A small molecule compound screening and molecular docking platform, method and storage medium for drug research and development

Abstract

The application discloses a virtual screening and molecular docking platform and method for drug structure optimization. The platform comprises a molecular data management module, a fingerprint similarity screening module, a deep learning fine screening module, a three-dimensional conformation generation and format conversion module, a molecular docking module and a result display module. The application adopts a screening strategy of Morgan fingerprint and deep learning model based on Transformer in series, first performs coarse screening through molecular fingerprint, and then performs fine screening through a deep learning model to calculate the semantic similarity between candidate molecules and reference molecules. After screening, the molecules are subjected to three-dimensional conformation generation and format conversion, and then subjected to molecular docking calculation of binding energy with target proteins. The application also provides a unified system parameter configuration mechanism to support flexible adjustment of parameters in each stage. The application improves the automation degree, screening accuracy and result visualization degree of the virtual screening process, and reduces the technical use threshold in the field of drug structure optimization.

Description

Technical Field

[0001] This invention relates to the field of drug development technology, specifically to a small molecule compound screening and molecular docking platform, method, and storage medium for drug development. Background Technology

[0002] Traditional drug development is extremely costly in terms of both financial and human resources. However, with the advancement of computer technology, computer-aided drug design can significantly save time and costs. Virtual screening methods are increasingly being applied in research areas such as small molecule drug discovery and molecular probe optimization. However, current virtual drug screening processes are complex, fragmented, lacking continuity, and inconvenient for parameter adjustments. Market platforms often focus on developing novel drugs from scratch, rather than based on existing structures, relying entirely on database selection. This results in impressive computer-generated data, but poor performance in real-world applications.

[0003] Therefore, existing virtual screening processes are fragmented, lack operational continuity, and are inconvenient for parameter adjustment. Furthermore, most platforms are not optimized based on existing structures, resulting in poor practical application of screening results. Therefore, it is necessary to design a unified system that integrates molecular data processing, similarity screening, deep learning-based fine screening, 3D conformation generation, format conversion, molecular docking, and result visualization for candidate compound screening. This will improve the automation, visualization, and efficiency of result management within the screening process. Summary of the Invention

[0004] The purpose of this invention is to provide a small molecule compound screening and molecular docking platform, method and storage medium for drug development, so as to solve the problems mentioned in the background art that the previous virtual drug screening process is relatively complicated, not only is the process scattered, the operation is not continuous, and the parameter adjustment is not convenient.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A virtual screening and molecular docking platform for drug structure optimization includes: The molecular data management module is used to store and manage the structural information of candidate small molecule compounds; The fingerprint similarity screening module is used to calculate the molecular fingerprint of candidate molecules based on reference molecules for the first-level coarse screening. The deep learning fine screening module is used to calculate the structural or semantic similarity between candidate molecules and reference molecules through a deep learning model, and to perform a second-level fine screening. The 3D conformation generation and format conversion module is used to generate the 3D structure of the candidate molecules after screening and convert it into the format required for molecular docking. The molecular docking module is used to perform docking calculations between candidate molecules and target proteins and output the binding energy. The results display module is used to visualize the screening results at each level and the molecular docking results.

[0007] As a further preferred technical solution, the reference molecule is F0502B, which is used for the structural optimization of the early identification probe for Parkinson's disease.

[0008] As a further preferred technical solution, the fingerprint similarity screening module and the deep learning fine screening module are connected in series for screening, and the deep learning fine screening module further screens based on the fingerprint screening results.

[0009] As a further preferred technical solution, a unified system parameter configuration mechanism is also included. Users can adjust data loading, fingerprint screening thresholds, deep learning model parameters and molecular docking parameters in the system configuration interface. After the configuration is saved, it is automatically synchronized to the background configuration file.

[0010] As a further preferred technical solution, the molecular fingerprint is the Morgan fingerprint, which is generated by iteratively expanding each non-hydrogen atom in the molecule, mapping the substructures within its radius range to integer identifiers, and using hash encoding to map the variable-length features to a fixed-length bit vector, which is used to measure the structural similarity between the candidate molecule and the reference molecule.

