A training data construction method and system for a protein pocket-small molecule contrast learning model

Abstract

The application discloses a kind of protein pocket-small molecule contrast learning model training data construction method and system, it is related to training dataset construction technical field, including: obtaining protein-small molecule activity data, existing structure coverage data and reference data, respectively construct active pairing set, structure coverage pairing set and reference exclusion set, and then filter out candidate pairing with active record but lack of structure coverage;Candidate pairing and reference exclusion set are compared and filtered in turn protein homology and small molecule overlap;Supplement complex structure information to the candidate pairing filtered, construct protein pocket feature and small molecule structure feature accordingly and organize as positive pairing training sample, write in training dataset.The application makes the pairing with active evidence but lack of structure coverage can be organized as the positive pairing training sample required by contrast learning model, and completes homology leak control, ligand overlap control and quality check before writing in dataset.

Description

Technical Field

[0001] This invention belongs to the technical fields of computer-aided drug design, bioinformatics, artificial intelligence-assisted drug development data processing and training dataset construction, and particularly relates to a method and system for constructing training data for protein pocket-small molecule contrastive learning models. Background Technology

[0002] Virtual screening is an important technique in early drug discovery, aiming to identify small molecules that may bind to specific protein targets from a large library of candidate compounds. With the development of deep learning and 3D structural characterization methods, a class of protein-small molecule contrastive learning models has been increasingly used in virtual screening tasks. These models typically encode the protein binding pocket and the small molecule structure separately, learning their matching relationship in the representation space through positively paired samples, thus achieving rapid ranking of candidate small molecules.

[0003] The training performance of such models relies on high-quality pocket-liquid pairing training data. Existing structural data largely comes from experimentally resolved complex structures or compiled pocket-ligand pairing data, offering high reliability, but their scale, target coverage, and chemical spatial coverage remain limited. Meanwhile, while a large number of protein-small molecule activity records have been accumulated in publicly available activity data, most records do not directly provide the corresponding three-dimensional complex structures or pocket information that can be directly used for model training.

[0004] In existing technologies, some solutions focus on improving the complex structure prediction, molecular docking, or virtual screening model itself, while others involve extracting pockets from the complex structure and generating predicted conformations. However, none of these solutions systematically address the connection between activity data and pocket-small molecule pairing training data, which are required for the training data of the contrastive learning model.

[0005] Furthermore, if newly added training pairs have high protein sequence homology with existing evaluation data, or if small molecule ligands completely or highly overlap, it may cause data leakage, leading to an overestimation of model performance. Therefore, it is necessary to control the homology and ligand overlap of candidate pairs when constructing training data.

[0006] The aforementioned prior art has at least the following problems: (1) The existing training data for pocket-small molecule pairing is limited in scale and insufficient in coverage; (2) There is a data format gap between the live data and the positively paired samples required for contrastive learning; (3) After generating the predicted structure of the complex, field alignment, pairing relationship organization and quality control are still required; (4) Lack of homology and ligand overlap control may affect the objectivity of the assessment; (5) The data from multiple sources differ in terms of identification and format, resulting in poor compatibility with existing training processes.

[0007] Therefore, it is necessary to provide a training data construction method for protein pocket-small molecule contrastive learning models, which pairs proteins and small molecules with activity evidence but lacking structural coverage, and transforms them into datasets suitable for contrastive learning training through candidate identification, normalization, leakage control, structural completion, sample organization, and quality control. Summary of the Invention

[0008] To address the challenges of directly converting activity data into pocket-small molecule positive pairing training samples in existing technologies, the limited scale and coverage of existing structured training data, potential homology or ligand overlap leakage between candidate pairings and reference data, and insufficient compatibility between new samples and existing contrastive learning training processes, this invention provides a training data construction method and system for protein pocket-small molecule contrastive learning models. This method enables protein-small molecule pairings with activity evidence but lacking structural coverage to be organized into protein pocket-small molecule positive pairing training samples, and performs necessary homology leakage control, ligand overlap control, and quality checks before writing them into the training dataset.

