Reverse design method applied to virus protein drug molecules by taking molecular spectrum as medium

Abstract

The invention relates to the technical field of molecular generation, and discloses a reverse design method using a molecular spectrum as a medium to be applied to virus protein drug molecules, which comprises the following steps: firstly, constructing a data set containing a molecular SMILES identifier and a corresponding molecular spectrum, and carrying out drug-likeness and false positive preliminary screening on compounds by adopting Lipinski five rules and a PAINS mode; then training a conditional diffusion model, and generating a target molecular spectrum by taking specific drug properties as conditions; the spectrum is mapped into a molecular SMILES structure through a Transform decoder, so that accurate conversion from the spectrum to the molecule is realized; finally, through conformation optimization and stability verification, candidate drug molecules meeting specific treatment requirements are output. The method effectively avoids the invalid structure problem caused by cross-dimension generation, significantly improves the synthesizability and patent medicine potential of generated molecules, and is suitable for efficient molecular design of antiviral drugs, inhibitors and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of molecular generation, and particularly relates to a reverse design method for virus protein drug molecules using molecular spectrum as an intermediary. BACKGROUND

[0002] With the continuous spread and variation of viral diseases, it has become a core challenge in the field of drug research and development to quickly design efficient antiviral drugs targeting specific viral protein targets. Traditional drug discovery methods rely heavily on large-scale experimental screening and chemical modification, which is time-consuming and laborious and has limited success rate. In recent years, as a new paradigm of computationally driven drug discovery, reverse molecular design has the potential to accelerate the development of antiviral drugs by inversely deducing molecular structures from target biological activities. In existing reverse design methods, the generation model plays a key role. The mainstream method is usually committed to directly generating two-dimensional or three-dimensional representations of molecules, such as generation based on SMILES strings, molecular graphs or three-dimensional coordinates. Existing studies attempt to control the physicochemical properties or binding affinity of generated molecules through a conditional diffusion process, to a certain extent, realizing the guidance of molecular properties. However, such methods generally face the problem of "dimensional gap", i.e. directly generating high-dimensional structure data from one-dimensional property conditions, resulting in high uncertainty in sampling, low generation efficiency, and a large number of generated molecules with unreasonable structure-activity relationship or difficulty in synthesis. Especially for viral protein targets, the spatial structure of the binding site is complex, and the shape complementarity, bonding mode and dynamic behavior of the ligand are demanding, further exacerbating the failure rate of existing generation methods. SUMMARY

[0003] The present application aims to provide a reverse design method for virus protein drug molecules using molecular spectrum as an intermediary, which uses molecular spectrum as a key bridging representation between properties and structures, effectively avoiding the uncertainty in direct generation of high-dimensional structures, and providing a reliable new approach for reverse design in the fields of drug molecules, functional materials, etc.

[0004] The above technical purpose of the present application is achieved by the following technical solution: a reverse design method for virus protein drug molecules using molecular spectrum as an intermediary, comprising the following steps: S1: Construct a data set containing antiviral molecule SMILES and corresponding molecular spectrum, calculate the molecular spectrum using quantum chemical calculation method to ensure the accuracy and physical consistency of the spectrum data; S2: Construct and train a conditional diffusion model, input the user-set molecular properties and molecular spectrum as conditions, the model learns to generate spectrum data that meets the conditions, and saves the model after training to loss convergence, which is used to generate candidate spectra according to the target properties; S3: Train a Transformer model of an encoder-decoder architecture, whose input is the spectrum data and output is the corresponding molecular SMILES sequence, which realizes accurate decoding from spectrum to molecular representation by capturing the mapping relationship between spectrum and molecular structure, and terminates training when the validation loss reaches the minimum and saves the optimal model; S4: Integrate the two trained models to build an end-to-end molecular generation process, given the target properties of the antiviral molecule, first generate a molecule spectrum that meets the conditions through the diffusion model, and then decode the spectrum corresponding to the SMILES expression through the Transformer model to obtain a candidate molecule with specified physical and chemical properties and reasonable structure; S5: Effectiveness screening and structure optimization of the generated SMILES, based on chemical rules and energy calculation, eliminate invalid or unstable structures, and perform conformation search and geometry optimization on the remaining molecules to finally output a stable molecule model that meets the user's needs.

