A protein ligand binding prediction and screening method and system based on multi-source data fusion

Abstract

The application provides a protein ligand binding prediction and screening method and system based on multi-source data fusion, wherein the method comprises the following steps: S1, candidate ligand retrieval; S2, structure prediction input construction; S3, batch structure prediction; S4, binding site analysis; S5, comprehensive scoring and sorting. Through the design of a five-step automatic pipeline, the application realizes a complete process from the input of a microorganism name or a target protein name by a user to the output of the ranking of candidate ligands with multi-dimensional comprehensive scores. The application combines experimental data of a ligand activity database and structure confidence of a protein structure prediction model, and comprehensively evaluates the candidate ligands through a multi-dimensional scoring mechanism.

Description

Technical Field

[0001] This invention relates to the field of computer-aided drug design, specifically to a method and system for predicting and screening protein ligand binding based on multi-source data fusion, applicable to the discovery and prioritization of candidate drugs based on microbial targets. Background Technology

[0002] Drug discovery is a high-investment, high-risk, and long-term process. Traditional drug screening mainly relies on experimental methods, including high-throughput screening and virtual screening. With the development of omics data and computational biology, using computational methods to predict protein-ligand binding affinity has become an important tool for early drug discovery.

[0003] In the field of protein-ligand interaction research, there are two main types of data resources:

[0004] Ligand activity databases (such as BindingDB and ChEMBL) contain a wealth of experimentally validated ligand-target binding activity data, including key indicators such as inhibition constant Ki, half-maximal inhibitory concentration (IC50), and dissociation constant Kd. These data serve as an important starting point for drug screening.

[0005] Protein structure prediction tools (such as AlphaFold, RoseTTAFold, Protenix, etc.) can predict the three-dimensional structure of proteins based on amino acid sequences. Some tools can also predict the binding mode and confidence level of protein-ligand complexes at the same time.

[0006] Although both types of data resources have matured in their respective development, they still present the following significant problems in practical applications:

[0007] (1) Data silo effect

[0008] The ligand activity database and protein structure prediction tools operate independently. Ligand activity data is stored in a dedicated database system, while protein structure prediction is typically performed on a separate computing platform. Researchers need to manually export data from the database, convert the format, input it into the structure prediction tool, and parse the output. This process is cumbersome, error-prone, and makes it difficult to ensure data consistency.

[0009] (2) Single screening criteria

[0010] Traditional drug screening primarily ranks ligands based on experimentally determined activity values ​​(such as Ki and IC50). However, highly active ligands do not necessarily have favorable binding conformations and may exhibit off-target effects or pharmacokinetic issues. Relying solely on activity values ​​for screening may lead to a higher failure rate in subsequent experiments.

[0011] (3) Lack of end-to-end automated processes

[0012] Starting with a given microorganism or target name, and then obtaining a ranking of candidate ligands after multi-dimensional evaluation, there is currently a lack of complete automated solutions. Researchers need expertise in multiple fields, including database querying, structure prediction, and results analysis, to complete cross-platform analysis, which presents a high barrier to entry.

[0013] (4) The ranking of ligands lacks a comprehensive evaluation system.

[0014] Existing methods struggle to organically integrate experimental activity data with calculated structural confidence. The lack of a comprehensive scoring mechanism that simultaneously considers binding affinity, interface prediction quality, and binding site coverage hinders a full assessment of the drug potential of candidate ligands.

