Biological macromolecule candidate molecular property prediction and enhancement discovery method and system based on adaptive combination molecular property cross-attention model

Abstract

The application discloses a biological macromolecule candidate molecule property prediction and enhancement discovery method and system based on an adaptive binding molecule property cross attention model, which sequentially performs eight steps of data input, sequence preprocessing and CDR region annotation, protein language model (PLM) hierarchical fine-tuning, ABPA cross attention modeling and reinforcement learning optimization, dynamic threshold virtual pre-screening, wet experiment verification, experimental data backflow iteration, standardized candidate list output. According to different prediction tasks, the application automatically allocates feature weights, simultaneously predicts affinity, expression, stability, cell activity, aggregation risk and other molecular properties; differentiates early screening and lead optimization to set differentiated screening thresholds, forms a model self-iteration closed loop through wet experiment data backflow, and is combined with MLOps to realize full-process traceability, which can shorten the research and development cycle of biological macromolecule drugs such as antibodies, TCE and CAR, and is suitable for industrialized multi-target protein drug virtual screening platforms.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics, and in particular to a method and system for predicting and enhancing the discovery of candidate molecular properties of biological macromolecules based on an adaptive cross-attention model of molecular properties. Background Technology

[0002] In the discovery of biopharmaceutical drugs, the number of candidate molecules, including antibodies, nanobodies (VHH), Fab, scFv, T-cell binders (TCE), chimeric antigen receptor (CAR) binding domains, and other protein-binding molecules, is typically large. In existing technologies, the screening and ranking of candidate binding molecules mainly rely on wet experimental data such as affinity assays, cell function experiments, expression level tests, stability and development experiments. For example, in TCE or other biotherapy development scenarios, researchers typically first construct a library of candidate binding molecules, then obtain affinity, activity, or other property data for these molecules through high-throughput screening, SPR / BLI, ELISA, cell killing or activation assays, etc., and subsequently perform manual prioritization and further optimization based on the experimental results.

[0003] In recent years, protein language models (PLMs) and deep learning models have been developed to learn protein representations from amino acid sequences and to predict protein structure, function, interactions, or properties. Some existing methods can classify or regress binding affinity using antibody sequences, antigen sequences, or protein embeddings, while others combine structure prediction, molecular docking, traditional machine learning, or deep learning models for candidate screening.

[0004] However, existing machine learning-based molecular property prediction methods are usually still single-point models or offline analysis tools: they lack a unified closed loop between input data, sequence preprocessing, model fine-tuning, inference, candidate ranking, experimental feedback, and retraining; the models mostly focus on a single endpoint of affinity, making it difficult to extend to other properties such as specificity, expression level, stability, aggregation risk, exploitability, cross-reactivity, functional activity, or manufacturing-related attributes within the same algorithmic framework; at the same time, existing methods often lack cross-attention mechanisms that can adaptively allocate feature weights for relationships between heavy chains, light chains, different CDR regions, antigens, or application contexts.

[0005] The ABPA (AdaptiveBinderPropertyX-attention) approach uses protein language models to directly predict the binding efficiency or properties of molecules from amino acid sequences, and achieves similar ROCAUC on both the training and test sets (e.g., approximately 0.78 on the training set and approximately 0.75 on the test set), indicating that the model has a certain discriminative ability and does not show obvious overfitting or underfitting. This approach can also be integrated into an enhanced biomacromolecule discovery workflow that includes sequence preparation, PLM fine-tuning, prediction, reinforcement learning, cloud training management, experimental validation, and model updates.

[0006] (1) The experimental screening is costly and time-consuming. Traditional candidate binding molecule screening requires the construction and testing of a large number of molecules. Properties such as affinity, functional activity and development potential all depend on different experimental platforms, which leads to a lot of time spent on sample registration, data processing, experimental scheduling and result analysis.

[0007] (2) Insufficient model-experiment closed loop. Existing machine learning models are usually trained offline on historical data. There is a lack of standardized process between prediction results and subsequent candidate selection, experimental verification, experimental result feedback and model iteration, making it difficult to continuously improve prediction performance as project data accumulates.

[0008] (3) The prediction target is too singular. Many models only model binding affinity or binary hit rate, and it is difficult to extend to properties other than affinity, such as specificity, selectivity, expression level, thermostability, aggregation tendency, immunogenicity risk, cross-reactivity, cellular functional activity, manufacturability or multi-property comprehensive score.