[0009] This invention provides a method for constructing training data for a protein pocket-small molecule contrastive learning model, comprising: Acquire protein-small molecule activity data, existing structure coverage data, and reference training or evaluation data; An activity pairing set is constructed based on the protein-small molecule activity data, a structure coverage pairing set is constructed based on the existing structure coverage data, and a reference exclusion set is constructed based on the reference training data or reference evaluation data. Based on the active pairing set and the structure coverage pairing set, candidate pairs with active records but lacking structure coverage are selected; The candidate pairs are compared with the reference exclusion set for protein homology, and the candidate pairs are screened based on the results of the protein homology comparison. The candidate pairs after protein homology comparison are compared with the reference exclusion set for small molecule overlap, and the candidate pairs are screened based on the results of the small molecule overlap comparison. For candidate pairings after small molecule overlap comparison, supplement the complex structure information; Based on the complex structural information, protein pocket features and small molecule structural features are constructed; The protein pocket features and small molecule structure features of the same candidate pair are organized into positive pair training samples and written into the training dataset.

[0010] Optionally, based on the active pairing set and the structure-covering pairing set, candidate pairs with active records but lacking structure coverage are selected, including: Compare the active pairing set with the structure coverage pairing set; Protein-small molecule pairs that exist in the active pairing set but not in the structure-covering pairing set are selected as candidate pairs.

[0011] Optionally, before comparing the candidate pairs with the reference exclusion set for protein homology, the method further includes normalizing and quality screening the candidate pairs, wherein the normalization and quality screening include: The small molecule structure representations in the candidate pairings are normalized and a unique identifier for the small molecule is generated. The protein identifiers in the candidate pairings are sorted out, abnormal ligands are removed, activity screening is performed according to a preset activity threshold, the distribution ratio of target sources is controlled, and at least one of the following processing methods is used to remove duplicate records.

[0012] Optionally, the candidate pairs are compared with the reference exclusion set for protein homology, and the candidate pairs are screened based on the results of the protein homology comparison, including: The protein sequences in the candidate pairings are compared with the protein sequences in the reference exclusion set; Calculate sequence consistency index, sequence coverage index, structural similarity index, or functional domain similarity index based on the alignment results; Based on at least one of the indicators and its preset threshold, the candidate pairings are subjected to deletion, marking, or hierarchical processing.

[0013] Optionally, the candidate pairs after protein homology comparison are compared with the reference exclusion set for small molecule overlap, and the candidate pairs are screened based on the results of the small molecule overlap comparison, including: The small molecules in the candidate pairings after protein homology comparison are compared with the small molecules in the reference exclusion set. The comparison method adopts at least one of the following: normalized molecule representation comparison, molecule unique identifier comparison, molecule fingerprint similarity comparison, or molecule backbone similarity comparison. Based on the comparison results and preset overlap conditions, the candidate pairs are subjected to deletion, marking, or layering processes.

[0014] Optionally, for candidate pairings after small molecule overlap comparison, supplementary complex structural information is provided, including: For candidate pairings retained after small molecule overlap comparison, a structure prediction model or molecular docking program is invoked to generate the predicted structure of the protein-small molecule complex corresponding to the candidate pairing; or the existing experimental complex structure corresponding to the candidate pairing is obtained from the experimental structure database.

[0015] Optionally, protein pocket features and small molecule structural features are constructed based on the complex structural information, including: The spatial positions of the small molecules in the candidate pairings within the three-dimensional structure of the protein are determined based on the complex structure information. Centered on the aforementioned spatial location, protein atoms, amino acid residues, and corresponding coordinate information within a preset spatial range are selected as the protein pocket features; The atom types, atomic coordinates, and molecular connections of the small molecules are extracted as structural features of the small molecules.

[0016] Optionally, the protein pocket features and small molecule structure features of the same candidate pair are organized into positive pair training samples and written into the training dataset, including: The protein pocket features, small molecule structure features, pairing identification information, protein identification information, and small molecule identification information corresponding to the same candidate pair are aligned according to the input field format of the target contrastive learning model. The aligned information is combined into a positive pair training sample, which contains matching tags indicating a positive pairing relationship between the protein pocket and the small molecule. Write the positively paired training samples into the training dataset.

[0017] Optionally, after writing the positively paired training samples into the training dataset, the method further includes: The training dataset after writing is subjected to field integrity checks, sample readability checks, coordinate validity checks, abnormal record checks, structural scale distribution checks, and environmental compatibility checks. The tested training dataset is output in a storage format suitable for batch reading in deep learning.