[0015] Furthermore, with the continuous emergence of drug-resistant pathogens, the need for research and development of novel antibacterial and antiviral drugs is becoming increasingly urgent. Rapidly identifying high-potential candidate drugs targeting specific pathogens is of significant strategic importance in responding to emerging infectious diseases. Current technologies are insufficient to meet the rapid response requirements from pathogen naming to candidate drug recommendation, necessitating an efficient and reliable automated screening method. Summary of the Invention

[0016] This invention provides a method and system for protein ligand binding prediction and screening based on multi-source data fusion, addressing the technical problems of data silos, single screening criteria, lack of end-to-end automation, and lack of comprehensive evaluation in ligand ranking in existing technologies. This invention designs a five-step automated pipeline to achieve a complete process from user input of microbial or target protein names to output of candidate ligand rankings based on multi-dimensional comprehensive scoring. The method integrates experimental data from ligand activity databases with the structural confidence of protein structure prediction models, and comprehensively evaluates candidate ligands through a multi-dimensional scoring mechanism.

[0017] The specific plan is as follows:

[0018] A method for predicting and screening protein ligand binding based on multi-source data fusion includes the following steps:

[0019] S1. Candidate Ligand Retrieval: Receives user-inputted query conditions, including microbial name (e.g., "Staphylococcus aureus") or target protein name (e.g., "PBP2a"); calls the ligand activity database API to perform a search, obtaining ligand-target binding activity records matching the query conditions, including the ligand's SMILES structural string, the target protein's amino acid sequence, experimental activity value (Ki, IC50, or Kd), and the source database record identifier; preprocesses and filters the search results, with filtering conditions including: non-empty ligand SMILES, non-empty target protein sequence with a length within a preset range, and activity value meeting a preset threshold; performs deduplication on the filtered results by target protein, for multiple records with the same target protein name, retaining only the top N ligands with the best activity values, and outputting a candidate ligand list;

[0020] S2. Construction of Structure Prediction Input: Read each record in the candidate ligand list and extract the ligand SMILES string and the target protein amino acid sequence; combine the SMILES and sequence into the standard input format of the structure prediction tool according to the predefined JSON schema, including the complex name field, the sequence object composed of protein chains, and the ligand object composed of ligand SMILES; organize multiple candidate ligands of the same target protein into the same input file and process them in batches.

[0021] S3. Batch Structure Prediction: The protein structure prediction model is invoked to perform three-dimensional structure prediction on all protein-ligand complexes in the input file. The prediction model adopts a deep learning-based protein-ligand complex structure prediction algorithm and outputs the predicted three-dimensional structure file and the structure confidence index file. After the prediction is completed, the output includes a three-dimensional structure file and a confidence index file. The confidence index includes pTM (overall prediction confidence), ipTM (interface prediction confidence, used to characterize the prediction quality of the protein-ligand binding interface) and pLDDT (local prediction confidence).

[0022] S4. Binding Site Analysis: Read the predicted 3D structure file and confidence index file; for each protein-ligand complex, extract the coordinates of ligand atoms and the coordinates of all protein residue atoms; calculate the Euclidean distance between ligand atoms and protein residues, and construct a distance matrix; identify contact residues based on a preset distance threshold, count the number of contact residues, and extract detailed information for each contact residue; extract the interface prediction confidence index ipTM value of the complex.

[0023] S5. Comprehensive scoring and ranking: A comprehensive score is calculated for each candidate ligand. The comprehensive score is obtained by weighted summation of indicators from multiple dimensions, including interface prediction confidence, activity normalization value, and contact residue coverage.

[0024] Further, in step S1, the screening priority of the activity value is as follows: inhibition constant Ki is less than 100 nM, half-maximal inhibitory concentration (IC50) is less than 1000 nM, and dissociation constant Kd is less than 500 nM; the length of the target protein sequence is within a preset range of 50 to 2000 amino acid residues; and N is 3 in the first N ligands.

[0025] Furthermore, in step S2, the number of complexes contained in each batch is set to no more than 10 complexes per batch based on computing resources.

[0026] Furthermore, in step S3, the structure prediction process uses preset model parameter configuration, and the model type is selected as a benchmark model that supports protein-ligand complex prediction; multiple random seeds are used to sample each complex multiple times; the three-dimensional structure file is in CIF format, and the structure confidence index file is in JSON format.