[0009] (4) Insufficient utilization of sequence regions and contextual information. Different regions of the antibody or binding protein contribute differently to the properties. For example, the effects of CDR-H3, CDR-H2, light chain CDR, framework region, and antigen contact region are not the same. Existing simple embedding splicing or global averaging methods are difficult to adaptively identify key regions according to specific properties and application scenarios.

[0010] (5) The candidate priority ranking lacks a decision-making mechanism oriented towards actual R&D costs. In real projects, the costs of false positives and false negatives are different. Early screening scenarios can accept more false positives to reduce the loss of potential good molecules, while later optimization scenarios focus more on high-confidence candidates. Existing methods usually output a single score and lack a mechanism to adjust thresholds, rankings, and experimental recommendations based on application constraints.

[0011] (6) Lack of integrated implementation with cloud-based MLOps and enhanced discovery processes. Existing script-based models are difficult to support large-scale virtual screening of candidate libraries, model version management, training experiment tracking, manual review, wet experiment feedback and automatic retraining, which limits their industrial application in biomacromolecule discovery projects. Summary of the Invention

[0012] To address the aforementioned issues, the present invention aims to provide a method and system for predicting and enhancing the discovery of candidate molecules in biomacromolecules based on an adaptive cross-attention model of molecular properties. This method effectively improves the efficiency of experimental resource utilization, enhances the interpretability and adaptability of candidate ranking, and has strong scalability. It can be adapted to different R&D stages, reduces manual data processing and candidate ranking, improves model traceability and reproducibility, helps shorten the time for generating and evaluating bound molecules, and improves the overall quality of development.

[0013] To solve the above-mentioned technical problems, the first technical solution adopted by the present invention is as follows: To address the aforementioned technical problems, the first technical solution adopted by this invention is as follows: A method for predicting and enhancing the discovery of candidate molecular properties of biomacromolecules based on an adaptive molecular property cross-attention model, comprising the following steps: S1. Input candidate binding molecule data, which includes the amino acid sequence of the binding molecule, antigen / target sequence, structural features, experimental tags, and project research and development constraints. S2. Perform data preprocessing and region labeling on candidate binding molecule data: Clean and deduplicate sequences, remove illegal amino acids and invalid records, and classify and label the main chain, light chain, CDR region, framework region, linker peptide, and functional domain based on the protein numbering system. Standardize multi-source experimental labels into binary labels, regression labels, ordination labels, or weak labels with confidence, while encoding species, experimental platform, and target family contextual features. S3. Perform supervised fine-tuning of the protein language model (PLM) using publicly available protein datasets. Input the labeled sequences into the pre-trained protein language model to generate residue-level, CDR region-level, and full-sequence multi-level embedding representations. Adaptively select fine-tuning strategies according to the project sample size. For small samples, freeze the PLM backbone and train only the adapter layer, low-rank adapter module, or prediction head. For large samples, fine-tune some or all parameters of the PLM layer by layer. S4. Construct an ABPA adaptive molecular property cross-attention model, and optimize the ABPA model by constructing a multi-objective reward function based on the project's internal experimental data. The ABPA adaptive molecular property cross-attention model takes a multi-level PLM embedding representation as input, and fuses the interaction features of heavy chain, light chain, each CDR region, combined molecule and antigen, sequence and experimental context through a multi-head cross-attention layer. It dynamically and adaptively adjusts the attention weights of each feature region for different target molecular properties, and outputs multiple molecular property prediction values, prediction confidence and uncertainty scores simultaneously for multiple tasks. S5. Apply the optimized ABPA model to candidate molecule pruning prediction. Configure dynamic screening thresholds based on prediction scores, prediction uncertainties, and false positive / false negative R&D costs. Differentiate between early high-throughput screening and later pilot optimization scenarios to stratify candidate molecules, screen out low-quality candidate molecules, and complete virtual pre-screening. S6. Submit the priority candidate molecules after stratified screening to wet experimental verification, perform protein construction and assembly, binding detection, cell function, stability, expression development experiments, and support researchers to manually review and adjust the candidate set based on sequence diversity and epitope information; S7. Collect the new experimental label data produced by the wet experiment and feed it back to the dataset. Input the newly added internal experimental data into S4 and reconstruct the reward function to achieve iterative enhancement of the ABPA model. S8. Output a standardized candidate list, which includes multi-dimensional predicted properties, confidence level, experimental priority, model version, and training metric information for each molecule.