[0011] A virtual screening and molecular docking method based on the aforementioned platform includes the following steps: Step S1: Input check; Step S2: Load the candidate molecule database, read the structural information of the molecules and perform preprocessing; Step S3: Data cleaning; Step S4: Based on the reference molecule, calculate the molecular fingerprint of the candidate molecules for coarse screening; Step S5: Calculate the similarity between candidate molecules and reference molecules using a deep learning model to refine the results from the previous step; Step S6: Perform structurally resolvable cleaning on the finely sieved molecules; Step S7: Generate the three-dimensional conformation of the candidate molecule and convert it into the docking standard format; Step S8: Perform molecular docking between the candidate molecule and the target protein, and calculate the binding energy; Step S9: Save and visualize the fingerprint screening, deep learning screening, and molecular docking results.

[0012] As a further preferred technical solution, the deep learning model is a molecular language model based on the Transformer architecture, which is used to encode SMILES sequences into contextual feature vectors of fixed dimensions; the similarity calculation adopts cosine similarity, which realizes semantic-level screening of molecular structure-activity relationship by comparing the vector angle between candidate molecules and reference molecules in the embedding space.

[0013] As a further preferred technical solution, the deep learning model adopts a contrastive pre-training strategy: in the pre-training stage, the semantic representation of molecular structure is learned by minimizing the embedding distance between positive sample pairs and maximizing the embedding distance between negative sample pairs; in the fine screening stage, the model parameters are fixed, only the cosine similarity between candidate molecules and reference molecules is calculated, and a threshold is set for screening.

[0014] As a further preferred technical solution, the three-dimensional conformation generation in step S7 adopts a two-stage method based on distance geometry and force field optimization: firstly, the distance geometry algorithm is used to generate initial three-dimensional coordinates from the molecular topology, and then the energy is minimized through the molecular force field to obtain a stable low-energy conformation; the format conversion step automatically allocates atom type, bond level information and partial charge to generate a standard molecular docking input file.

[0015] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned method for screening and molecular docking of small molecule compounds for drug development.

[0016] Compared with the prior art, the beneficial effects of the present invention are: (1) This solution integrates molecular data management, fingerprint screening, deep learning screening, three-dimensional conformation generation, format conversion, molecular docking and result visualization into the same platform, realizing full automation of the virtual screening process, reducing manual intervention and improving screening efficiency; (2) A screening strategy combining Morgan fingerprint and Transformer-based deep learning model was adopted. First, irrelevant molecules were quickly filtered out using structural fingerprint, and then the false positive rate was reduced by semantic similarity assessment, thereby improving the screening accuracy and biological relevance. (3) Provides a unified parameter configuration interface, supporting flexible adjustment of data loading, filtering thresholds, model parameters and docking parameters, automatic configuration synchronization, reducing dependence on code modification, and improving system adaptability and ease of use. The three-dimensional conformation generation adopts a two-stage method of distance geometry and force field optimization, automatically assigns atom type and charge, and outputs standardized docking files to ensure the repeatability and stability of docking calculations; (4) The visualization of multi-dimensional results facilitates comparison and traceability, reduces the technical threshold for drug structure optimization, and is conducive to its promotion in scientific research and R&D teams. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of a small molecule compound screening and molecular docking platform for drug development according to the present invention;

[0018] Figure 2 This is a flowchart of a method for screening and molecular docking of small molecule compounds for drug development according to the present invention;

[0019] Figure 3 This is a main flow diagram of a small molecule compound screening and molecular docking platform for drug development according to the present invention.

[0020] Figure 4 This is a similarity analysis interface diagram of a small molecule compound screening and molecular docking platform for drug development according to the present invention;

[0021] Figure 5 This is a diagram showing the docking results of a small molecule compound screening and molecular docking platform for drug development according to the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] This embodiment uses the structural optimization of the Parkinson's disease early identification probe F0502B as an application scenario, but the application of the present invention is not limited to this.

[0024] I. System Overall Architecture

[0025] See Figure 1 and Figures 3 to 5 This invention provides a virtual screening and molecular docking platform for drug structure optimization. The platform includes a molecular data management module, a fingerprint similarity screening module, a deep learning fine screening module, a 3D conformation generation and format conversion module, a molecular docking module, and a results display module. All modules work collaboratively through a unified configuration management mechanism. Users can adjust parameters at each stage in a graphical interface, and the configuration is automatically synchronized to the background configuration file after saving.

[0026] II. Specific Implementation Process

[0027] See Figure 2 The main steps for drug structure optimization using the aforementioned virtual screening and molecular docking platform for drug structure optimization include:

[0028] Step 1: Input Check

[0029] The system first checks the user-input reference molecule (F0502B molecule in this example) and candidate molecule library to confirm that the SMILES expression format is correct, the atom types are valid, and there are no unclosed rings or abnormal valence states. If the input is abnormal, the system prompts the user to correct it.