[0027] Further, in step S4: the preset distance threshold is 4.0 Å; the method for identifying contact residues is: if the distance between any atom in a protein residue and any ligand atom is less than the threshold, then the residue is marked as a contact residue; the detailed information of each contact residue includes chain identifier, residue name, residue number and nearest contact distance; the ipTM value reflects the prediction confidence of the protein-ligand binding interface, and the value ranges from 0 to 1, with a higher value indicating better interface prediction quality.

[0028] Furthermore, in step S5, the formula for calculating the comprehensive score is as follows:

[0029]

[0030] in, This represents the overall score of the candidate ligand, with a value ranging from 0 to 1; This represents the confidence level of the interface prediction, reflecting the structural confidence of the protein-ligand binding interface, with a value ranging from 0 to 1. This represents the activity normalization value, which converts the experimental activity of the ligand into a normalized score. This indicates the contact residue coverage, reflecting the degree to which the ligand covers the target protein binding site; , , The weighting coefficients for the three indicators mentioned above are respectively, satisfying... The weighting coefficients can be adjusted according to specific application scenarios;

[0031] Activity normalized value The calculation formula is:

[0032]

[0033] in, This indicates the negative logarithm of the ligand's activity. This is the suppression constant, expressed in moles per liter. and These are the lower and upper bounds of the activity normalization interval, corresponding to pKi values ​​of 5 and 9, respectively.

[0034] Contact residue coverage The calculation formula is:

[0035]

[0036] in, Indicates the number of residues in contact with the ligand; A baseline value representing the number of contact residues (e.g., set to 20) is used to standardize coverage. This indicates a function that takes the minimum value, ensuring that the coverage does not exceed 1.0.

[0037] The candidate ligands are sorted in descending order based on their comprehensive scores to generate a ligand ranking table. A complete analysis report is generated by combining contact residue information and structural confidence index. The report includes: target protein information, candidate ligand ranking table (including SMILES, activity value, structural confidence, and comprehensive score), list of key contact residues, and recommended candidate molecules.

[0038] A protein ligand binding prediction and screening system based on multi-source data fusion is used to implement the above method. It includes a data retrieval module, an input construction module, a prediction execution module, a result analysis module, and a report generation module. The modules transfer data to each other through standardized data files to form a complete automated pipeline.

[0039] The data retrieval module is used to interface with the ligand activity database API, receive the microbial name or target protein name input by the user, execute the retrieval query, filter and deduplicate the data according to preset rules, and output a candidate ligand list CSV file.

[0040] The input building module is used to read the candidate ligand list CSV file, format and organize the ligand SMILES and target protein sequences according to the JSON schema of the structure prediction tool, and generate a JSON file that meets the model input requirements;

[0041] The prediction execution module is used to call the protein structure prediction model, perform batch inference on all complexes in the input JSON file, and output a three-dimensional structure file and a confidence index file.

[0042] The results analysis module is used to parse the structure file and confidence file, extract the interface prediction confidence ipTM and contact residue information, and calculate the comprehensive score of each candidate ligand.

[0043] The report generation module is used to sort the analysis results in descending order of comprehensive score and generate a structured Markdown format analysis report.

[0044] The beneficial effects of this invention are as follows: Through modular design, five processing steps are connected in series to form a complete automated pipeline. Users only need to provide the name of the microorganism or the target protein to obtain a complete analysis report containing ligand ranking, structural confidence, and contact residue information. The standardized process ensures consistent analytical results for different researchers, significantly reducing the technical threshold for drug screening. Specific advantages are as follows:

[0045] (1) Traditional drug screening processes require researchers to possess expertise in multiple fields, such as database querying, structure prediction, and result analysis, and manual data transfer is required between each step. The one-click automated process implemented in this invention reduces the traditional manual analysis of several days to several hours, significantly improving the screening efficiency in the early stages of drug discovery.