[0014] In the above technical solution, the candidate biomolecules for binding include any one or more of the following: antibodies, nanobodies (VHH), Fab, scFv, DARPin, TCR, CAR binding domains, T cell binders (TCE), bispecific / multispecific engineered proteins, and receptor ligand proteins. The protein numbering system can be any one of IMGT, Kabat, Chothia, or AHo; for non-antibody binding molecules, sequence regions are segmented and labeled according to loops, binding motifs, and conserved domains.

[0015] In the above technical solution, the target molecule properties include binding affinity, binding efficiency, specificity, cross-reactivity, cell activation / inhibition activity, expression level, solubility, thermal stability, aggregation tendency, immunogenicity risk, serum stability, endocytosis capacity, manufacturability, or a comprehensive candidate quality score obtained by weighting multiple properties; the same ABPA model framework can achieve parallel prediction of different molecular properties by replacing training labels, model output heads, and loss functions.

[0016] In the above technical solution, the pre-trained protein language model includes ESM, ESM2, ESM3, ProtBERT, ProtT5, TAPE, IgLM, antibody-specific protein language model, and structure-aware protein large model; the adapter layer and low-rank adapter module are at least one of Adapter and LoRA.

[0017] In the above technical solutions, the multi-head cross-attention can be replaced by collaborative co-attention (Co-Attention), graph attention network, hybrid expert MoE fusion network, and multi-task dual-tower fusion attention network; for affinity prediction tasks, the weights of CDR and antigen interaction features are increased; for expression and stability prediction tasks, the weights of frame region and hydrophobic sequence features are increased; and for TCE and CAR function prediction tasks, the context weights of the construct structure domain are fused.

[0018] In the above technical solution, the dynamic screening threshold configuration logic in S5 is as follows: in the early high-throughput screening scenario, a loose threshold is used to improve the recall rate and reduce the missed detection of high-quality molecules; in the later lead optimization scenario, a strict threshold is used to improve the precision rate and reduce the number of invalid wet experiments; candidate molecules are divided into four categories: recommended experimental verification, retained for observation, sequence redesign, and direct elimination.

[0019] In the above technical solution, the multi-objective reward function described in S4 simultaneously constrains affinity enhancement, specificity enhancement, aggregation risk reduction, expression level enhancement, sequence diversity preservation, and undesirable motif avoidance; mutation enumeration, genetic algorithm, and Bayesian optimization are used to generate the next generation of candidate variants, and the ABPA model is used to complete the rapid pre-evaluation of variants, thereby realizing SAR iterative optimization.

[0020] The above technical solution also includes MLOps full-process management steps: using at least one of the tools SageMaker, MLflow, Kubeflow, Weights & Biases, and Databricks ML to record the original sequence dataset, PLM and ABPA model parameters, training metrics, virtual screening records, and raw wet experiment data, so as to achieve full-link traceability and reproducibility of model training, inference, and iteration; the system deployment environment is public cloud, private cloud, local HPC, or hybrid computing cluster.

[0021] In the above technical solution, the publicly available protein dataset mentioned in S3 includes the SAbDab antibody structure library and provides publicly available scFv, Fab, and VHH sequence data for general protein characterization pre-training; S4 uses project-specific internal wet experimental data to complete the ABPA model reinforcement learning optimization, with publicly available data and internal data working together to complete model construction.

[0022] To address the aforementioned technical problems, the second technical solution adopted by this invention is as follows: a biological macromolecule candidate molecular property prediction and enhanced discovery system based on an adaptive cross-attention model combining molecular properties, used to execute the above method, comprising: a data access and standardization module, a sequence preparation and region labeling module, a PLM fine-tuning module, an ABPA adaptive cross-attention prediction and reinforcement learning optimization module, a candidate ranking and pruning decision module, a virtual screening and wet experiment transfer module, an experimental data feedback iteration module, and an MLOps cloud management module; the ABPA adaptive cross-attention prediction and reinforcement learning optimization module is the core computing unit, used for multi-level sequence embedding feature adaptive weighted fusion, simultaneous prediction of multiple molecular properties, and model iterative optimization based on an internal experimental reward function.

[0023] In summary, the advantages of the technical solution of this invention compared to traditional methods are as follows: This invention, by combining PLM with an adaptive cross-attention model, can directly predict the properties of candidate binding molecules from amino acid sequences, enabling a large number of candidates to undergo computational pre-screening before entering wet experiments, thereby reducing the proportion of low-quality candidates entering high-cost experimental processes. According to technical documentation, this type of process can remove approximately 30% of undesirable candidates using sequence and structure models before high-throughput hit screening, thereby improving the efficiency of experimental resource utilization.