[0005] The further setting of the application is that the S1 constructs a data set containing antiviral molecule SMILES and corresponding molecular spectrum, specifically including the following steps: S1.1: Molecule source and screening, obtain the initial SMILES data set of antiviral drug molecules from public compound databases such as PubChem, ChEMBL, ZINC, etc., in order to ensure the diversity and drug relevance of the data, the following criteria will be used for screening: Drug-likeness: preliminary filtering is performed using Lipinski's five rules, PAINS screening rules, etc.; Synthetic accessibility (SA): use tools such as RDKit to calculate SA Score, preferentially select molecules with SA Score<4.5 to ensure the synthesizability of candidate molecules; S1.2: Molecular spectrum calculation, using Gaussian 16 quantum chemistry calculation software package; using density functional theory (DFT), specifically using B3LYP functional and TZVP basis set for geometry optimization and frequency calculation to ensure the balance between calculation accuracy and efficiency; S1.3: Post-processing: standardize, broaden the peak width, and normalize the intensity of the original spectrum obtained by calculation to make the spectrum data closer to the spectrum form measured by experiment for model learning; S1.4: Data storage: build a structured database, each record contains molecular SMILES, optimized 3D conformation, and calculated spectrum data.

[0006] The further setting of the application is that the S2 constructs and trains the conditional diffusion model, specifically including the following steps: S2.1: A conditional diffusion model based on U-Net is adopted, and the encoder-decoder structure of U-Net can effectively capture the multi-scale features of the spectrum; S2.2: Targeted conditional injection: Convert the key antiviral property parameters into a condition vector (Condition Vector), and inject them into the model at multiple levels of U-Net through cross-attention (Cross-Attention) mechanism or feature concatenation (Feature Concatenation), to ensure that the generated spectrum meets multiple constraints while generating the spectrum; Further settings of the present application are that the S3 trains a Transformer model with an encoder-decoder architecture, specifically including the following model architecture: Encoder (Encoder): responsible for processing the spectrum sequence, first using a one-dimensional convolutional layer (1D CNN) to downsample and extract local features of the spectrum, and then inputting the feature sequence into the Transformer encoder layer to obtain the context-aware representation of the spectrum; Decoder (Decoder): an autoregressive (Autoregressive) Transformer decoder is adopted, which takes the spectrum representation output by the encoder as the context (Key / Value), and takes the already generated SMILES character sequence as the query (Query) to predict the next most likely character.

[0007] Further settings of the present application are that the S4 integrates the two trained models above to build an end-to-end molecule generation process, specifically including the following steps: S4.1: Targeted generation process: according to the property requirements of a specific viral target, the property conditions are converted into a condition vector, which is input into the trained conditional diffusion model, and the diffusion model generates a virtual spectrum that meets the conditions through iterative denoising. Then, the generated spectrum is input into the trained Transformer encoder-decoder model, and the decoder generates the corresponding SMILES string in a self-recurrent manner. S4.2: Batch processing and diversity: by setting different random seeds for the diffusion model, multiple different spectra can be generated for the same set of property conditions, thereby generating multiple candidate molecules with different structures but meeting the requirements, and increasing the diversity of the output.

[0008] Further settings of the present application are that the S5 performs effectiveness screening and structure optimization on the generated SMILES, specifically including the following steps: S5.1 initial screening, specifically including the following steps: S5.1.1: Chemical validity check: using RDKit to check if the generated SMILES can be successfully converted into a legal molecular object, immediately filtering out syntax invalid strings; S5.1.2: De-duplication: removing duplicate molecules based on the normalized SMILES (Canonical SMILES); S5.2 Conformation generation and optimization, specifically including the following steps: S5.2.1: Using the ETKDG method of RDKit to generate multiple initial three-dimensional conformations for each candidate molecule; S5.2.2: Using molecular force fields such as MMFF94 or UFF, energy minimization is performed on these conformations through molecular mechanics (Molecular Mechanics) optimization to find the most stable conformation with the lowest energy; S5.3: Final screening and output, specifically including the following steps: S5.3.1: For the optimized stable conformation, perform molecular docking simulation to evaluate the binding mode with the specific viral protein target, and use ADMET prediction tools to evaluate the in vivo pharmacokinetic properties of the generated molecules; S5.3.2: Further screening of candidate molecules according to expert experience, output format suitable for subsequent in vitro antiviral experimental verification of molecular structure information.

