Noninvasive multi-cancer-species AI screening method based on tumor reactive TCR analysis

Abstract

The invention relates to a non-invasive multi-cancer AI screening method based on tumor reactive TCR analysis. The method comprises the following steps: extracting a peripheral blood sample of a subject, and sequencing on a machine to obtain sequencing data; pairing an alpha chain and a beta chain of a T cell receptor (TCR) in the sequencing data, and constructing a peripheral blood TCR library of the subject; extracting and splicing amino acid sequences of each pair of alpha chain and beta chain according to a preset sequence to obtain a structured full-chain TCR sequence; inputting the structured amino acid sequence into a trained model to obtain a tumor reactivity probability value corresponding to the TCR clone; calculating a sample-level tumor score based on the tumor reactivity probability value of each TCR clone; and according to the sample-level tumor score, generating a tumor risk screening result of the generic cancer species. The method has the technical effect that accurate, stable and non-invasive assessment of cancer risk can be realized only by using full-chain TCR sequence information with extremely low proportion and high noise in peripheral blood.

Description

Technical Field

[0001] This invention relates to the field of medical technology applications, and in particular to a non-invasive AI screening method for multiple cancers based on tumor reactive TCR analysis. Background Technology

[0002] In existing technologies, peripheral blood T cell receptor (TCR) sequencing data is typically used directly to extract the TCR β chain CDR3 sequence. Feature vectors are constructed by generating k-mer word frequencies and combining them with manual features such as amino acid physicochemical properties. These vectors are then input into a deep learning model to directly classify samples as "cancer / health" and identify cancer types.

[0003] However, the limitations of this technology are as follows: First, its modeling is directly based on statistical differences in the overall TCR library, failing to reveal the essence of cancer screening through TCR—namely, the enrichment of tumor-reactive TCRs. Second, it only utilizes the CDR3 region fragment of the TCR β chain, discarding information from the paired α chain and key structural information such as the V and J regions that constitute the complete variable region of the TCR. This limits the model's ability to fully characterize the TCR recognition function, affecting its expressive power and generalization performance across different cancer types. Furthermore, this method heavily relies on domain knowledge for manual feature engineering and employs a relatively general neural network architecture, failing to fully utilize the deep sequence pattern learning capabilities of more advanced large-scale models pre-trained on massive biological sequences. When faced with weak tumor signals and complex background noise in peripheral blood, the model's robustness and accuracy suffer.

[0004] Currently, no effective solutions have been proposed to address the shortcomings in sensitivity, specificity, and generalization ability across multiple cancer types in the cancer screening phase of related technologies. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing a non-invasive multi-cancer AI screening method based on tumor reactive TCR analysis, in order to solve the technical problems of insufficient sensitivity, specificity, and multi-cancer generalization ability in the cancer screening stage.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: This invention provides a non-invasive multi-cancer AI screening method based on tumor reactivity TCR analysis, comprising: extracting peripheral blood from a subject to obtain sequencing data; pairing the α and β chains of the T cell receptor in the sequencing data to generate a peripheral blood TCR library; extracting the amino acid sequences of the α and β chains from the peripheral blood TCR library according to a preset order, and splicing the obtained amino acid sequences of the α and β chains to obtain a structured full-chain amino acid sequence; inputting the structured full-chain amino acid sequence into a pre-trained model to obtain the tumor reactivity probability value of the corresponding TCR clone; calculating a sample-level tumor score based on the tumor reactivity probability value; and generating pan-cancer tumor risk screening results based on the sample-level tumor score.

[0007] Optionally, the peripheral blood obtained from the subjects is extracted to obtain sequencing data, including: separating peripheral blood mononuclear cells from the peripheral blood obtained from the subjects through a preset procedure; extracting RNA or DNA from the peripheral blood mononuclear cells to construct a TCR sequencing library; and obtaining sequencing data from the TCR sequencing library through a sequencing platform.

[0008] Optionally, the peripheral blood TCR library of the subject is generated by pairing the α and β chains of the T cell receptor in the sequencing data. This includes: aligning and assembling the T cell receptor in the sequencing data to obtain the variable region sequences of the α and β chains of the TCR; pairing the variable region sequences of the α and β chains of the TCR; removing clones with low quality, low read counts, and incomplete V or J annotations; and counting the clonal frequency of paired clones for each α chain and each paired clone for each β chain to construct the peripheral blood TCR library of the subject.

[0009] Further, optionally, the amino acid sequences of the α chain and β chain are extracted from the peripheral blood TCR library according to a preset order, and the obtained amino acid sequences of the α chain and β chain are spliced ​​together to obtain a structured full-chain amino acid sequence, including: extracting the amino acid sequences of the α chain and β chain respectively in a preset order including first extracting the V region, then extracting the CDR3 region, and finally extracting the J region, to obtain the amino acid sequences of the α chain and β chain; connecting the amino acid sequences of the α chain and β chain with connectors to generate a single structured amino acid sequence; and identifying the single structured amino acid sequence as the structured full-chain amino acid sequence.