[0018] On the other hand, this invention also proposes a training data construction system for protein pocket-small molecule contrastive learning models, used to implement the method, including: The data acquisition module is used to acquire protein-small molecule activity data, existing structure coverage data, and reference training data or reference evaluation data. The pairing set construction module is used to construct an activity pairing set based on the protein-small molecule activity data, construct a structure coverage pairing set based on the existing structure coverage data, and construct a reference exclusion set based on the reference training data or reference evaluation data. The candidate pairing identification module is used to filter out candidate pairs that have active records but lack structural coverage based on the active pairing set and the structural coverage pairing set; A homology control module is used to compare the candidate pairings with the reference exclusion set for protein homology, and to screen the candidate pairings based on the results of the protein homology comparison. An overlap control module is used to perform small molecule overlap comparison between the candidate pairs after protein homology comparison and the reference exclusion set, and to screen the candidate pairs based on the results of the small molecule overlap comparison. The structure completion module is used to supplement the complex structure information after candidate pairings have undergone small molecule overlap comparison; The sample construction module is used to construct protein pocket features and small molecule structure features based on the complex structure information, and to organize the protein pocket features and small molecule structure features with the same candidate pairing into positive pairing training samples. The dataset writing module is used to write the positively paired training samples into the training dataset.

[0019] Compared with the prior art, the present invention has the following advantages and technical effects: This invention addresses the problems in existing technologies, such as formal gaps between activity data and training data for contrastive learning models, insufficient coverage of structural training data, and the leakage of homology or ligand overlap due to newly added data. It provides a systematic data construction scheme. This method first acquires protein-small molecule activity data, existing structural coverage data, and reference training or evaluation data. Based on these, it constructs activity pairing sets, structural coverage pairing sets, and reference exclusion sets, respectively. Through set comparison, candidate pairings with activity records but lacking structural coverage are screened, thereby achieving effective identification and mining of potentially usable pairings from massive amounts of publicly available activity data, compensating for the deficiencies in target and chemical space coverage of existing structural data.

[0020] Building upon this foundation, this invention compares candidate pairings with the reference exclusion set for protein sequence homology and small molecule overlap. Based on the comparison results, candidate pairings are screened, avoiding information leakage between new training data and existing evaluation data due to sequence homology or molecular overlap, thus ensuring the objectivity and reliability of subsequent model performance evaluation. Subsequently, complex structure information is supplemented to the screened candidate pairings, and protein pocket features and small molecule structure features are constructed based on this information. Pocket features and molecular features of the same pair are organized into positive pairing training samples. This transforms the protein-small molecule relationships, which originally only contained activity records but lacked three-dimensional structural information, into standard positive pairing samples that can be directly used for training dual-encoder contrastive learning models, filling the gap between activity data and contrastive learning training data.

[0021] Furthermore, when generating positively paired training samples, this invention aligns information such as pairing identifiers, protein identifiers, small molecule identifiers, and matching labels with the input fields of the target contrastive learning model. This ensures that the constructed training data is compatible with existing training processes, avoiding integration difficulties caused by differences in identifiers and formats among multi-source data. The resulting training dataset can be widely applied in downstream tasks such as virtual screening, where candidate small molecules are matched, retrieved, and ranked given protein pockets. This provides contrastive learning models with a larger-scale, more comprehensive, and quality-controlled source of training data. Attached Figure Description

[0022] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the system structure according to an embodiment of the present invention. Detailed Implementation

[0023] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0024] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0025] Example 1 This invention provides a method for constructing training data for a protein pocket-small molecule contrastive learning model. Based on protein-small molecule activity data, existing structure coverage data, and reference training or evaluation data, this method first identifies candidate pairings with activity records but lacking structure coverage, then performs normalization, quality screening, homology leakage control, and ligand overlap control. Subsequently, it supplements the candidate pairings with complex structure information, extracts protein pocket features and small molecule features, and organizes them into positive pairing training samples suitable for a dual-encoder contrastive learning model. Figure 1 As shown, the method includes the following steps: S1. Data Acquisition. Acquire protein-small molecule activity data, existing structure coverage data, and reference training or evaluation data. The protein-small molecule activity data includes one or more of the following: experimental assay number, protein identifier, protein sequence, small molecule structure representation, activity value, and activity relation symbol; wherein, the small molecule structure representation may include SMILES (Simplified Molecular Input Line Entry System) or other equivalent representation methods.