[0030] Step 2: Database Loading and Preprocessing

[0031] The system loads the candidate molecule database and reads the molecule's identifier and SMILES structure information. Preprocessing includes standardizing atomic representations (e.g., unifying aromatic nitrogen atoms to specific symbols), normalizing charges, and removing salt ions or solvent molecules.

[0032] Step 3: Data Cleaning

[0033] Molecules that do not contain common drug atoms such as carbon, hydrogen, oxygen, nitrogen, phosphorus, and sulfur, molecules with molecular weights that are too small or too large (e.g., less than 100 Da or greater than 800 Da), and molecules containing heavy metals or non-synthetic fragments are excluded.

[0034] Step 4: Morgan fingerprint screening

[0035] The system uses molecule F0502B as a reference to calculate the Morgan molecular fingerprint of candidate molecules. This fingerprint is generated by iteratively expanding each non-hydrogen atom in the molecule, mapping substructures within a radius of two chemical bonds to integer identifiers, and using hash encoding to map variable-length features to a fixed-length bit vector (2048 bits in this embodiment). Molecules with a similarity greater than a first threshold (default 0.4, adjustable) are selected by calculating the Tanimoto coefficient between the candidate and reference molecule fingerprints.

[0036] Step 5: Deep Learning Fine Screening

[0037] The molecules selected in the previous step are input into a molecular language model based on the Transformer architecture (ChemBERTa is used in this embodiment, but it is not limited to this). The model encodes the SMILES sequence of each molecule into a contextual feature vector of fixed dimensions (e.g., 768 dimensions), and then calculates the cosine similarity between the candidate molecule and the F0502B molecule. The system selects molecules whose cosine similarity is greater than a second threshold (default 0.7, adjustable).

[0038] This deep learning model employs a contrastive pre-training strategy: during the pre-training phase, it learns the semantic representation of molecular structures by minimizing the embedding distance between structurally similar molecules (positive sample pairs) and maximizing the embedding distance between structurally different molecules (negative sample pairs). In the actual fine-tuning phase, the model parameters are fixed, and only forward inference computation is performed.

[0039] Step 6: Second cleaning

[0040] The molecules selected by deep learning are cleaned for structural resolvability. Cheminformatics tools such as RDKit are used to check whether SMILES can be successfully resolved into molecular objects and generate three-dimensional conformations. Molecules that fail to resolve are removed.

[0041] Step 7: 3D Image Generation and Format Conversion

[0042] A two-stage method is used to generate the three-dimensional conformation of candidate molecules. First, an initial three-dimensional coordinate system is generated from the molecular topology using a distance geometry algorithm. Then, energy minimization optimization is performed using a molecular force field (MMFF94 force field in this embodiment) to obtain a stable low-energy conformation. Subsequently, atom types, bond order information, and partial charges are automatically assigned and converted into a standard molecular docking input file (PDBQT format in this embodiment, but not limited to this).

[0043] Step 8: Molecular docking

[0044] The candidate molecules obtained in the previous step are then molecularly docked with the target protein (in this example, a Parkinson's disease-related target protein). A docking engine based on simulated annealing or Lamarck genetic algorithms (such as AutoDockVina, but not limited to this) is used to calculate the binding energy (unit: kcal / mol) of each candidate molecule. The position and size of the docking box are pre-set according to the active pocket of the target protein.

[0045] Step 9: Saving and Visualizing Results The system will save and visualize the following results: Morgan fingerprint screening results: similarity distribution is displayed as a bar chart; Deep learning screening results: The cosine similarity matrix is ​​displayed as a heatmap; Molecular docking results: Binding energy ranking is displayed in tabular form, and a 3D structure overlay view is supported.

[0046] III. Parameter Configuration Mechanism

[0047] The system provides a unified configuration interface, allowing users to adjust the following parameters: Data loading: Candidate molecule library path, output path; Fingerprint filtering: fingerprint radius (default 2), bit vector length (default 2048), similarity threshold; Deep learning models: model path, batch size, similarity threshold; Molecular docking: docking engine path, docking box coordinates and dimensions, docking accuracy (exhaustiveness default 8).

[0048] After the parameters are saved, they are automatically written to the backend configuration file (such as JSON or YAML format), and each module reads the configuration when it starts.