[0010] Optionally, the training process of the pre-trained model includes: collecting paired TCR sequences from multiple cancer tumor samples that have been functionally validated as tumor reactivity, as positive samples; collecting paired TCR sequences from peripheral blood of healthy individuals, deleting sequences that are repetitive with or highly similar to the positive samples, as high-confidence negative samples; balancing positive and negative samples according to a preset ratio to determine the TCR sequences; splicing the α and β chains in the TCR sequences into a structured full-chain amino acid sequence according to a preset order; encoding the full-chain amino acid sequence using a word segmenter, truncating or padding to the maximum length; performing end-to-end adjustments using cross-entropy loss and Adam optimization, extracting hidden layer representations for t-SNE visualization and comprehensive gradient interpretation, verifying that the model can identify key amino acid residues and domains related to tumor reactivity, and obtaining a pre-trained model that has converged to training.

[0011] Further, optionally, calculating the sample-level tumor score based on the tumor reactivity probability value includes: sorting the tumor reactivity probability values, selecting the Top N clones, where N is a natural number; multiplying the frequency of each clone by the tumor reactivity probability value and summing the results to obtain the tumor reactivity intensity of the sample; and determining the tumor reactivity intensity of the sample as the sample-level tumor score.

[0012] Optionally, the method further includes: selecting the Top N clones; calculating the tumor reactivity intensity of the sample according to a preset formula; and mapping the tumor reactivity intensity to a sample-level tumor score in the range (0,1) through a preset transformation function.

[0013] Furthermore, optional, preset formulas include: Where S is the aggregation fraction, P i The tumor reactivity prediction score output for each clonus, F i The predicted frequency of tumor reactivity output for each clone; The preset transformation functions include: Where e is the natural constant; k is the coefficient, calculated based on a baseline established for healthy individuals; and Score is the score constructed in the healthy population.

[0014] Optionally, generating pan-cancer cancer risk screening results based on sample-level tumor scores includes: identifying individuals with a risk of developing cancer who have scores above a baseline threshold by setting a baseline threshold.

[0015] This invention employs the above technical solution, extracting sequencing data from peripheral blood of the subjects; pairing the α and β chains of the T cell receptor in the sequencing data to generate a peripheral blood TCR library; extracting the amino acid sequences of the α and β chains from the peripheral blood TCR library according to a preset order, and splicing the obtained α and β chain amino acid sequences to obtain a structured full-chain amino acid sequence; inputting the structured full-chain amino acid sequence into a pre-trained model to obtain the tumor reactivity probability value of the corresponding TCR clone; calculating a sample-level tumor score based on the tumor reactivity probability value; and generating pan-cancer tumor risk screening results based on the sample-level tumor score. Compared with existing technologies, this invention has the following technical effects: it achieves accurate, stable, and non-invasive assessment of cancer risk using only a very low proportion of TCR full-chain sequence information with a high-noise background in peripheral blood, and also achieves a lower false positive rate in multi-tumor screening. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a non-invasive multi-cancer AI screening method based on tumor reactive TCR analysis according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the gene structure of the α chain and β chain in a non-invasive multi-cancer AI screening method based on tumor reactive TCR analysis according to an embodiment of the present invention. Figure 3 This is a schematic diagram of a non-invasive multi-cancer AI screening method based on tumor reactive TCR analysis according to an embodiment of the present invention, after fine-tuning the ESM-2 protein language model; Figure 4 This is a flowchart illustrating another non-invasive multi-cancer AI screening method based on tumor reactive TCR analysis according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the algorithm usage interface in a non-invasive multi-cancer AI screening method based on tumor reactive TCR analysis according to an embodiment of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this application clearer, the application is described and illustrated below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. All other embodiments obtained by those skilled in the art based on the embodiments provided in this application without inventive effort are within the scope of protection of this application.

[0018] Obviously, the accompanying drawings described below are merely some examples or embodiments of this application. Those skilled in the art can apply this application to other similar scenarios based on these drawings without any inventive effort. Furthermore, it is understood that although the efforts made in this development process may be complex and lengthy, for those skilled in the art related to the content disclosed in this application, any changes to design, manufacturing, or production based on the technical content disclosed in this application are merely conventional technical means and should not be construed as insufficient disclosure of the content of this application.

[0019] In this application, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment that is mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this application may be combined with other embodiments without conflict.