[0026] S2. Constructing Pairing Sets. Using protein identifiers and normalized small molecule structural representations as basic units, construct activity pairing sets from the activity data; construct structure coverage pairing sets from existing complex structure data, pocket structure data, or other data that can characterize structure coverage relationships; and construct reference exclusion sets from existing training sets, validation sets, test sets, or benchmark data.

[0027] S3. Identify candidate missing structure pairs. Compare the active pair set and the structure coverage pair set, and screen out protein-small molecule pairs that exist in the active pair set but not in the structure coverage pair set as candidate missing structure pairs.

[0028] S4. Perform standardization and quality screening. The candidate missing structure pairs are subjected to one or more of the following treatments: small molecule representation standardization, aberrant ligand removal, small molecule unique identifier generation, protein identifier organization, activity threshold screening, target distribution control, and duplicate record removal.

[0029] S5. Perform homology leakage control and ligand overlap control. Compare the protein sequences in candidate pairs with those in the reference training or evaluation data for homology, and delete, label, or stratify candidate pairs that may cause leakage according to preset homology thresholds. Compare the small molecules in candidate pairs with those in the reference data for normalized labeling or structural similarity comparison, and delete, label, or stratify completely overlapping or highly similar small molecules according to preset rules. For candidate pairs that have passed, been deleted, labeled, or stratified, retain their homology level, ligand overlap level, or screening status information with the reference exclusion set as part of the training sample quality control information.

[0030] The homology leakage control includes: comparing the protein sequences in the candidate missing structure pair with the protein sequences in the reference exclusion set, and deleting, marking or stratifying the candidate pairs that meet the preset homology conditions based on one or more indicators such as sequence consistency, sequence coverage, structural similarity or functional domain similarity.

[0031] The ligand overlap control includes: comparing the normalized molecular representation, molecular unique identifier, molecular fingerprint similarity, or backbone similarity of small molecules in the candidate missing structure pair with small molecules in the reference exclusion set, and deleting, marking, or stratifying small molecules that are completely overlapping or highly similar according to preset overlap conditions.

[0032] S6. Supplementing Complex Structural Information. For candidate pairings after the above screening and control, generate or obtain protein-small molecule complex structural information. The complex structural information can be obtained from structural prediction models, docking procedures, experimental structures, or other data sources that can provide the relative spatial relationship between proteins and small molecules. This invention is not limited to a specific structure generation method.

[0033] S7. Construct protein pocket features and small molecule features. Based on the complex structure information, determine the local protein binding region around the small molecule to obtain pocket features such as protein pocket atoms, coordinates, and residues; and obtain small molecule atoms, coordinates, or other structural features based on the small molecule structure representation.

[0034] S8. Organize pocket-small molecule positive pairing training samples and write them into the training dataset. Organize the small molecule features and protein pocket features corresponding to the same candidate pair into a pocket-small molecule positive pairing training sample, and align it with the input field of the existing protein pocket-small molecule contrastive learning model; then write it into the training dataset, and perform field integrity, coordinate validity, sample readability, outlier records, structural scale distribution, and environmental compatibility checks.

[0035] The pocket-small molecule positive pairing training samples include one or more of the following: small molecule features, protein pocket features, pairing identifiers, protein identifiers, small molecule identifiers, and positive pairing relationship information used to represent the matching relationship between pockets and small molecules; the training samples are aligned with the input fields of the dual encoder contrastive learning model.

[0036] After writing the pocket-small molecule positive pair training samples into the training dataset, the method further includes quality control of the training dataset, which includes field integrity checks, sample readability checks, coordinate validity checks, abnormal record checks, structural scale distribution checks, and environmental compatibility checks; the training dataset that passes the checks is output in a storage format suitable for batch reading in deep learning.

[0037] In one alternative implementation, the homology leakage control may employ one or more indicators selected from protein sequence identity, sequence coverage, structural similarity, or functional domain similarity; the ligand overlap control may employ one or more indicators selected from normalized molecular representation, molecular unique identifier, molecular fingerprint similarity, or backbone similarity.

[0038] In one alternative implementation, the training dataset can be a key-value database, a file database, tabular data, tensor data, or other data formats suitable for batch reading in deep learning. This invention is not limited to a specific database format.

[0039] In one alternative implementation, the protein pocket-small molecule contrastive learning model can be a dual encoder model, in which one encoder encodes the protein binding pocket and the other encoder encodes the small molecule structure. The model learns the matching relationship between the two through positive pairing samples of pocket and small molecule.