[0049] Using molecule F0502B as a reference, a candidate library containing 100,000 small molecules was screened, ultimately yielding 47 candidate molecules with binding energies below -9.0 kcal / mol. Compared to the traditional one-to-one docking method, the screening time was reduced from approximately 72 hours to 6 hours, and the hit rate was increased from approximately 0.3% to 1.2%.

[0050] Those skilled in the art will understand that the reference molecule is not limited to F0502B and can be replaced with other drug molecules or probes requiring structural optimization. The deep learning model can employ other pre-trained molecular models (such as MolBERT or GPT-Mol). The docking engine can also be replaced with other mainstream molecular docking software. All these substitutions fall within the scope of this invention.

[0051] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A virtual screening and molecular docking platform for drug structure optimization, characterized in that, include: The molecular data management module is used to store and manage the structural information of candidate small molecule compounds; The fingerprint similarity screening module is used to calculate the molecular fingerprint of candidate molecules based on reference molecules for the first-level coarse screening. The deep learning fine screening module is used to calculate the structural or semantic similarity between candidate molecules and reference molecules through a deep learning model, and to perform a second-level fine screening. The 3D conformation generation and format conversion module is used to generate the 3D structure of the candidate molecules after screening and convert it into the format required for molecular docking. The molecular docking module is used to perform docking calculations between candidate molecules and target proteins and output the binding energy. The results display module is used to visualize the screening results at each level and the molecular docking results.

2. The platform according to claim 1, characterized in that, The reference molecule is F0502B, which is used for the structural optimization of early identification probes for Parkinson's disease.

3. The platform according to claim 1, characterized in that, The fingerprint similarity screening module and the deep learning fine screening module are connected in series. The deep learning fine screening module further filters the fingerprint screening results.

4. The platform according to claim 1, characterized in that, It also includes a unified system parameter configuration mechanism, where users can adjust data loading, fingerprint screening thresholds, deep learning model parameters, and molecular docking parameters in the system configuration interface. After the configuration is saved, it is automatically synchronized to the background configuration file.

5. The platform according to claim 1, characterized in that, The molecular fingerprint is the Morgan fingerprint, which is generated by iteratively expanding each non-hydrogen atom in the molecule, mapping the substructures within its radius to integer identifiers, and using hash encoding to map variable-length features to fixed-length bit vectors to measure the structural similarity between candidate molecules and reference molecules.

6. A virtual screening and molecular docking method based on the platform described in any one of claims 1 to 5, characterized in that, Includes the following steps: Step S1: Input check; Step S2: Load the candidate molecule database, read the structural information of the molecules and perform preprocessing; Step S3: Data cleaning; Step S4: Based on the reference molecule, calculate the molecular fingerprint of the candidate molecules for coarse screening; Step S5: Calculate the similarity between candidate molecules and reference molecules using a deep learning model to refine the results from the previous step; Step S6: Perform structurally resolvable cleaning on the finely sieved molecules; Step S7: Generate the three-dimensional conformation of the candidate molecule and convert it into the docking standard format; Step S8: Perform molecular docking between the candidate molecule and the target protein, and calculate the binding energy; Step S9: Save and visualize the fingerprint screening, deep learning screening, and molecular docking results.

7. The method according to claim 6, characterized in that, The deep learning model is a molecular language model based on the Transformer architecture, used to encode SMILES sequences into fixed-dimensional contextual feature vectors; the similarity calculation adopts cosine similarity, which realizes semantic-level screening of molecular structure-activity relationship by comparing the vector angle between candidate molecules and reference molecules in the embedding space.

8. The method according to claim 7, characterized in that, The deep learning model employs a contrastive pre-training strategy: during the pre-training phase, it learns the semantic representation of molecular structures by minimizing the embedding distance between positive sample pairs and maximizing the embedding distance between negative sample pairs; during the fine screening phase, it fixes the model parameters, calculates only the cosine similarity between candidate molecules and reference molecules, and sets a threshold for screening.

9. The method according to claim 6, characterized in that, The three-dimensional conformation generation in step S7 adopts a two-stage method based on distance geometry and force field optimization: first, the distance geometry algorithm is used to generate initial three-dimensional coordinates from the molecular topology, and then the energy is minimized through molecular force field optimization to obtain a stable low-energy conformation; the format conversion step automatically allocates atom type, bond order information and partial charge to generate a standard molecular docking input file.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 6 to 9.

Images