[0020] Unless otherwise defined, the technical or scientific terms used in this application shall have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms “a,” “an,” “an,” “the,” and similar words used in this application do not indicate quantity limitation and may indicate singular or plural. The terms “comprising,” “including,” “having,” and any variations thereof used in this application are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that includes a series of steps or units (components) is not limited to the listed steps or units, but may also include steps or units not listed, or may include other steps or units inherent to these processes, methods, products, or apparatus. The terms “multiple” / “several” used in this application refer to two or more. “And / or” describes the relationship between related objects, indicating that three relationships may exist; for example, “A and / or B” can represent: A alone, A and B simultaneously, and B alone. The character “ / ” generally indicates that the preceding and following related objects are in an “or” relationship. The terms “first,” “second,” “third,” etc., used in this application are merely to distinguish similar objects and do not represent a specific ordering of the objects.

[0021] An illustrative embodiment of the present invention, such as Figure 1 As shown, Figure 1 This is a flowchart illustrating a non-invasive multi-cancer AI screening method based on tumor reactive TCR analysis according to an embodiment of the present invention. The non-invasive multi-cancer AI screening method based on tumor reactive TCR analysis provided in this application includes: Step S102: Extract peripheral blood from the obtained subjects to obtain sequencing data; Optionally, in step S102, the extraction of peripheral blood from the obtained subjects to obtain sequencing data includes: separating peripheral blood mononuclear cells from the obtained peripheral blood through a preset procedure; extracting RNA or DNA from the peripheral blood mononuclear cells to construct a TCR sequencing library; and obtaining sequencing data from the TCR sequencing library through a sequencing platform.

[0022] Step S104: Pair the α and β chains of the T cell receptor in the sequencing data to generate a peripheral blood TCR library of the subject. Optionally, step S104, which involves pairing the α and β chains of the T cell receptor in the sequencing data to generate the subject's peripheral blood TCR library, includes: aligning and assembling the T cell receptor in the sequencing data to obtain the variable region sequences of the α and β chains of the TCR; pairing the variable region sequences of the α and β chains of the TCR; removing clones with low quality, low read counts, and incomplete V or J annotations; and statistically analyzing the cloning frequencies of paired clones for each α chain and each paired clone for each β chain to construct the subject's peripheral blood TCR library.

[0023] Step S106: Extract the amino acid sequences of the α chain and β chain from the peripheral blood TCR library according to a preset order, and splice the obtained amino acid sequences of the α chain and β chain to obtain a structured full-chain amino acid sequence. Optionally, in step S106, the amino acid sequences of the α-chain and β-chain are extracted from the peripheral blood TCR library according to a preset order, and the obtained amino acid sequences of the α-chain and β-chain are spliced ​​together to obtain a structured full-chain amino acid sequence. This includes: extracting the amino acid sequences of the α-chain and β-chain respectively in a preset order including first extracting the V region, then extracting the CDR3 region, and finally extracting the J region; connecting the amino acid sequences of the α-chain and β-chain with connectors to generate a single structured amino acid sequence; and identifying the single structured amino acid sequence as the structured full-chain amino acid sequence.

[0024] Specifically, the TCR representation based on the paired V–CDR3–J full-chain structure is as follows: The TCR consists of an α chain and a β chain. The TCR α chain is composed of the following gene segments: V (Variable) refers to the variable region gene segment; J (Joining) region gene fragment; C (Constant) constant region gene fragments; The TCRβ chain is composed of the following gene segments: V (Variable) refers to the variable region gene segment; D (Diversity) gene fragments J (Joining) region gene fragment; C (Constant) constant region gene fragments; The diversity of TCRs primarily originates from the VJ(D) rearrangement region. Due to the structural differences between the α and β chains, a novel data organization method, "V-CDR3-J," has been proposed. In the α chain, the CDR3 region is formed by the end of the V fragment and the beginning of the J fragment; while in the β chain, the CDR3 region is composed of the end, the D fragment, and the beginning of the J fragment. Therefore, using a unified "V-CDR3-J" structure can accommodate both the α and β chains simultaneously and retain key information related to TCR recognition to the greatest extent possible. Figure 2 As shown, Figure 2 This is a schematic diagram of the gene structure of the α and β chains in a non-invasive multi-cancer AI screening method based on tumor reactive TCR analysis according to an embodiment of the present invention. For each TCR clone, the paired α and β chain variable region amino acid sequences are first obtained using a TCR reconstruction tool, and the order information of the V region, CDR3 region, and J region within the chain is preserved. A fixed rule is adopted, that is, "α chain V region – α chain CDR3 – α chain J region – separator – β chain V region – β chain CDR3 – β chain J region" is spliced ​​into a structured amino acid sequence, which preserves the intra-chain motifs and explicitly encodes the inter-chain pairing relationship. The structured full-chain amino acid sequence was directly input into a large-scale protein language model, without relying on manually designed k-mer word frequency and AAindex features.

