Molecular generation methods, systems, media, and apparatuses for drug discovery

Abstract

The application relates to the technical field of computer-aided drug design, and provides a molecule generation method, system, medium and equipment for drug discovery, which comprises the following steps: acquiring a SMILES string of a ligand and an amino acid sequence of a target protein, respectively extracting a ligand feature vector and a protein feature vector, splicing and fusing, and then predicting a binding affinity value through a multilayer perception machine; taking the protein feature vector as a condition, generating a new molecular potential representation through a reverse denoising process of a conditional diffusion model; wherein, at each step of the reverse denoising process, a graph-level readout operation is performed on pure noise, an affinity value is predicted through the multilayer perception machine, a gradient of an affinity guidance loss is returned to a noise prediction network, and the new molecule generation is guided to a high affinity area; and the new molecular potential representation is decoded into a SMILES string through a pre-trained molecular language decoder. Novel molecules with high binding potential can be quickly generated.

Description

Technical Field

[0001] This invention belongs to the field of computer-aided drug design technology, and particularly relates to molecular generation methods, systems, media and devices for drug discovery. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] New drug development is central to driving medical progress, but the process is typically lengthy, costly, and has a very low clinical success rate. In this process, the core task of early drug discovery is to rapidly and accurately identify small molecule lead compounds from a vast chemical space that have high binding affinity to specific disease targets (usually proteins) and exhibit good drug resistance.

[0004] In recent years, deep learning technology has deeply penetrated all aspects of drug discovery and formed two major technological paradigms:

[0005] The first category is predictive models, such as DeepDTA (a deep learning-based drug-target affinity prediction model) and GraphDTA (a graph neural network-based drug-target affinity prediction model). These models learn from known drug-target interaction data to build a model from molecular structure (such as SMILES (Simplified molecular input lineentry system) strings or molecular graphs) and protein sequences to binding affinity (such as the negative logarithmic constant value pIC of drug activity). 50 Mapping functions of ) can serve as efficient "virtual filters" to quickly assess the binding strength of a given molecule to a target site, but they do not have the ability to create new molecular structures.

[0006] The second category is generative models, including variational autoencoders (VAEs), generative adversarial networks (GANs), autoregressive models (MolGPT), and emerging diffusion models. These models can generate a large number of novel molecules from scratch. However, these models are usually trained with the goal of optimizing sequence likelihood or structural rationality, and do not incorporate quantitative, continuous binding affinity signals as a direct optimization target into the generation process. Therefore, although the generated molecules may be chemically effective, their binding potential to specific targets is unknown, like a "blind box".

[0007] In current actual R&D processes, the two types of models mentioned above are disconnected from each other. Researchers are forced to adopt an expensive iterative cycle of "generation → virtual screening → optimization → regeneration," which has inherent flaws.

[0008] Inefficient: Each iteration involves large-scale molecule generation and computationally intensive virtual screening (such as molecule docking), which is time-consuming and computationally expensive.

[0009] Distribution shift: The chemical space explored by the generative model may be far from the distribution covered by the training data of the predictive model, resulting in unreliable scores from the predictive model, i.e., "the evaluation criterion fails".

[0010] Guidance lag: The generation process itself is goal-blind, and optimization can only rely on posterior, discrete screening results, rather than real-time, continuous affinity gradient guidance during the generation process.

[0011] While existing research has attempted to connect predictors and generators using reinforcement learning or multi-task recurrent neural networks (RNNs), these methods typically achieve only shallow coupling. For example, reinforcement learning uses a frozen predictor to provide the reward signal, whose gradients cannot penetrate the generator; multi-task RNNs only perform joint training at the loss function level, failing to achieve gradient-level guidance for each step of the generation process. Some recent diffusion models (such as the deep learning algorithm DiffAffinity designed specifically for antibodies), while introducing affinity, are primarily aimed at generating three-dimensional (3D) structures and still require a post-processing SMILES transformation step. Summary of the Invention

[0012] To address the technical problems mentioned above, this invention provides a molecular generation method, system, medium, and device for drug discovery. It incorporates the gradient signal of affinity prediction as a guiding term into the training process of a conditional diffusion model. Through affinity-guided loss, it achieves real-time optimization guidance for molecular generation, ensuring that the generated molecules possess both chemical efficacy and high theoretical binding affinity. This overcomes the inefficiency and distribution shift problems caused by the separation of prediction and generation tasks in traditional drug design, enabling the rapid generation of novel molecular structures with high binding potential.

