VJ diversity score Defined as:

[0292] when >1 hour:

[0293]

[0294] when When =1:

[0295]

[0296] The higher the value, the more dispersed the TCR sources for identifying the candidate peptide are, rather than being concentrated in a single VJ combination, which is more in line with the goal of screening candidate peptides with broad TCR recognition potential.

[0297] Example 12: Population TCR Identification: Comprehensive Score and Ranking Output

[0298] For the j-th candidate peptide, the overall score is calculated. Defined as:

[0299]

[0300] in, To achieve a strong combination of scoring, Score for breadth. λ is the normalized VJ diversity score, where λ is the diversity weight and μ is the breadth weight.

[0301] In one embodiment, the parameter is set as follows:

[0302]

[0303] In other embodiments, α, λ, μ and It can be determined through validation sets, expert experience, wet experiment feedback, or Bayesian optimization.

[0304] The final output candidate peptide ranking table may include:

[0305]

[0306] The top-ranked candidate peptides are those that are more likely to be identified by HCC-related TCR libraries in model predictions with higher binding probability, broader TCR coverage, and greater VJ diversity.

[0307] Example 13: System Example

[0308] like Figure 5As shown, this embodiment provides a population TCR broad identification candidate peptide screening system 10 for hepatocellular carcinoma, including a training data acquisition module 11, an encoder pre-training module 12, a model building module 13, a supervised training module 14, a candidate data acquisition module 15, a combinatorial construction module 16, a binding probability prediction module 17, a scoring module 18, a sorting output module 19, and a data storage module 20.

[0309] The training data acquisition module 11 is used to acquire HLA sequences, immune peptide sequences, TCRCDR3 sequences, and HCC-related antigens or TCR data.

[0310] The encoder pre-training module 12 is used to perform domain-adaptive pre-training on the HLA encoder, peptide encoder and TCR encoder respectively.

[0311] The model building module 13 is used to connect the pre-trained encoder to the cross-attention module through the adaptation layer to form a hepatocellular carcinoma targeted immune binding prediction model.

[0312] The supervised training module 14 is used to perform at least one of peptide-HLA training, peptide-TCR training, HCC targeted fine-tuning, virtual adversarial training, supervised contrastive learning, and ranking training.

[0313] The candidate data acquisition module 15 is used to acquire the candidate peptide set and the HCC-related TCRCDR3 library.

[0314] Combinatorial construction module 16 is used to construct full or batch combinations between candidate peptides and TCRCDR3.

[0315] The probability prediction module 17 is used to call the trained model to output the combination probability of each combination and construct the combination probability matrix.

[0316] The scoring module 18 is used to perform low probability filtering, TCR weight calculation, strong binding score calculation, breadth score calculation, VJ diversity score calculation, and comprehensive score calculation.

[0317] The sorting output module 19 is used to generate a ranking table of candidate peptides and output it to a file, database, or visualization interface.

[0318] The data storage module 20 is used to store training data, model weights, candidate peptides, TCRCDR3 library, prediction probability matrix, scoring results, model version and parameter configuration.

[0319] Example 14: Electronic Devices and Storage Media

[0320] The present invention also provides an electronic device. This electronic device includes a processor, a memory, a communication interface, and a bus. The memory stores a computer program, and when the processor executes the computer program, it implements the population TCR broad-recognition candidate peptide screening method for hepatocellular carcinoma as described in any embodiment of this specification.

[0321] Electronic devices can be local workstations, servers, cloud computing nodes, GPU servers, edge computing devices, or computing clusters composed of multiple devices.

[0322] The present invention also provides a computer-readable storage medium. This computer-readable storage medium stores a computer program, which, when executed by a processor, implements the methods described in any embodiment of this specification. The computer-readable storage medium may include a magnetic disk, optical disk, solid-state drive, flash memory, read-only memory, random access memory, or other media capable of storing program code.

[0323] Example 15: Data Partitioning and Model Evaluation

[0324] To reduce the risk of data leakage, the following division method can be adopted in model training and evaluation:

[0325] 1. Peptide segmentation: The test peptide does not appear in the supervised training set.

[0326] 2. TCR-level partitioning: Test that TCRCDR3 does not appear in the supervised training set.

[0327] 3. Patient-level segmentation: Data from the same patient are not included in the training set and the test set at the same time.

[0328] 4. Source research level division: Use independent studies or independent cohorts as external validation sets.

[0329] 5. HLA allele classification: testing the generalization ability of low-frequency or unseen HLA alleles.

[0330] Evaluation metrics may include pHLAAUROC, pHLAAUPR, pTCRAUROC, pTCRAUPR, MCC, F1, PPV@K, peptide-unseenAUROC, TCR-unseenAUROC, patient-unseenAUROC, HLC mass spectrometry-based HLAligand recall rate, vaccine reactive peptide sequencing metrics, and amplified TCR enrichment metrics.