[0025] Step S108: Input the structured full-chain amino acid sequence into the pre-trained model to obtain the tumor reactivity probability value of the corresponding TCR clone; Optionally, the training process of the pre-trained model in step S108 includes: collecting paired TCR sequences from multiple cancer tumor samples that have been functionally validated as tumor reactivity, as positive samples; collecting paired TCR sequences from peripheral blood of healthy individuals, deleting sequences that are repetitive with or highly similar to the positive samples, as high-confidence negative samples; balancing positive and negative samples according to a preset ratio to determine the TCR sequences; splicing the α and β chains in the TCR sequences into a structured full-chain amino acid sequence according to a preset order; encoding the full-chain amino acid sequence using a word segmenter, truncating or padding to the maximum length; performing end-to-end adjustments using cross-entropy loss and Adam optimization, extracting hidden layer representations for t-SNE visualization and comprehensive gradient interpretation, verifying that the model can identify key amino acid residues and domains related to tumor reactivity, and obtaining a pre-trained model that has converged to training.

[0026] The training process for the pre-trained model is as follows: Step 1, Dataset Construction: Collect paired TCR sequences from tumor samples of multiple cancer types that have been functionally validated as tumor-reactive, as positive samples; collect paired TCR sequences from peripheral blood of healthy individuals, and delete sequences that are duplicates or highly similar to the positive samples, as high-confidence negative samples; balance the positive and negative samples in a ratio of approximately 1:1, with a total of 30,372 TCR sequences.

[0027] Step 2, Sequence Encoding and Model Structure: Assemble the V–CDR3–J regions of the α / β chain into a structured sequence according to the aforementioned rules; use the ESM-2 tokenizer to encode the amino acid sequence, truncating or padding to the maximum length (e.g., 512 tokens).

[0028] Step 3, Model Training: End-to-end fine-tuning was performed using cross-entropy loss and Adam optimization. Hidden layer representations were extracted for t-SNE visualization and comprehensive gradient interpretation to verify that the model can identify key amino acid residues and domains related to tumor reactivity.

[0029] The model is fine-tuned end-to-end using cross-entropy loss and Adam optimization, and the training and test sets are divided in an 8:2 ratio, with the model's performance on the test set serving as the evaluation criterion for model training.

[0030] Specifically, in this embodiment, experimentally validated tumor-reactive TCRs from various cancer types and tumor-infiltrating lymphocytes are selected as positive samples, and high-confidence negative TCRs from the peripheral blood TCR database of healthy individuals are selected as negative samples to construct a balanced training set; such as Figure 3 As shown, Figure 3 This is a schematic diagram of a non-invasive multi-cancer AI screening method based on tumor reactivity TCR analysis according to an embodiment of the present invention, after fine-tuning the model using the ESM-2 protein language model. Using the pre-trained protein language model ESM-2 as the backbone, the structured TCR sequence is fine-tuned end-to-end, context-sensitive sequence embeddings are extracted, and the tumor reactivity probability between 0 and 1 is output to characterize the likelihood of a single TCR recognizing a tumor antigen.

[0031] Step S110: Calculate the sample-level tumor score based on the tumor reactivity probability value; Optionally, step S110, which calculates the sample-level tumor score based on the tumor reactivity probability value, includes: sorting the tumor reactivity probability values, selecting the Top N clones, where N is a natural number; multiplying the frequency of each clone by the tumor reactivity probability value and summing the results to obtain the tumor reactivity intensity of the sample; and determining the tumor reactivity intensity of the sample as the sample-level tumor score.

[0032] Optionally, the non-invasive multi-cancer AI screening method based on tumor reactive TCR analysis provided in this application embodiment further includes: selecting the Top N clones; calculating the tumor reactive intensity of the sample according to a preset formula; and mapping the tumor reactive intensity to a sample-level tumor score in the range (0,1) through a preset transformation function.

[0033] Furthermore, optional, preset formulas include: Where S is the aggregate score, which is the tumor reactivity prediction score for each clone output and the tumor reactivity prediction frequency for each clone output; The preset transformation functions include: Where e is the natural constant, k is the coefficient, k is the value calculated based on the baseline established for healthy subjects, and Score is the sample-level tumor score; It should be noted that in this embodiment, k is a coefficient constructed using the scores of healthy individuals (i.e., healthy subjects in this embodiment), and in this embodiment, k=48.