[0009] In summary, the present application has the following beneficial effects: The present application introduces molecular spectrum as a key intermediary between molecular "fingerprint" and properties, structure, effectively avoiding the "dimensional disaster" problem faced by traditional methods of generating three-dimensional molecular structure from one-dimensional properties, significantly improving the stability and reliability of the generation process; The high-dimensional structure generation task is decomposed into a sequential mapping process based on molecular spectrum, as the spectrum itself has a clear physical meaning and structure sensitivity, the molecules generated by this method not only conform to chemical rules, but also effectively reduce the appearance of invalid or difficult to synthesize structures, at the same time, the spectrum as an intermediate representation reduces the complexity of the generation space, significantly reduces the computational resource consumption and iteration time, making it possible to conduct targeted molecular exploration under limited computing power conditions; This method also has good scalability, can introduce other spectral data or property constraints, and is suitable for various practical scenarios such as drug lead compound optimization and functional molecule design. BRIEF DESCRIPTION OF DRAWINGS

[0010] Fig. 1 The flowchart of the present application based on the molecular spectrum intermediary applied to the reverse design method of drug molecules for viral proteins; Fig. 2 The schematic diagram of the present application training data preparation and preprocessing; Fig. 3A schematic diagram of the condition diffusion model structure and training of the present application; Fig. 4 A schematic diagram of the spectrum-to-SMILES Transformer model architecture of the present application; Fig. 5 A distribution diagram of the generated molecular property results of the present application; Fig. 6 A schematic diagram of the candidate drug molecules generated by the present application binding to viral proteins. DETAILED DESCRIPTION

[0011] The present application will be further described below with reference to the accompanying drawings of the embodiments of the present application.

[0012] Please refer to Figs. 1-6 In the embodiments of the present application, a reverse design method for viral protein drug molecules using molecular spectra as an intermediary includes the following steps: Step one: construction and preprocessing of high-quality antiviral molecule dataset Molecule source and screening: obtain the initial SMILES dataset of molecules with antiviral activity from public compound libraries such as PubChem, ChEMBL, ZINC, etc. To ensure the diversity and drug relevance of the data, the following criteria will be used for screening: 1) Drug-like: apply Lipinski's five rules, PAINS screening rules, etc. for preliminary filtering; 2) Synthetic accessibility (SA): use tools such as RDKit to calculate SA Score, and preferentially select molecules with SA Score <4.5 to ensure the synthesizability of candidate molecules; Molecular spectrum calculation: 1) Calculation software: use Gaussian 16 quantum chemistry calculation software package.