[0024] This invention uses regional representation and cross-attention fusion of heavy chain, light chain, CDR, frame region, antigen, and context features. Compared with simple whole-sequence embedding or artificial feature models, it can better identify sequence regions that contribute significantly to specific properties, thereby improving the interpretability and adaptability of candidate ranking.

[0025] This invention is not only applicable to affinity prediction, but can also be extended to a variety of properties beyond affinity. By replacing the label, output header, and loss function, the same ABPA framework can be used for prediction of specificity, expression level, stability, aggregation risk, exploitability, immunogenicity, functional activity, or a comprehensive multi-property score, thereby avoiding the need to develop completely different models for each property.

[0026] This invention connects model training, inference, candidate selection, scientist review, experimental verification, and model updating into a closed loop, enabling the model to continuously iterate as project experimental data accumulates. This reduces the need for manual data processing and candidate sorting, allowing researchers to dedicate more time to experimental design, result interpretation, and the identification of high-quality leads.

[0027] This invention adapts to different stages of research and development by setting application-specific thresholds based on the costs of false positives and false negatives. For example, in the early screening stage, higher recall can reduce missed potential candidates; in the later optimization stage, higher precision can reduce invalid experiments. This mechanism makes the model's predictions closer to real research and development decisions, rather than just outputting static scores.

[0028] This invention manages model versions, training metrics, datasets, inference results, and experimental feedback through cloud-based MLOps, which improves model traceability and reproducibility, making it suitable for deployment in biomacromolecule discovery platforms with multiple projects, targets, and candidate libraries.

[0029] This invention supports multiple scenarios such as virtual screening, lead optimization, de novo design, SAR model iteration, and manufacturing-related optimization, which helps to shorten the time for generating conjugated molecules, shorten the pre-manufacturing evaluation time for biomacromolecules, and improve the overall quality of candidates entering subsequent development. Attached Figure Description

[0030] The foregoing and other objects, features and advantages of the present invention will become apparent from the following detailed description taken in conjunction with the accompanying drawings.

[0031] Figure 1 This is a schematic diagram of the process of the present invention; Figure 2 This is a complete dry-wet closed-loop flowchart of the ABPA process of the present invention. Detailed Implementation

[0032] Based on the preferred embodiments of the present invention, and through the following description, those skilled in the art can make various changes and modifications without departing from the technical concept of the present invention. The technical scope of the present invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

[0033] The invention will be further described with reference to the following figures: Example 1: S1, input candidate binding molecule sequences, optional antigen sequences, structural features, experimental labels, and project constraints; S2, clean, remove duplicates, label regions, and standardize labels for the sequences. S3, input the sequence into the pre-trained PLM to generate residue-level and region-level representations, and fine-tune them on project data; S4 utilizes the ABPA cross-attention model to fuse heavy chain, light chain, CDR, antigen, and contextual features to output target property predictions; S5: Based on the application scenario, set thresholds and decision rules to sort and stratify candidate molecules; S6 will submit priority candidates for experimental verification or build assembly; S7 feeds the experimental results back to the dataset and updates the model through active learning, reinforcement learning, or retraining. S8 outputs a candidate list, prediction properties, confidence scores, recommended experiment priorities, and model version information.

[0034] Example 2: S1 data input imports all scFv sequences, antigen sequences, physicochemical experimental labels, and early screening project constraints.

[0035] S2 Data Preprocessing and Region Labeling Sequence cleaning: Empty sequences, sequences containing illegal amino acids, and duplicate clone sequences were removed, leaving 9264 valid sequences; Region division: The IMGT numbering system is used to automatically label CDR-H1 / H2 / H3, CDR-L1 / L2 / L3, FR1-4, and Linker connection areas; Label standardization: affinity is converted into continuous regression values, cluster risk binary classification labels, and expression level regression labels; target species and SPR experimental platform are used as contextual features.

[0036] S3PLM Monitoring and Fine-tuning Public datasets: The basic representation of the ESM2 pre-trained model was completed using the SAbDab public scFv and Fab antibody libraries; With a moderate sample size, a LoRA low-rank adaptation fine-tuning strategy was adopted, the ESM2 backbone weights were frozen, and only the LoRA adaptation layer and the multi-task prediction head were trained. Output three levels of embedding vectors for each sequence: residue level, CDR region level, and whole sequence level.