[0046] (2) Traditional screening methods that rely solely on activity values ​​may miss moderately active candidate ligands with good binding conformations. The multi-dimensional comprehensive scoring mechanism of this invention can screen out candidate ligands that possess both activity and structural quality.

[0047] (3) The standardized process of this invention ensures consistent analytical results for each run, facilitating comparison and verification between different laboratories. The generated Markdown format report contains complete scoring details, a list of contact residues, SMILES structures, and other information, providing a clear decision-making basis for subsequent medicinal chemistry optimization.

[0048] (4) The technical solution of this invention is universal and applicable to the screening of candidate drugs for various pathogen targets. Only the search conditions of the ligand activity database need to be changed to screen drugs for different types of pathogens such as bacteria, viruses, and fungi. With the accumulation of more pathogenic target data, this invention is expected to play an important role in the fields of infectious disease prevention and control, and public health security. Attached Figure Description

[0049] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0050] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0051] As shown in the figure, this invention provides a method for predicting and screening protein ligand binding based on multi-source data fusion. This embodiment uses a coronavirus target as an example to verify the effectiveness of the method.

[0052] The user enters the following query criteria: the microorganism name is "Human * coronavirus", and searches the ligand activity database for candidate ligands targeting the target protein of this pathogen.

[0053] Step 1: Candidate Ligand Search

[0054] Call the ligand activity database API and execute the search query: GET / api / search?q=Human *coronavirus&limit=50

[0055] After obtaining the search results, filter them according to the following criteria:

[0056] Filter records where SMILES is empty;

[0057] Records containing empty target protein sequences or those whose length is not in the range of 50 to 2000 were filtered out.

[0058] Activity value screening: records with Ki < 100 nM are retained first, followed by IC50 < 1000 nM, and then Kd < 500 nM;

[0059] Duplicates are removed by target protein name, and only the top 3 ligands with the best activity values ​​are retained for each target protein.

[0060] Nine candidate ligands were ultimately selected, all targeting the protein "Replicase polyprotein 1ab". A CSV file containing the candidate ligand list was output, including fields such as record_id, organism, target_name, smiles, ki_nm, and ic50_nm.

[0061] Step 2: Construction of Structural Prediction Input

[0062] Read the candidate ligand list from the CSV file and extract the SMILES and target protein sequences for each ligand. Construct a JSON input file, with an example format as follows:

[0063] [{

[0064] "name": "*_CoV_Replicase_1ab_ligand_001",

[0065] "sequences": [

[0066] {"proteinChain": {"sequence": "MKSFFC... (target protein amino acid sequence)", "count": 1}},

[0067] {"ligand": {"ligand": "CC(C)C[C@@H](N=C(O)...(ligands SMILES)", "count": 1}} ]

[0069] }]

[0070] The nine candidate ligands were organized into three batches (three complexes per batch), generating three JSON input files.

[0071] Step 3: Batch Structure Prediction

[0072] The Protenix structural prediction model is invoked to perform inference on each batch of input files:

[0073] protenix pred -i . / input_batch_001.json -o . / structures / -n protenix_base_default_v1.0.0 --use_default_params -s 101 -e 3

[0074] The parameter -s 101 sets the random seed to 101, and -e 3 indicates that each complex is sampled 3 times.

[0075] After the prediction is completed, collect all output CIF structure files and JSON confidence files.

[0076] Step 4: Binding Site Analysis

[0077] For each output protein-ligand complex structure, the coordinates of ligand atoms and protein residue atoms are extracted. Contact residues are identified using a distance threshold of 4.0 Å.

[0078] Taking the top-ranked candidate ligand as an example, its contact residues include: HIS163, ALA173, THR175, LEU30, GLY174, LEU32, TRP31, HIS172, GLY29, MET17, ILE281, THR280, LEU167, ASN28, TYR37, LEU177, MET162, MET165, ASP33, ARG279, ASP176, VAL171, and LEU282, totaling 23 residues.

[0079] The confidence level for the interface prediction of the complex was extracted: ipTM = 0.532.