[0040] On the other hand, in this embodiment, such as Figure 2 As shown, a training data construction system for a protein pocket-small molecule contrastive learning model is provided to implement the above-described method. The system includes a data acquisition module, a pairing set construction module, a candidate pairing identification module, a homology control module, an overlap control module, a structure completion module, a sample construction module, and a dataset writing module.

[0041] Each module sequentially transmits data and forms a processing flow: the data acquisition module acquires protein-small molecule activity data, existing structure coverage data, and reference training data or reference evaluation data; the pairing set construction module constructs an activity pairing set based on the protein-small molecule activity data, a structure coverage pairing set based on the existing structure coverage data, and a reference exclusion set based on the reference training data or reference evaluation data; the candidate pairing identification module filters out candidate pairs with activity records but lacking structure coverage based on the activity pairing set and the structure coverage pairing set; the homology control module performs a protein homology ratio between the candidate pairs and the reference exclusion set. The candidate pairs are screened based on the results of the protein homology comparison; the overlap control module performs small molecule overlap comparison between the candidate pairs after protein homology comparison and the reference exclusion set, and screens the candidate pairs based on the results of the small molecule overlap comparison; the structure completion module supplements the complex structure information of the candidate pairs after small molecule overlap comparison; the sample construction module constructs protein pocket features and small molecule structure features based on the complex structure information, and organizes the protein pocket features and small molecule structure features of the same candidate pair into positive pair training samples; the dataset writing module writes the positive pair training samples into the training dataset.

[0042] Compared with the prior art, the present invention has at least the following technical effects: By comparing activity pairing sets and structure coverage pairing sets, we can systematically identify candidate protein-small molecule pairings that have evidence of activity but lack structural coverage. This allows us to fully explore the available pairing information contained in publicly available activity data and existing experimental data, improve the utilization efficiency of existing activity data, and effectively make up for the shortcomings of existing structural data in terms of scale, target coverage, and chemical spatial coverage.

[0043] By supplementing the identified candidate pairings with complex structure information and constructing protein pocket features and small molecule features accordingly, the pocket features and molecular features of the same candidate pairing are organized into positive pairing training samples. This allows the protein-small molecule relationship, which originally only had activity records but lacked three-dimensional structural information, to be transformed into standard positive pairing samples suitable for training the dual encoder contrastive learning model, rather than just remaining at the level of complex structure prediction results. This fills the gap between activity data and contrastive learning training data.

[0044] By comparing candidate pairings with the reference exclusion set for protein homology and small molecule overlap, and then filtering candidate pairings based on the comparison results, the risk of information leakage caused by sequence homology or molecular overlap between new training data and existing training or evaluation data can be effectively reduced, ensuring the objectivity and reliability of subsequent model training and performance evaluation.

[0045] By normalizing candidate pairings to small molecule representations, removing anomalous ligands, controlling target distribution, and performing quality checks, the impact of anomalous molecules, duplicate records, missing fields, and structural scale anomalies on training data quality can be reduced, thereby improving the overall usability and robustness of the dataset.

[0046] Furthermore, when constructing positively paired training samples, this invention aligns information such as pairing identifiers, protein identifiers, small molecule identifiers, and matching tags with the input fields of the target contrastive learning model. This ensures that the constructed training data is compatible with existing training processes, avoiding integration difficulties caused by differences in identifiers and formats among multi-source data. The training data constructed by this invention can serve as a supplementary data source for downstream computational tasks, particularly useful for tasks in virtual screening scenarios involving matching, retrieving, and ranking candidate small molecules within a given protein pocket.

[0047] Example 2 This embodiment provides a specific implementation process for constructing training data for a protein pocket-small molecule contrastive learning model. The process will be described in detail below.

[0048] In this embodiment, protein-small molecule activity data, existing structure coverage data, and reference training data or reference evaluation data are acquired. An activity pairing set is constructed based on the protein-small molecule activity data, a structure coverage pairing set is constructed based on the existing structure coverage data, and a reference exclusion set is constructed based on the reference training data or reference evaluation data. The activity pairing set represents protein-small molecule relationships supported by existing activity evidence, the structure coverage pairing set represents protein-small molecule relationships covered by existing complex structure, pocket structure, or equivalent structure information, and the reference exclusion set is used for subsequent protein homology comparisons and small molecule overlap comparisons.