[0013] 2) Calculation method: use density functional theory (DFT), specifically use B3LYP functional and TZVP basis set for geometry optimization and frequency calculation to ensure the balance between calculation accuracy and efficiency; Post-processing: standardize, broaden the peak width (usually use Lorentzian or Gaussian function, half-width is set to 5-10 cm⁻¹) and intensity normalization of the original spectrum obtained by calculation to make the spectrum data closer to the spectrum form measured by experiment, which is convenient for model learning; Data storage: construct a structured database, each record contains: molecular SMILES, optimized 3D conformation, and calculated spectrum data; In this embodiment, preferably, in the construction of the training data set stage, strict preliminary screening of the initial compound library is the key first step to ensure the generation of molecules with drug potential; the Lipinski five rules (Rule of Five, Ro5) and PAINS (Pan-Assay Interference Compounds) screening rules are used for preliminary filtering to eliminate molecules with insufficient oral bioavailability potential and common false positive interference risk from the source; Lipinski five rule screening (property-based preliminary screening): this rule is used to quickly evaluate the drug-likeness of a compound and predict its potential to become an oral drug; the specific screening criteria are as follows (the molecule is required to violate no more than two rules at the same time): Molecular weight (MW): less than 500 Da LogP (logarithm of the octanol-water partition coefficient): less than 5 Number of hydrogen bond donors (HBD): less than 5 Number of hydrogen bond acceptors (HBA): less than 10 Number of rotatable bonds: as a supplementary reference, less than 10 Using chemical informatics toolkits such as RDKit, OpenBabel, etc., parse each molecule SMILES in the database and automatically calculate the above-mentioned physicochemical parameters, and only keep the molecules that meet the Ro5 standard for the next stage; PAINS screening (structure-based preliminary screening): PAINS filter aims to identify molecules containing "bad" substructures, which are chemically reactive, fluorescent, or colloidal aggregating, etc., and often exhibit non-specific interference activity in biological assays, leading to false positive results, which can significantly increase the cost and time of drug research and development; PAINS provides a series of substructure definitions based on SMARTS patterns, each pattern corresponds to a known interfering chemical structure fragment (such as enone, Michael acceptor, quinone, thiobarbituric acid, etc.); the structure of each molecule in the database is matched with the SMARTS pattern library of PAINS (Substructure Matching); if a molecule is detected to contain any one of the PAINS characteristic substructures, it is marked and excluded from the training set, and the FilterCatalog module of RDKit is used to load the PAINS pattern for batch processing.

[0014] Step two: specific training and application of conditional diffusion model The conditional diffusion model based on U-Net can effectively capture the multi-scale features of the spectrum with its encoder-decoder structure; Targeted Conditional Injection, convert the antiviral key property parameters (such as predicted virus protein binding energy, drug-likeness, etc.) into a condition vector (Condition Vector), and inject it into the model through Cross-Attention mechanism or Feature Concatenation at multiple levels of U-Net, to ensure that the generated spectrum meets multiple constraints at the same time; In this embodiment, the calculation details are as follows: The input of the diffusion model is the calculated spectrum data (plus the gradually added Gaussian noise) and the condition vector; the output is the predicted noise; Loss function: use Mean Squared Error Loss (MSE Loss) to calculate the difference between the predicted noise and the real added noise; Training goal: the model learns to gradually denoise from random noise under the constraint of given property conditions, and generate the target spectrum that meets the requirements; Step three: Spectrum-SMILES conversion model for antiviral molecule generation The encoder (Encoder) is responsible for processing the spectrum sequence. First, use a one-dimensional convolutional layer (1D CNN) to downsample and extract local features of the spectrum, and then input the feature sequence into the Transformer encoder layer to obtain the context-aware representation of the spectrum.

[0015] The decoder (Decoder) uses an autoregressive (Autoregressive) Transformer decoder. It takes the spectrum representation output by the encoder as the context (Key / Value), and takes the already generated SMILES character sequence as the query (Query), to predict the next most likely character; In this embodiment, the calculation details are as follows: Transformer model input and output: the encoder input is the standardized spectrum data sequence generated in step one; the decoder input is the SMILES sequence shifted right (Shifted Right), and the output is the original SMILES sequence; Loss function: use Cross-Entropy Loss to calculate the difference between the predicted character and the real character; Regularization: use Dropout and Label Smoothing to prevent overfitting.

[0016] Step four: end-to-end generation and optimization of antiviral molecules Targeted generation process: the properties of the specific viral target site, such as binding affinity, drug-like properties, and other property conditions, are converted into a condition vector and input into the trained conditional diffusion model; the diffusion model generates a virtual spectrum that meets the conditions through iterative denoising (e.g., 1000-step sampling); then, the generated spectrum is input into the trained Transformer encoder-decoder model, and the decoder generates the corresponding SMILES string step by step in an autoregressive manner; Batch processing and diversity: by setting different random seeds for the diffusion model, multiple different spectra can be generated for the same set of property conditions, thereby producing multiple candidate molecules with different structures but meeting the requirements, increasing the diversity of the output.