[0080] Step 5: Overall Scoring and Ranking

[0081] A comprehensive score was calculated for each of the nine candidate ligands. Weighting coefficients were set as follows: (Interface prediction confidence weight) (Activity normalized value weight) (Contact residue coverage weight).

[0082] Taking the top-ranked candidate ligand as an example, its parameter values ​​are:

[0083] -

[0084] - (Corresponding to Ki = 3.00 nM), calculate

[0085] - , ,calculate

[0086] Overall rating:

[0087]

[0088] The ligands were sorted in descending order of their overall scores, and an analysis report was generated. The overall scores of the top 9 ligands are shown in the table below:

[0089]

[0090] As verified by the above embodiments, the present invention is capable of:

[0091] (1) Successfully realized the end-to-end automated process from microbial name to comprehensive evaluation of candidate ligands;

[0092] (2) Integrate activity experimental data with structure prediction confidence to conduct multi-dimensional evaluation of candidate ligands;

[0093] (3) The candidate ligand with the highest comprehensive score (0.7771) was identified, and its SMILES is: CC(C)C[C@@H](N=C(O)[C@H](N=C(O)COc1cccc(N(C)C)c1)C(C)C)C(O)=N[C@H](C[C@H]1CCN=C1O)C(=O)c1nc2ccccc2s1. This ligand can be used as the starting point for subsequent drug optimization.

[0094] The technical means disclosed in this invention are not limited to those disclosed in the above embodiments, but also include technical solutions composed of any combination of the above technical features. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications are also considered within the scope of protection of this invention.

Claims

1. A method for predicting and screening protein ligand binding based on multi-source data fusion, characterized in that, Includes the following steps: S1, Candidate Ligand Search: Receives user-inputted query conditions, including microbial name or target protein name; Call the ligand activity database API to perform a search and obtain ligand-target binding activity records that match the query conditions, including the SMILES structure string of the ligand, the amino acid sequence of the target protein, the experimental activity value, and the source database record identifier; The search results are preprocessed and filtered. The filtering conditions include: non-empty ligand SMILES, non-empty target protein sequence with length within a preset range, and activity value meeting a preset threshold. The filtered results are deduplicated according to target protein. For multiple records with the same target protein name, only the top N ligands with the best activity values ​​are retained, and a candidate ligand list is output. S2. Construction of Structure Prediction Input: Read each record in the candidate ligand list and extract the ligand SMILES string and the target protein amino acid sequence; combine the SMILES and sequence into the standard input format of the structure prediction tool according to the predefined JSON schema, including the complex name field, the sequence object composed of protein chains, and the ligand object composed of ligand SMILES; organize multiple candidate ligands of the same target protein into the same input file and process them in batches. S3. Batch Structure Prediction: The protein structure prediction model is invoked to perform three-dimensional structure prediction on all protein-ligand complexes in the input file. The prediction model adopts a deep learning-based protein-ligand complex structure prediction algorithm and outputs the predicted three-dimensional structure file and the structure confidence index file. After the prediction is completed, the output includes a three-dimensional structure file and a confidence index file. The confidence index includes pTM, ipTM and pLDDT. S4. Binding Site Analysis: Read the predicted 3D structure file and confidence index file; for each protein-ligand complex, extract the coordinates of ligand atoms and the coordinates of all protein residue atoms; calculate the Euclidean distance between ligand atoms and protein residues, and construct a distance matrix; identify contact residues based on a preset distance threshold, count the number of contact residues, and extract detailed information for each contact residue; extract the interface prediction confidence index ipTM value of the complex. S5. Comprehensive scoring and ranking: A comprehensive score is calculated for each candidate ligand. The comprehensive score is obtained by weighted summation of indicators from multiple dimensions, including interface prediction confidence, activity normalization value, and contact residue coverage.