[0037] S4ABPA Model Construction and Reinforcement Learning Optimization An ABPA multi-head cross-attention network is constructed: self-attention completes the interaction of features within the light and heavy chains and each CDR; X cross-attention realizes the fusion of cross-regional features between antibody CDRs and antigen sequences; adaptive weight gating automatically increases the CDR-antigen interaction weight for affinity tasks and amplifies the weight of the FR frame region for stability tasks. A multi-objective reward function was constructed using historical experimental data from within the project. The reward items included: high affinity, high expression, low aggregation, and sequence diversity. The ABPA multi-task output head is trained iteratively using reinforcement learning, and the affinity prediction value, expression level, clustering probability, and prediction confidence are output simultaneously. Model metrics: ROCAUC=0.78 for training set, ROCAUC=0.75 for test set, indicating no significant overfitting.

[0038] S5 dynamic threshold virtual pruning pre-screening is currently an early high-throughput screening process, with a relaxed recall threshold set, and stratification based on baseline prediction scores and uncertainty. Recommended experimental candidates: 2100; Observations retained: 3300; 3,864 low-quality candidates were directly eliminated, and approximately 31% of inferior candidates were pre-screened and removed.

[0039] S6 wet experiments validated plasmid construction, protein expression and purification, SPR affinity detection, DSF thermal stability, and SEC aggregation detection for 2100 recommended candidates. Researchers manually reviewed the CDR sequence diversity, removed 120 duplicate clones, and finally completed 1980 wet experiments and generated new experimental tags.

[0040] The S7 experimental data backflow iteration incorporates 1980 newly added wet experimental labels into the training dataset, backflows them to the S4 module, updates the multi-objective reward function, and fine-tunes the ABPA model; after iteration, the test set AUC improves to 0.79.

[0041] S8 outputs a standardized candidate list, which includes the predicted affinity, expression level, aggregation risk, confidence score, experimental priority, current ABPA model version, and training AUC metric for each scFv, and delivers it to researchers for lead molecule optimization.

[0042] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting and enhancing the discovery of candidate molecular properties in biological macromolecules based on an adaptive cross-attention model of molecular properties, characterized in that, Includes the following steps: S1. Input candidate binding molecule data, which includes the amino acid sequence of the binding molecule, antigen / target sequence, structural features, experimental tags, and project research and development constraints. S2. Perform data preprocessing and region labeling on candidate binding molecule data: Clean and deduplicate sequences, remove illegal amino acids and invalid records, and classify and label the main chain, light chain, CDR region, framework region, linker peptide, and functional domain based on the protein numbering system. Standardize multi-source experimental labels into binary labels, regression labels, ordination labels, or weak labels with confidence, while encoding species, experimental platform, and target family contextual features. S3. Perform supervised fine-tuning of the protein language model (PLM) using publicly available protein datasets. Input the labeled sequences into the pre-trained protein language model to generate residue-level, CDR region-level, and full-sequence multi-level embedding representations. Adaptively select fine-tuning strategies according to the project sample size. For small samples, freeze the PLM backbone and train only the adapter layer, low-rank adapter module, or prediction head. For large samples, fine-tune some or all parameters of the PLM layer by layer. S4. Construct an ABPA adaptive molecular property cross-attention model, and optimize the ABPA model by constructing a multi-objective reward function based on the project's internal experimental data. The ABPA adaptive molecular property cross-attention model takes a multi-level PLM embedding representation as input, and fuses the interaction features of heavy chain, light chain, each CDR region, combined molecule and antigen, sequence and experimental context through a multi-head cross-attention layer. It dynamically and adaptively adjusts the attention weights of each feature region for different target molecular properties, and outputs multiple molecular property prediction values, prediction confidence and uncertainty scores simultaneously for multiple tasks. S5. Apply the optimized ABPA model to candidate molecule pruning prediction. Configure dynamic screening thresholds based on prediction scores, prediction uncertainties, and false positive / false negative R&D costs. Differentiate between early high-throughput screening and later pilot optimization scenarios to stratify candidate molecules, screen out low-quality candidate molecules, and complete virtual pre-screening. S6. Submit the priority candidate molecules after stratified screening to wet experimental verification, perform protein construction and assembly, binding detection, cell function, stability, expression development experiments, and support researchers to manually review and adjust the candidate set based on sequence diversity and epitope information; S7. Collect the new experimental label data produced by the wet experiment and feed it back to the dataset. Input the newly added internal experimental data into S4 and reconstruct the reward function to achieve iterative enhancement of the ABPA model. S8. Output a standardized candidate list, which includes multi-dimensional predicted properties, confidence level, experimental priority, model version, and training metric information for each molecule.