[0049] Based on the activity pairing set and the structure coverage pairing set, candidate pairings with activity records but lacking structure coverage are screened. These candidate pairings are then normalized and quality-screened, including at least one of the following methods: small molecule representation normalization, anomalous ligand removal, protein labeling refinement, activity threshold screening, and target distribution control, to reduce the impact of duplicate records, anomalous molecules, inconsistent labeling, or excessive target distribution bias on subsequent processing.

[0050] Candidate pairs that have undergone normalization and quality screening are compared with the reference exclusion set for protein homology. Based on the results of the protein homology comparison, the candidate pairs are deleted, labeled, or stratified. Candidate pairs that have undergone protein homology comparison are compared with the reference exclusion set for small molecule overlap. Based on the results of the small molecule overlap comparison, the candidate pairs are deleted, labeled, or stratified.

[0051] For candidate pairings retained after protein homology comparison and small molecule overlap comparison, complex structural information is supplemented. Based on the complex structural information, the spatial position of the small molecule in the three-dimensional protein structure of the candidate pairing is determined. Protein atoms, amino acid residues, and corresponding coordinates within a preset spatial range centered on the spatial position are selected as protein pocket features. The atom type, atomic coordinates, and molecular connectivity of the small molecule are extracted as small molecule structural features. The protein pocket features and small molecule structural features corresponding to the same candidate pairing are organized into a positive pairing training sample, and field alignment is performed with the input field of the target contrastive learning model. The positive pairing training samples are written into the training dataset, and the written training dataset undergoes field integrity checks, sample readability checks, coordinate validity checks, anomaly record checks, structural scale distribution checks, and environmental compatibility checks to obtain a structured dataset suitable for training the contrastive learning model.

[0052] In a specific application example, the protein pocket-small molecule contrastive learning model can employ DrugCLIP or a model with the same dual-encoder training logic. Such models typically include a protein pocket encoder and a small molecule encoder, which respectively receive the three-dimensional structural features of the protein pocket and the small molecule, and use a contrastive learning objective to bring real pocket-small molecule pairs closer together in the representation space.

[0053] The training data constructed in this embodiment does not change the DrugCLIP model structure, loss function, or encoder design. Instead, it organizes the protein pocket features, small molecule features, and pairing identifiers obtained from candidate pairings into pocket-small molecule positive pairing samples that can be read by the DrugCLIP training process.

[0054] DrugCLIP is merely an example of a downstream contrastive learning model of this invention and does not constitute a limitation on the scope of protection of this invention; this invention can also be applied to other protein pocket-small molecule dual encoder contrastive learning models. The core of this invention lies in the training data construction process, not in the network structure, parameters, or training loss form of a specific contrastive learning model.

[0055] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for constructing training data for a protein pocket-small molecule contrastive learning model, characterized in that, include: Acquire protein-small molecule activity data, existing structure coverage data, and reference training or evaluation data; An activity pairing set is constructed based on the protein-small molecule activity data, a structure coverage pairing set is constructed based on the existing structure coverage data, and a reference exclusion set is constructed based on the reference training data or reference evaluation data. Based on the active pairing set and the structure coverage pairing set, candidate pairs with active records but lacking structure coverage are selected; The candidate pairs are compared with the reference exclusion set for protein homology, and the candidate pairs are screened based on the results of the protein homology comparison. The candidate pairs after protein homology comparison are compared with the reference exclusion set for small molecule overlap, and the candidate pairs are screened based on the results of the small molecule overlap comparison. For candidate pairings after small molecule overlap comparison, supplement the complex structure information; Based on the complex structural information, protein pocket features and small molecule structural features are constructed; The protein pocket features and small molecule structure features of the same candidate pair are organized into positive pair training samples and written into the training dataset.

2. The method according to claim 1, characterized in that, Based on the active pairing set and the structure-covering pairing set, candidate pairs with active records but lacking structure coverage are selected, including: Compare the active pairing set with the structure coverage pairing set; Protein-small molecule pairs that exist in the active pairing set but not in the structure-covering pairing set are selected as candidate pairs.

3. The method according to claim 1, characterized in that, Before comparing the candidate pairs with the reference exclusion set for protein homology, the method further includes normalizing and quality screening the candidate pairs, wherein the normalization and quality screening include: The small molecule structure representations in the candidate pairings are normalized and a unique identifier for the small molecule is generated. The protein identifiers in the candidate pairings are sorted out, abnormal ligands are removed, activity screening is performed according to a preset activity threshold, the distribution ratio of target sources is controlled, and at least one of the following processing methods is used to remove duplicate records.