2. The protein ligand binding prediction and screening method based on multi-source data fusion according to claim 1, characterized in that, In step S1, the screening priority of the activity value is as follows: inhibition constant Ki is less than 100 nM, half-maximal inhibitory concentration (IC50) is less than 1000 nM, and dissociation constant Kd is less than 500 nM; the length of the target protein sequence is within a preset range of 50 to 2000 amino acid residues; and N is 3 in the first N ligands.

3. The protein ligand binding prediction and screening method based on multi-source data fusion according to claim 1, characterized in that, In step S2, the number of complexes in each batch is set to no more than 10 complexes per batch based on computing resources.

4. The protein ligand binding prediction and screening method based on multi-source data fusion according to claim 1, characterized in that, In step S3, the structure prediction process uses preset model parameter configuration, and the model type is selected as a benchmark model that supports protein-ligand complex prediction; multiple random seeds are used to sample each complex multiple times. The three-dimensional structure file is in CIF format, and the structure confidence index file is in JSON format.

5. The protein ligand binding prediction and screening method based on multi-source data fusion according to claim 1, characterized in that, In step S4: the preset distance threshold is 4.0 Å; the method for identifying contact residues is: if the distance between any atom in a protein residue and any ligand atom is less than the threshold, then the residue is marked as a contact residue; the detailed information of each contact residue includes chain identifier, residue name, residue number and nearest contact distance; the ipTM value ranges from 0 to 1, and the higher the value, the better the interface prediction quality.

6. The protein ligand binding prediction and screening method based on multi-source data fusion according to claim 1, characterized in that, In step S5, the formula for calculating the comprehensive score is as follows: in, This represents the overall score of the candidate ligand, with a value ranging from 0 to 1; This represents the confidence level of the interface prediction, reflecting the structural confidence of the protein-ligand binding interface, with a value ranging from 0 to 1. This represents the activity normalization value, which converts the experimental activity of the ligand into a normalized score. This indicates the contact residue coverage, reflecting the degree to which the ligand covers the target protein binding site; , , The weighting coefficients for the three indicators mentioned above are respectively, satisfying... ; Activity normalized value The calculation formula is: in, This indicates the negative logarithm of the ligand's activity. This is the suppression constant, expressed in moles per liter. and These are the lower and upper bounds of the activity normalization interval, corresponding to pKi values ​​of 5 and 9, respectively. Contact residue coverage The calculation formula is: in, Indicates the number of residues in contact with the ligand; A baseline value representing the number of contact residues, used to standardize coverage; This indicates a function that takes the minimum value, ensuring that the coverage does not exceed 1.0; The candidate ligands are sorted in descending order based on their comprehensive scores to generate a ligand ranking table. A complete analysis report is generated by combining contact residue information and structural confidence index. The report includes: target protein information, candidate ligand ranking table, list of key contact residues, and recommended candidate molecules.

7. A protein ligand binding prediction and screening system based on multi-source data fusion, used to implement the method described in any one of claims 1-6, characterized in that, It includes a data retrieval module, an input construction module, a prediction execution module, a result analysis module, and a report generation module. The modules transfer data to each other through standardized data files, forming a complete automated pipeline. The data retrieval module is used to interface with the ligand activity database API, receive the microbial name or target protein name input by the user, execute the retrieval query, filter and deduplicate the data according to preset rules, and output a candidate ligand list CSV file. The input building module is used to read the candidate ligand list CSV file, format and organize the ligand SMILES and target protein sequences according to the JSON schema of the structure prediction tool, and generate a JSON file that meets the model input requirements; The prediction execution module is used to call the protein structure prediction model, perform batch inference on all complexes in the input JSON file, and output a three-dimensional structure file and a confidence index file. The results analysis module is used to parse the structure file and confidence file, extract the interface prediction confidence ipTM and contact residue information, and calculate the comprehensive score of each candidate ligand. The report generation module is used to sort the analysis results in descending order of comprehensive score and generate a structured Markdown format analysis report.

Images