[0017] Step five: post-processing and optimization of antiviral candidate molecules Initial screening, including: Chemical validity check: use RDKit to check whether the generated SMILES can be successfully converted into a legal molecule object, and immediately filter out invalid syntax strings; De-duplication: remove duplicate molecules based on normalized SMILES (Canonical SMILES); Conformation generation and optimization, including: Use the ETKDG method of RDKit to generate multiple initial three-dimensional conformations for each candidate molecule; Use molecular force fields such as MMFF94 or UFF to perform energy minimization on these conformations through molecular mechanics (Molecular Mechanics) optimization to find the most stable conformation with the lowest energy; Final screening and output, including: For the optimized stable conformation, perform molecular docking simulation to evaluate the binding mode with the specific viral protein target site, and use ADMET prediction tools to evaluate the in vivo pharmacokinetic properties of the generated molecule.

[0018] Further screening of candidate molecules according to expert experience, and output of molecular structure information suitable for subsequent in vitro antiviral experimental verification.

[0019] The above only describes the preferred embodiments of the present application, and any equivalent changes or modifications made to the structure, features, and principles described in the scope of the present patent application are included in the scope of the present patent application.

Claims

1. A method for reverse design of viral protein drug molecules using molecular spectroscopy as an intermediary, characterized in that, Includes the following steps: S1: Construct a dataset containing antiviral molecules SMILES and their corresponding molecular spectra, and use quantum chemical calculation methods to calculate molecular spectra to ensure the accuracy and physical consistency of the spectral data; S2: Build and train a conditional diffusion model, using user-defined molecular properties and molecular spectra as conditional inputs. The model learns to generate spectral data that meets the conditions. After training until the loss converges, the model is saved and used to generate candidate spectra based on the target properties. S3: Train a Transformer model with an encoder-decoder architecture. Its input is spectral data and its output is the corresponding molecular SMILES sequence. This model achieves accurate decoding from spectrum to molecular representation by capturing the mapping relationship between spectrum and molecular structure. Training is terminated when the validation loss is minimized and the optimal model is saved. S4: Integrate the two trained models above to construct an end-to-end molecular generation process. Given the target properties of the antiviral molecule, first generate the molecular spectrum that meets the conditions through the diffusion model, and then decode it into the SMILES expression corresponding to the spectrum through the Transformer model, thereby obtaining candidate molecules that have both specified physicochemical properties and structural rationality. S5: Perform validity screening and structural optimization on the generated SMILES. Based on chemical rules and energy calculations, eliminate invalid or unstable structures, perform conformational search and geometric optimization on the retained molecules, and finally output a molecular model that meets the user's needs and has a stable structure.

2. The reverse design method for viral protein drug molecules using molecular spectroscopy as an intermediary, as described in claim 1, is characterized in that: The S1 step involves constructing a dataset containing the antiviral molecule SMILES and its corresponding molecular spectra, specifically including the following steps: S1.1: Molecular Source and Screening. Initial SMILES datasets containing antiviral drug molecules were obtained from public compound libraries such as PubChem, ChEMBL, and ZINC. To ensure data diversity and drug relevance, screening was conducted based on the following criteria: Drug-like properties: Preliminary filtering was performed using Lipinski's five rules and PAINS screening rules; Synthetic accessibility (SA): The SA score is calculated using tools such as RDKit. Molecules with an SA score < 4.5 are preferred to ensure the synthetic accessibility of candidate molecules. S1.2: Molecular spectral calculations were performed using the Gaussian 16 quantum chemistry calculation software package. Density functional theory (DFT) was employed, specifically using the B3LYP functional and TZVP basis set for geometry optimization and frequency calculations to ensure a balance between computational accuracy and efficiency. S1.3: Post-processing: The calculated raw spectrum is standardized, peak width broadened and intensity normalized to make the spectral data closer to the experimentally measured spectral shape, and then the model is learned. S1.4: Data storage: Construct a structured database, with each record containing molecular SMILES, optimized 3D conformation, and calculated spectral data.