[0034] Specifically, for peripheral blood samples, the clonal frequency of each unique α / β paired clonal type was first counted; For each clone, the first-level model described above is invoked to obtain its tumor reactivity probability; Select a number of clones with the highest cloning frequency (such as the top 25, i.e. N in the embodiments of this application), and sum the weighted values ​​of “cloning frequency × tumor reactivity probability” to obtain the original tumor reactivity intensity of the sample. The original intensity is then mapped to the (0,1) interval using the transformation function described above, forming a standardized "tumor score". The baseline threshold (e.g., 0.5) is determined by the distribution in the healthy population to distinguish between negative and positive samples.

[0035] The multi-cancer non-invasive screening process design in this application embodiment is as follows: by integrating peripheral blood TCR data from various cancers (such as colorectal cancer, melanoma, non-small cell lung cancer, Hodgkin lymphoma, etc.), the distribution of tumor scores is modeled. By adopting a unified tumor score threshold or a cancer type-specific stratification strategy, we can achieve the "presence or absence of tumor" determination in pan-cancer (major / few / zero-shot) scenarios, and perform risk stratification and auxiliary diagnosis based on the prior information of different cancer types in the training set. The same procedure can be repeatedly applied to longitudinal follow-up samples to achieve dynamic monitoring of the proportion of tumor-responsive TCRs during immunotherapy.

[0036] The model evaluation and tumor score threshold construction in this embodiment are as follows: In a preferred example, the model is evaluated holistically on queued data: The model was used to predict the proportion of non-tumor reactive TCRs in peripheral blood TCRs from healthy individuals (83,901 entries), non-small cell lung cancer (NSCLC) patients (23,761 entries), melanoma patients (8,216 entries), and COVID-19 patients (103,532 entries), revealing the model's ability to predict tumor-reactive TCRs at the overall peripheral blood environment. Results showed that the model's predicted proportion of non-reactive TCRs was almost identical in the healthy group (0.9059, σ=0.0021) and the COVID-19 group (0.9033, σ=0.0022), with minimal difference between the two groups. In contrast, the proportion of non-reactive TCRs was significantly higher in the NSCLC (0.9234, σ=0.0038) and melanoma (0.9185, σ=0.0055) cohorts, showing clear differences from the healthy population and the COVID-19 population. This demonstrates that the model possesses overall tumor recognition capabilities and high robustness against interference.

[0037] In a preferred example, the model evaluates multiple individuals independently: The model was used to predict cancer in 9 melanoma patients (with a large number of TCRs reactive to this cancer type in the training set) using a Major shot (5 patients were successfully predicted to have cancer), 5 colorectal cancer patients (with a small number of TCRs reactive to this cancer type in the training set) using a Few shot (4 patients were successfully predicted to have cancer), and 8 rare Hodgkin's lymphoma patients (with no TCRs reactive to this cancer type in the training set) using a Zero shot (2 patients were successfully predicted to have cancer). The model was also used to predict cancer in 12 healthy individuals; none of them were diagnosed with cancer. In 42 COVID-19 patients, only one patient was predicted to have cancer, but the score was low, only 0.03 above the threshold. The model demonstrates strong robustness against interference and a low false negative rate, showcasing its potential for early cancer screening.

[0038] In a preferred example, the tumor score threshold is constructed as follows: The tumor reactivity prediction score output by each clone in the model is multiplied by its corresponding frequency to obtain a weighted reactivity score. To focus on dominant clones in the immune response, the 25 most frequent TCR clones in the sample were selected, and their weighted reactivity scores were summed to form a preliminary aggregate score S. Finally, to ensure the scores are comparable and to perform normalization, the following nonlinear transformation function is applied to transform the aggregated scores: Here, e is a natural constant, and a baseline score is used to construct a score that can distinguish between patients with cancer and those with good health, ensuring a clear boundary for classifying cancer patients. This threshold is set to 0.5. To determine k, a score is constructed in the healthy population, and k is finally calculated to be 48.

[0039] Step S112: Generate pan-cancer tumor risk screening results based on sample-level tumor scores.

[0040] Optionally, the step S112 of generating pan-cancer tumor risk screening results based on sample-level tumor scores includes: determining the risk of developing tumors by setting a baseline threshold, with higher scores indicating a greater probability of developing tumors.

[0041] Specifically, the non-invasive multi-cancer AI screening method based on tumor reactive TCR analysis provided in this application embodiment can be deployed on a local server or cloud platform to provide standardized peripheral blood multi-cancer AI screening services for hospitals or testing institutions.

[0042] In summary, combining steps S102 to S112, such as Figure 4 As shown, Figure 4 This is a flowchart illustrating another non-invasive multi-cancer AI screening method based on tumor reactive TCR analysis according to an embodiment of the present invention.