2. The method according to claim 1, characterized in that, The candidate biomolecules for binding include any one or more of the following: antibodies, nanobodies (VHH), Fab, scFv, DARPin, TCR, CAR binding domains, T cell binders (TCE), bispecific / multispecific engineered proteins, and receptor ligand proteins. The protein numbering system can be any one of IMGT, Kabat, Chothia, or AHo; for non-antibody binding molecules, sequence regions are segmented and labeled according to loops, binding motifs, and conserved domains.

3. The method according to claim 1, characterized in that, The target molecule properties include binding affinity, binding efficiency, specificity, cross-reactivity, cell activation / inhibition activity, expression level, solubility, thermal stability, aggregation tendency, immunogenicity risk, serum stability, endocytosis capacity, manufacturability, or a comprehensive candidate quality score obtained by weighting multiple properties; the same ABPA model framework can achieve parallel prediction of different molecular properties by replacing training labels, model output heads, and loss functions.

4. The method according to claim 1, characterized in that, The pre-trained protein language models include ESM, ESM2, ESM3, ProtBERT, ProtT5, TAPE, IgLM, antibody-specific protein language models, and structure-aware protein large models; the adapter layer and low-rank adapter module are at least one of Adapter and LoRA.

5. The method according to claim 1, characterized in that, The multi-head cross-attention can be replaced by collaborative co-attention (Co-Attention), graph attention network, hybrid expert MoE fusion network, or multi-task dual-tower fusion attention network; for affinity prediction tasks, the weights of CDR and antigen interaction features are increased; for expression and stability prediction tasks, the weights of frame region and hydrophobic sequence features are increased; and for TCE and CAR function prediction tasks, the context weights of the construct domain are fused.

6. The method according to claim 1, characterized in that, The dynamic screening threshold configuration logic in S5 is as follows: in the early high-throughput screening scenario, a loose threshold is used to improve the recall rate and reduce the missed detection of high-quality molecules; in the later lead optimization scenario, a strict threshold is used to improve the precision rate and reduce the number of invalid wet experiments; candidate molecules are divided into four categories: recommended experimental verification, retained for observation, sequence redesign, and direct elimination.

7. The method according to claim 1, characterized in that, The multi-objective reward function described in S4 simultaneously constrains affinity enhancement, specificity enhancement, aggregation risk reduction, expression level enhancement, sequence diversity preservation, and undesirable motif avoidance. It uses mutation enumeration, genetic algorithm, and Bayesian optimization to generate the next generation of candidate variants, and completes rapid pre-evaluation of variants through the ABPA model to achieve SAR iterative optimization.

8. The method according to claim 1, characterized in that, It also includes MLOps full-process management steps: using at least one of the following tools, SageMaker, MLflow, Kubeflow, Weights & Biases, and Databricks ML, to record the original sequence dataset, PLM and ABPA model parameters, training metrics, virtual screening records, and raw wet experiment data, to achieve full-link traceability and reproducibility of model training, inference, and iteration; the system deployment environment is a public cloud, private cloud, local HPC, or hybrid computing cluster.

9. The method according to claim 1, characterized in that, The publicly available protein dataset mentioned in S3 includes the SAbDab antibody structure library and provides publicly available scFv, Fab, and VHH sequence data for general protein characterization pre-training; S4 uses project-specific internal wet experimental data to complete the ABPA model reinforcement learning optimization, with publicly available data and internal data working together to complete the model construction.

10. A system for predicting and enhancing the discovery of candidate molecular properties of biological macromolecules based on an adaptive cross-attention model of molecular properties, characterized in that, The method for executing any one of claims 1 to 9 comprises: a data access and standardization module, a sequence preparation and region labeling module, a PLM fine-tuning module, an ABPA adaptive cross-attention prediction and reinforcement learning optimization module, a candidate ranking and pruning decision module, a virtual screening and wet experiment transfer module, an experimental data feedback iteration module, and an MLOps cloud management module; wherein the ABPA adaptive cross-attention prediction and reinforcement learning optimization module is the core computing unit, used for multi-level sequence embedding feature adaptive weighted fusion, multi-molecule property synchronous prediction, and model iterative optimization based on the internal experimental reward function.

Images