[0331] The above evaluation metrics are used to measure the performance of the model as a data processing and candidate peptide ranking tool, and are not used to make direct clinical diagnostic or treatment decisions.

Claims

1. A method for screening candidate peptides for broad TCR recognition in hepatocellular carcinoma populations, used for in vitro data processing for non-disease diagnosis and treatment purposes, characterized in that, include: Step 1: Obtain immune sequence training data, which includes HLA sequence data, immune peptide sequence data, T cell receptor CDR3 sequence data, and hepatocellular carcinoma-associated antigen or T cell receptor data. Step 2: Construct a hepatocellular carcinoma targeted immune binding prediction model based on the immune sequence training data. The hepatocellular carcinoma targeted immune binding prediction model includes an encoder, a cross-attention module, and a classification output module. The encoder includes an HLA encoder, a peptide encoder, and a TCR encoder. The cross-attention module includes a peptide-HLA cross-attention module and a peptide-TCR cross-attention module. The encoders were pre-trained for their respective domain adaptations using the immune sequence training data. The pre-trained encoder is connected to the cross-attention module through a projection adaptation layer, and two-stage supervised training is performed using experimentally validated peptide-HLA pairing data and peptide-TCR pairing data collected from four major public databases: IEDB, PIRD, McPAS-TCR, and VDJdb. This results in a hepatocellular carcinoma targeted immune binding prediction model that can output the probability of peptide binding to TCRCDR3. Step 3: Obtain the candidate peptide set and the hepatocellular carcinoma-related TCRCDR3 library. Each TCR record in the hepatocellular carcinoma-related TCRCDR3 library includes the CDR3 amino acid sequence, V gene fragment, J gene fragment, and clonal abundance information. Step 4: Construct peptide-TCR combinations for each candidate peptide in the candidate peptide set and each TCR record in the hepatocellular carcinoma-related TCRCDR3 library, and input the peptide-TCR combinations into the hepatocellular carcinoma-directed immune binding prediction model to obtain the binding probability between each candidate peptide and each TCR record, and construct a binding probability matrix based on the binding probability. Step 5: Filter the binding probabilities in the binding probability matrix that are below a preset probability threshold, determine the TCR weight of each TCR record based on the clonal abundance information, and calculate the strong binding score, breadth score and VJ diversity score for each candidate peptide. A comprehensive population TCR identification score for each candidate peptide is calculated based on the strong binding score, the breadth score, and the VJ diversity score. Step 6: Sort the candidate peptide set according to the population TCR identification comprehensive score and output the candidate peptide ranking results.

2. The method according to claim 1, characterized in that, The immune sequence training data in step 1 also includes HLA pseudo-sequences and HLA binding groove-related residue sequences; The respective domain adaptation pre-training in step 2 includes: Perform at least one of the following tasks on the HLA encoder: masking language modeling task, binding slot residue enhancement masking task, HLA allele family prediction task, and complete sequence vs. pseudo-sequence consistency comparison learning task. Perform at least one of the following on the peptide encoder: short peptide mask language modeling, continuous segment mask reconstruction, peptide length interval prediction, and peptide source type prediction; The TCR encoder performs at least one of the following: TCR mask language modeling, V gene fragment prediction, J gene fragment prediction, paired αβ chain contrastive learning, and tissue or disease origin-assisted prediction.

3. The method according to claim 1, characterized in that, The two-stage supervised training in step 2 includes: In the first stage, input peptide-HLA pairing data, freeze the TCR encoder and the adapter layer and peptide-TCR cross-attention module behind the TCR encoder, and update the peptide adapter layer, HLA adapter layer, peptide-HLA cross-attention module and peptide-HLA classification output module. In the second stage, peptide-TCR pairing data is input, the HLA encoder and the adapter layer behind the HLA encoder, the peptide-HLA cross-attention module and the peptide-HLA classification output module are frozen, and the peptide adapter layer, TCR adapter layer, peptide-TCR cross-attention module and peptide-TCR classification output module are updated; wherein, the peptide adapter layer is updated in both the first and second stages.

4. The method according to claim 1, characterized in that, The candidate peptide set in step 3 includes at least one of the following: mutant neoantigen candidate peptides generated from mutation sites in hepatocellular carcinoma samples, antigenic peptides derived from tumor-associated antigens or viral proteins, or hepatocellular carcinoma-associated HLA-presenting peptides obtained by mass spectrometry detection. The construction of the binding probability matrix in step 4 includes: assigning rows of the binding probability matrix to candidate peptides, columns to TCRCDR3, and matrix elements to the binding probabilities between the corresponding candidate peptides and the corresponding TCRCDR3.