[0043] Step 1. Sample collection and TCR sequencing: Peripheral blood was collected from the subjects, peripheral blood mononuclear cells (PBMCs) were separated using standard procedures, and RNA or DNA was extracted to construct a TCR sequencing library (such as 5'RACE or target enrichment protocol). Raw sequencing data were obtained using a high-throughput sequencing platform.

[0044] Step 2. TCR Reconstruction and Paired Clone Library Construction: TCR reconstruction software (such as TRUST4, Cell Ranger, etc.) is used to align and assemble sequencing data to obtain the variable region sequences of the TCR α and β chains and perform chain pairing; clones with low quality, low read count and incomplete V / J annotation are removed, the clone frequency of each unique α / β paired clone is counted, and the peripheral blood TCR library of the subjects is constructed.

[0045] Step 3. TCR full-chain structured encoding: For each clone, record the α chain and β chain sequences in the order of “V region – CDR3 – J region”; connect the α / β chains with a preset delimiter to generate a single structured full-chain amino acid sequence, which will be used as input for the subsequent large model.

[0046] Step 4. Tumor Reactivity Probability Prediction: Input the above structured sequence into the trained model and output the tumor reactivity probability value for each TCR clone.

[0047] Step 5. Calculate the sample-level tumor score: Sort the clones from high to low frequency and select the top 25 clones; multiply the frequency of each clone by its tumor reactivity probability and sum them to obtain the original tumor reactivity intensity of the sample.

[0048] Step 6. Output Results: Input the raw intensity into the self-built algorithm function to obtain a tumor score ranging from 0 to 1. The baseline threshold is 0.5; a score higher than 0.5 indicates a risk of developing tumors.

[0049] This application provides a non-invasive AI screening method for multiple cancer types based on tumor reactive TCR analysis, belonging to the interdisciplinary fields of bioinformatics, tumor immunology, and artificial intelligence. Specifically, it relates to a method and system for non-invasive screening and dynamic monitoring of multiple cancer types using high-throughput sequencing data of peripheral blood T cell receptors (TCRs) combined with a large-scale protein language model. TCRs are heterodimeric receptors on the surface of T cells, mainly composed of α and β chains, which mediate adaptive immune responses by recognizing antigenic peptides presented by MHC molecules. During tumor development, tumor antigens can induce clonal expansion of tumor reactive T cells, which accumulate in tumor tissue and peripheral blood. Therefore, analyzing the composition and clonal expansion characteristics of the TCR library can reflect an individual's anti-tumor immune status, providing novel non-invasive biomarkers for early tumor screening and efficacy monitoring.

[0050] Based on the above, this application takes the calculation of tumor scores from the peripheral blood of subjects as an example. The specific details of the non-invasive multi-cancer AI screening method based on tumor reactive TCR analysis provided by this application are as follows: Step 1, Peripheral Blood Mononuclear Cell Extraction (PBMC): Peripheral blood from the subject was collected in a vacuum blood collection tube containing an anticoagulant, and then peripheral blood mononuclear cells (PBMC) were separated. The obtained cells were stored at −80°C for use in subsequent experiments.

[0051] Step 2, Pre-sequencing Preparation: Total RNA was extracted from the preserved PBMCs, and its concentration and purity were assessed. Subsequently, a low-biased TCR cDNA library was constructed based on the obtained RNA using a 5' rapid amplification (5' RACE) strategy. Two rounds of PCR amplification were then performed using primers targeting the TCR region. The products from the second round of amplification were separated in an approximately 2% polyacrylamide gel, and a band of approximately 500 bp was excised. The gel band was recovered and purified using a commercial DNA purification kit (such as a domestic or other brand-name gel extraction system) and used as library material for subsequent sequencing.

[0052] Step 3, Sequencing: Dilute the constructed and quality-controlled library to the specified concentration as required, and load it onto a high-throughput sequencing platform for TCR / RNA sequencing. Illumina series sequencers (such as NovaSeq 6000, NextSeq2000, or MiSeq) are recommended. Select the appropriate sequencing mode according to the kit instructions (commonly used is PE150). Follow the instrument operation procedures to complete library preparation, sample mixing, cluster generation, and sequencing setup to ensure effective read length coverage of the TCR V(D)J region or target RNA fragment.

[0053] Step 4, TCR assembly and pairing: The original sequence obtained from sequencing is compared with the reference genome. Non-TCR derived reads are filtered out, and the reads of the correctly matched TCR gene fragments are assembled to obtain a complete full-length TCR single strand. The TCR single strand is then annotated to obtain a TCR single strand with the gene fragment name. After filtering the obtained full-strand TCR, the TCR α and β strands from the same cell are paired using barcodes (tools such as TRUST4 and Cell Ranger can be used).