3. The reverse design method for viral protein drug molecules using molecular spectroscopy as an intermediary, as described in claim 1, is characterized in that: The specific steps for constructing and training the conditional diffusion model in S2 include: S2.1: The conditional diffusion model based on U-Net is adopted. The encoder-decoder structure of U-Net can effectively capture the multi-scale features of the spectrum. S2.2: Targeted Condition Injection: The key antiviral property parameters are converted into a condition vector, which is then injected into the model at multiple levels of U-Net through cross-attention or feature concatenation to ensure that multiple constraints are met while generating the spectrum.

4. The reverse design method for viral protein drug molecules using molecular spectroscopy as an intermediary, as described in claim 1, is characterized in that: The specific steps for constructing and training the conditional diffusion model in S2 include: S2.1: The conditional diffusion model based on U-Net is adopted. The encoder-decoder structure of U-Net can effectively capture the multi-scale features of the spectrum. S2.2: Targeted Condition Injection: The key antiviral property parameters are converted into a condition vector, which is then injected into the model at multiple levels of U-Net through cross-attention or feature concatenation to ensure that multiple constraints are met while generating the spectrum.

5. The reverse design method for viral protein drug molecules using molecular spectroscopy as an intermediary, as described in claim 1, is characterized in that: The S3 trains a Transformer model with an encoder-decoder architecture, specifically including the following model architecture: Encoder: Responsible for processing spectral sequences. First, a one-dimensional convolutional layer (1D CNN) is used to downsample the spectrum and extract local features. Then, the feature sequence is input into the Transformer encoder layer to obtain a context-aware representation of the spectrum. Decoder: Employs an autoregressive Transformer decoder that uses the spectral representation of the encoder output as context (key / value) and the SMILES character sequence that has already begun to be generated as query to predict the next most likely character.

6. The reverse design method for viral protein drug molecules using molecular spectroscopy as an intermediary, as described in claim 1, is characterized in that: S4 integrates the two trained models mentioned above to construct an end-to-end molecular generation process, which specifically includes the following steps: S4.1: Target generation process: Based on the property requirements of specific viral targets, the property conditions are converted into a conditional vector and input into a pre-trained conditional diffusion model. The diffusion model generates a virtual spectrum that meets the conditions through iterative denoising. Then, this generated spectrum is input into a pre-trained Transformer encoder-decoder model. The decoder generates the corresponding SMILES string step by step in an autoregressive manner. S4.2: Batch processing and diversity: By setting different random seeds for the diffusion model, multiple different spectra can be generated for the same set of property conditions, thereby producing multiple candidate molecules with different structures but all meeting the requirements, increasing the diversity of output.

7. The reverse design method for viral protein drug molecules using molecular spectroscopy as an intermediary, as described in claim 1, is characterized in that: In step S5, the generated SMILES undergo validity screening and structural optimization, specifically including the following steps: S5.1 Initial Screening includes the following steps: S5.1.1: Chemical validity check: Use RDKit to check whether the generated SMILES can be successfully converted into a valid molecular object, and immediately filter out strings with invalid syntax; S5.1.2: Removal of duplicates: Removing duplicate molecules based on normalized SMILES (Canonical SMILES); S5.2 Conformation generation and optimization specifically includes the following steps: S5.2.1: Use the ETKDG method of RDKit to generate multiple initial three-dimensional conformations for each candidate molecule; S5.2.2: Using molecular force fields such as MMFF94 or UFF, the energy of these conformations is minimized through molecular mechanics optimization to find the most stable conformation with the lowest energy. S5.3: Final filtering and output, specifically including the following steps: S5.3.1: Perform molecular docking simulations on the optimized stable conformation to evaluate the binding mode with specific viral protein targets, and use the ADMET prediction tool to evaluate the in vivo pharmacokinetic properties of the generated molecule. S5.3.2: Further screen candidate molecules based on expert experience, and output the molecular structure information in a format suitable for subsequent in vitro antiviral experiments.

Images