4. The method according to claim 1, characterized in that, The candidate pairs are compared with the reference exclusion set for protein homology, and the candidate pairs are screened based on the results of the protein homology comparison, including: The protein sequences in the candidate pairings are compared with the protein sequences in the reference exclusion set; Calculate sequence consistency index, sequence coverage index, structural similarity index, or functional domain similarity index based on the alignment results; Based on at least one of the indicators and its preset threshold, the candidate pairings are subjected to deletion, marking, or hierarchical processing.

5. The method according to claim 1, characterized in that, Candidate pairings after protein homology comparison are compared with the reference exclusion set using small molecule overlap comparison. The candidate pairings are then screened based on the results of this small molecule overlap comparison, including: The small molecules in the candidate pairings after protein homology comparison are compared with the small molecules in the reference exclusion set. The comparison method adopts at least one of the following: normalized molecule representation comparison, molecule unique identifier comparison, molecule fingerprint similarity comparison, or molecule backbone similarity comparison. Based on the comparison results and preset overlap conditions, the candidate pairs are subjected to deletion, marking, or layering processes.

6. The method according to claim 1, characterized in that, For candidate pairings after small molecule overlap comparison, supplement the complex structure information, including: For candidate pairings retained after small molecule overlap comparison, a structure prediction model or molecular docking program is invoked to generate the predicted structure of the protein-small molecule complex corresponding to the candidate pairing; or the existing experimental complex structure corresponding to the candidate pairing is obtained from the experimental structure database.

7. The method according to claim 1, characterized in that, Based on the complex structural information, protein pocket features and small molecule structural features are constructed, including: The spatial positions of the small molecules in the candidate pairings within the three-dimensional structure of the protein are determined based on the complex structure information. Centered on the aforementioned spatial location, protein atoms, amino acid residues, and corresponding coordinate information within a preset spatial range are selected as the protein pocket features; The atom types, atomic coordinates, and molecular connections of the small molecules are extracted as structural features of the small molecules.

8. The method according to claim 1, characterized in that, The protein pocket features and small molecule structure features of the same candidate pair are organized into positive pair training samples and written into the training dataset, including: The protein pocket features, small molecule structure features, pairing identification information, protein identification information, and small molecule identification information corresponding to the same candidate pair are aligned according to the input field format of the target contrastive learning model. The aligned information is combined into a positive pair training sample, which contains matching tags indicating a positive pairing relationship between the protein pocket and the small molecule. Write the positively paired training samples into the training dataset.

9. The method according to claim 1, characterized in that, After writing the positively paired training samples into the training dataset, the process also includes: The training dataset after writing is subjected to field integrity checks, sample readability checks, coordinate validity checks, abnormal record checks, structural scale distribution checks, and environmental compatibility checks. The tested training dataset is output in a storage format suitable for batch reading in deep learning.

10. A training data construction system for a protein pocket-small molecule contrastive learning model, characterized in that, For implementing the method according to any one of claims 1-9, comprising: The data acquisition module is used to acquire protein-small molecule activity data, existing structure coverage data, and reference training data or reference evaluation data. The pairing set construction module is used to construct an activity pairing set based on the protein-small molecule activity data, construct a structure coverage pairing set based on the existing structure coverage data, and construct a reference exclusion set based on the reference training data or reference evaluation data. The candidate pairing identification module is used to filter out candidate pairs that have active records but lack structural coverage based on the active pairing set and the structural coverage pairing set; A homology control module is used to compare the candidate pairings with the reference exclusion set for protein homology, and to screen the candidate pairings based on the results of the protein homology comparison. An overlap control module is used to perform small molecule overlap comparison between the candidate pairs after protein homology comparison and the reference exclusion set, and to screen the candidate pairs based on the results of the small molecule overlap comparison. The structure completion module is used to supplement the complex structure information after candidate pairings have undergone small molecule overlap comparison; The sample construction module is used to construct protein pocket features and small molecule structure features based on the complex structure information, and to organize the protein pocket features and small molecule structure features with the same candidate pairing into positive pairing training samples. The dataset writing module is used to write the positively paired training samples into the training dataset.

Images