5. The method according to claim 4, characterized in that, Step 5 involves filtering binding probabilities in the binding probability matrix that are below a preset probability threshold, and determining the TCR weights for each TCR record based on the clonal abundance information, including: Definition of the first The first TCR record and the first The binding probability between the candidate peptides is The preset probability threshold is The effective combination probability after filtering is... satisfy: When the TCR record includes a clonal abundance ratio, the clonal abundance ratio or its normalized result is used as the TCR weight of the corresponding TCR record. When the TCR records include clonal abundance counts, the TCR weights are determined according to the proportion of the clonal abundance count of each TCR record to the total clonal abundance count. .

6. The method according to claim 5, characterized in that, Step 5 involves calculating strong binding score and breadth score for each candidate peptide, including: Regarding the first The strong binding score of the candidate peptides and breadth rating Calculate according to the following formulas respectively: in, Indicates the first The TCR weight of each TCR record It is a strong binding index.

7. The method according to claim 6, characterized in that, The calculation of the VJ diversity score in step 5 and the comprehensive score for population TCR identification in step 6 include: TCR records were divided into multiple VJ groups based on the combination of V and J gene fragments, and the number of fragments was calculated. The VJ group paired with the first Contribution ratio of each candidate peptide : Calculate the first VJ diversity entropy of candidate peptides : According to the first Number of VJ groups with non-zero contributions from candidate peptides ,right Normalization yields ,when When greater than 1: When Kⱼ equals 1: The population TCR identification comprehensive score Calculate according to the following formula: in, Indicates the first A VJ group, For VJ diversity weights, This is the breadth weight.

8. The method according to claim 2, characterized in that, The domain-adaptive pre-training and two-stage supervised training in step 2 are optimized using a multi-task joint loss function, which includes: The masking language modeling loss, binding slot residue enhancement masking loss, and complete sequence vs. pseudo-sequence consistency comparison learning loss are used for the HLA encoder. For the peptide encoder, there are continuous fragment mask reconstruction loss and source type prediction loss; for the TCR encoder, there are V gene fragment classification loss, J gene fragment classification loss and paired αβ chain contrastive learning loss. And the peptide-HLA binary classification loss, peptide-TCR binary classification loss, and virtual adversarial training loss in two-stage supervised training; The multi-task joint loss function sums the above losses using preset weight parameters.

9. The method according to claim 8, characterized in that, Regarding the various losses of the HLA encoder: in, Represents HLA sequence, Indicates the first One real residue, For the set of occluded locations, This represents the input after masking. Represents the set of binding groove residues. For residue weights, For a set of categories, and These are the true label and the predicted probability, respectively. and These are represented as complete sequences and pseudo-sequences, respectively. For the candidate set within the batch, For similarity function, Temperature coefficient; Regarding the various losses of the peptide encoder: Indicates peptide sequence, Given a set of random occlusion locations, For a set of continuous occlusion segments, It is a continuous segment. For the segment length, and These are respectively the length range category and the source type category; Regarding the various losses of the TCR encoder: Indicates TCR sequence, For the set of occlusion locations, and These are sets of V gene and J gene fragment categories, respectively. and Let them represent the α chain and the β chain, respectively. Category of origin for organization or disease; Losses during the two-phase supervised training: Among them, virtual adversarial perturbation The calculation formula is: and These are peptide-HLA and peptide-TCR training sets, respectively. and For sample weights, For binary categories, To predict probabilities, The total number of samples, Let KL divergence be the KL divergence. To predict the distribution for the model, For continuous embedding, The disturbance radius is... For gradient, For the anchor sample set, For the positive sample set, For the candidate sample set, To supervise and compare the learning temperature coefficient.

10. A population TCR broad-recognition candidate peptide screening system for hepatocellular carcinoma that implements the method as described in any one of claims 1-9, characterized in that, include: The training data acquisition module is used to acquire immune sequence training data; The encoder pre-training module is used to perform domain-adaptive pre-training on the HLA encoder, peptide encoder, and TCR encoder, respectively. The model building module is used to connect the pre-trained encoder to the cross-attention module through the adaptation layer to build a hepatocellular carcinoma targeted immune binding prediction model. The supervised training module is used for two-stage supervised training using paired data. The candidate data acquisition module is used to acquire a set of candidate peptides and a TCRCDR3 library related to hepatocellular carcinoma; the combination construction module is used to construct peptide-TCR combinations. Combined with the probability prediction module, it is used to output the combination probability and construct the combination probability matrix; The scoring module is used to filter the binding probability, determine the TCR weight, and calculate the strong binding score, breadth score, VJ diversity score, and overall TCR identification score for the population. The sorting output module is used to output the ranking results of candidate peptides.

Images