[0054] Step 5: Convert to TCR amino acid sequence: Translate the nucleic acid sequence into an amino acid sequence to determine the accurate V, D, and J gene fragments. At the same time, remove missing fragments or clones with extremely low abundance to reduce errors and improve sequence reliability.

[0055] Step 6, convert to input data format: organize the amino acid fragments of the V gene of the α and β chains of TCR, the amino acid fragments of CDR3, the amino acid fragments of the J gene, the number of reads, and put them in an Excel spreadsheet.

[0056] Step 7, Input Model and Calculate Results: Input the model data, and the algorithm will calculate and output the tumor score. (See attached image) Figure 5 As shown, Figure 5 This is a schematic diagram of the algorithm usage interface in a non-invasive multi-cancer AI screening method based on tumor reactive TCR analysis according to an embodiment of the present invention.

[0057] Compared to existing technologies, the non-invasive multi-cancer AI screening method based on tumor reactive TCR analysis provided in this application achieves significant technical effects in the following aspects by integrating a large tumor reactive TCR identification model, a peripheral blood tumor scoring model, and a multi-cancer workflow: 1. Improved the specificity and stability of non-invasive peripheral blood screening: In large-sample healthy cohorts and COVID-19 virus infection cohorts, the tumor score distribution calculated by this invention was highly consistent, with very few false positives; indicating that the model has good anti-interference ability against non-tumor-related immune activation and is suitable for large-scale pan-cancer screening.

[0058] 2. Achieved antigen-agnostic cross-cancer generalization capability: Through large-scale protein language models and full-chain structure input, this invention can give reasonable high tumor scores in a variety of cancers (including cancers that did not appear in the training) and remain stable in major / few / zero-shot scenarios.

[0059] 3. Significantly enhances the biological interpretability and scalability of the model: By visualizing the ESM-2 latent space and attention weights, this invention can identify key amino acid residues and domains that contribute significantly to tumor reactivity discrimination, providing candidate targets for subsequent functional validation and personalized TCR treatment; the unified "tumor score" index can be easily combined with other clinical indicators, extending to scenarios such as prognostic assessment, efficacy prediction, and dynamic monitoring of clinical treatment.

[0060] 4. It provides a standardized and easily engineered multi-cancer AI screening process: the entire process from raw sequencing data, TCR whole chain construction, large model inference, tumor score calculation to report generation is automated and can be integrated into hospital information systems or third-party testing platforms, reducing the threshold for clinical use; the same process can be used for baseline screening and follow-up monitoring, which is conducive to forming a closed-loop immune monitoring system.

[0061] 5. The technical approach differs fundamentally from existing TCR early screening patents in key implementation paths: The non-invasive multi-cancer AI screening method based on tumor reactive TCR analysis provided in this application does not rely on k-mer abundance statistics and AAindex features, but instead uses end-to-end representation learning based on the paired α / β chain V–CDR3–J full-chain structure and a large-scale protein language model; the screening logic changes from directly classifying whether a sample has cancer to first identifying tumor reactive TCR and then aggregating it into a sample-level tumor score, which is significantly different from existing schemes based on peripheral blood immune features in terms of modeling perspective and core algorithm.

[0062] This invention employs the above technical solution, extracting sequencing data from peripheral blood of the subjects; pairing the α and β chains of the T cell receptor in the sequencing data to generate a peripheral blood TCR library; extracting the amino acid sequences of the α and β chains from the peripheral blood TCR library according to a preset order, and splicing the obtained α and β chain amino acid sequences to obtain a structured full-chain amino acid sequence; inputting the structured full-chain amino acid sequence into a pre-trained model to obtain the tumor reactivity probability value of the corresponding TCR clone; calculating a sample-level tumor score based on the tumor reactivity probability value; and generating pan-cancer tumor risk screening results based on the sample-level tumor score. Compared with existing technologies, this invention has the following technical effects: it achieves accurate, stable, and non-invasive assessment of cancer risk using only a very low proportion of TCR full-chain sequence information with a high-noise background in peripheral blood, and also achieves a lower false positive rate in multi-tumor screening.

[0063] The above description is merely a preferred embodiment of the present invention and does not limit the implementation and protection scope of the present invention. Those skilled in the art should realize that any equivalent substitutions and obvious changes made based on the description and illustrations of the present invention should be included within the protection scope of the present invention.

Claims

1. A non-invasive multi-cancer AI screening method based on tumor reactive TCR analysis, characterized in that, include: Peripheral blood samples were extracted from the subjects to obtain sequencing data; The α and β chains of the T cell receptor in the sequencing data were paired to generate the peripheral blood TCR library of the subject; The amino acid sequences of the α chain and the β chain are extracted from the peripheral blood TCR library according to a preset order, and the amino acid sequences of the α chain and the β chain are spliced ​​together to obtain a structured full-chain amino acid sequence. The structured full-chain amino acid sequence is input into a pre-trained model to obtain the tumor reactivity probability value of the corresponding TCR clone; Calculate the sample-level tumor score based on the tumor reactivity probability value; Based on the sample-level tumor scores, pan-cancer tumor risk screening results are generated.

2. The non-invasive multi-cancer AI screening method based on tumor reactive TCR analysis according to claim 1, characterized in that, The extraction of peripheral blood from the subjects to obtain sequencing data includes: Peripheral blood mononuclear cells were separated from the obtained peripheral blood of the subject using a preset procedure; TCR sequencing libraries were constructed by extracting RNA or DNA from peripheral blood mononuclear cells. The sequencing data is obtained through a sequencing platform based on the TCR sequencing library.

3. The non-invasive multi-cancer AI screening method based on tumor reactive TCR analysis according to claim 1 or 2, characterized in that, The step of pairing the α and β chains of the T cell receptor in the sequencing data to generate the peripheral blood TCR library of the subject includes: The T cell receptor in the sequencing data was compared and assembled to obtain the variable region sequences of the α chain and β chain of the TCR; Chain pairing was performed on the variable region sequences of the α and β chains of the TCR. By removing clones with low quality, low read counts, and incomplete V or J annotations, the clone frequencies of paired clones for each α chain and each paired clone for each β chain were counted to construct the peripheral blood TCR library of the subject.

4. The non-invasive multi-cancer AI screening method based on tumor reactive TCR analysis according to claim 3, characterized in that, The step of extracting the amino acid sequences of the α chain and the β chain from the peripheral blood TCR library according to a preset order, and splicing the obtained amino acid sequences of the α chain and the β chain to obtain a structured full-chain amino acid sequence includes: When the preset order includes first extracting the V region, then extracting the CDR3 region, and finally extracting the J region, the amino acid sequences of the α chain and the β chain are extracted respectively to obtain the amino acid sequences of the α chain and the β chain. The amino acid sequences of the α chain and the β chain are connected by a connector to generate a single structured amino acid sequence. The single structured amino acid sequence is identified as the structured full-chain amino acid sequence.

5. The non-invasive multi-cancer AI screening method based on tumor reactive TCR analysis according to claim 4, characterized in that, The training process of the pre-trained model includes: Paired TCR sequences that were functionally validated as tumor-reactive from tumor samples of multiple cancer types were collected as positive samples; Paired TCR sequences were collected from peripheral blood of healthy individuals, and sequences with positive duplications and high similarity were removed to obtain high-confidence negative samples. Positive and negative samples were balanced according to a preset ratio to determine the TCR sequence; The α and β chains in the TCR sequence are assembled into a structured full-chain amino acid sequence according to the preset order; The full-chain amino acid sequence is encoded using a word segmenter, and then truncated or padded to the maximum length. End-to-end tuning was performed using cross-entropy loss and Adam optimization. Hidden layer representations were extracted for t-SNE visualization and comprehensive gradient interpretation. The model was validated to identify key amino acid residues and domains related to tumor reactivity, resulting in a pre-trained model that had achieved training convergence.

6. The non-invasive multi-cancer AI screening method based on tumor reactive TCR analysis according to claim 5, characterized in that, The calculation of the sample-level tumor score based on the tumor reactivity probability value includes: The tumor reactivity probability values ​​are sorted, and the Top N clones are selected, where N is a natural number; The tumor reactivity intensity of the sample is obtained by multiplying the frequency of each clone by the tumor reactivity probability value and summing the results. The tumor reactivity intensity of the sample is determined as the sample-level tumor score.

7. The non-invasive multi-cancer AI screening method based on tumor reactive TCR analysis according to claim 6, characterized in that, The method further includes: Select the top N clones; The tumor reactivity intensity of the sample is calculated according to a preset formula; The tumor reactivity intensity is mapped to the sample-level tumor score in the range (0,1) through a preset transformation function.

8. The non-invasive multi-cancer AI screening method based on tumor reactive TCR analysis according to claim 7, characterized in that, The preset formula includes: Wherein, S is the polymerization fraction, representing the tumor reactivity intensity; P i The tumor reactivity prediction score output for each clonus, F i The predicted frequency of tumor reactivity output for each clone; The preset transformation function includes: Where e is a natural constant; k is a coefficient, calculated based on a baseline established for healthy subjects; and Score is the sample-level tumor score.

9. The non-invasive multi-cancer AI screening method based on tumor reactive TCR analysis according to claim 6, characterized in that, The generation of pan-cancer tumor risk screening results based on the sample-level tumor scores includes: By setting a baseline threshold, individuals with a risk of developing cancer are identified whose levels exceed the baseline threshold.

Images