[0013] To achieve the above objectives, the present invention adopts the following technical solution:

[0014] A first aspect of the present invention provides a molecular generation method for drug discovery, comprising:

[0015] The SMILES string of the ligand and the amino acid sequence of the target protein are obtained, and the ligand feature vector and the protein feature vector are extracted respectively through a dual-branch feature encoding network.

[0016] After concatenating and fusing the ligand feature vector and the protein feature vector, the binding affinity value is predicted by a multilayer perceptron. The mean squared error loss is calculated based on the binding affinity value, and the multilayer perceptron is trained and optimized.

[0017] Using protein feature vectors as conditions, new molecular potential representations are generated through a reverse denoising process of a conditional diffusion model. In each step of the reverse denoising process, a graph-level readout operation is performed on the pure noise, and the readout vector is input into a multilayer perceptron to predict the affinity value. The mean square error between the predicted affinity value and the true value is calculated as the affinity guidance loss, and the gradient of the affinity guidance loss is backpropagated to the noise prediction network to guide the generation of new molecules toward the high affinity region.

[0018] The new molecular latent representations are decoded into the string SMILES using a pre-trained molecular language decoder.

[0019] Furthermore, the dual-branch feature encoding network includes processing the molecular graph using a parallel graph attention network and a graph convolutional network, and the molecular graph is obtained by parsing the SMILES string of the ligands using cheminformatics tools.

[0020] Furthermore, the dual-branch feature encoding network includes a hierarchical structure that processes the amino acid sequence of the target protein using a convolutional neural network, a bidirectional long short-term memory network, and a Transformer encoder.

[0021] Furthermore, the training of the noise prediction network uses a weighted sum of affinity-guided loss and noise prediction loss as the loss function.

[0022] A second aspect of the present invention provides a molecular generation system for drug discovery, comprising:

[0023] The feature encoding module is configured to: acquire the SMILES string of the ligand and the amino acid sequence of the target protein, and extract the ligand feature vector and the protein feature vector respectively through a two-branch feature encoding network;

[0024] The affinity prediction module is configured to: concatenate and fuse the ligand feature vector and the protein feature vector, predict the binding affinity value through a multilayer perceptron, calculate the mean squared error loss based on the binding affinity value, and train and optimize the multilayer perceptron.

[0025] The molecule generation module is configured to generate new potential molecular representations using protein feature vectors as conditions through a reverse denoising process of a conditional diffusion model. In each step of the reverse denoising process, a graph-level readout operation is performed on the pure noise, the readout vector is input into a multilayer perceptron to predict the affinity value, the mean square error between the predicted affinity value and the true value is calculated as the affinity guidance loss, and the gradient of the affinity guidance loss is backpropagated to the noise prediction network to guide the generation of new molecules toward the high affinity region.

[0026] The decoding module is configured to decode new molecular latent representations into the string SMILES using a pre-trained molecular language decoder.

[0027] Furthermore, the dual-branch feature encoding network includes processing the molecular graph using a parallel graph attention network and a graph convolutional network, and the molecular graph is obtained by parsing the SMILES string of the ligands using cheminformatics tools.

[0028] Furthermore, the dual-branch feature encoding network includes a hierarchical structure that processes the amino acid sequence of the target protein using a convolutional neural network, a bidirectional long short-term memory network, and a Transformer encoder.

[0029] Furthermore, the training of the noise prediction network uses a weighted sum of affinity-guided loss and noise prediction loss as the loss function.

[0030] A third aspect of the invention provides a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of the molecular generation method for drug discovery described above.

[0031] A fourth aspect of the present invention provides a computer device including a computer-readable storage medium, a processor, and a computer program stored on the computer-readable storage medium and executable on the processor, wherein the processor executes the program to implement the steps of the molecular generation method for drug discovery described above.

[0032] Compared with the prior art, the beneficial effects of the present invention are:

[0033] This invention introduces the gradient signal of affinity prediction as a guiding term into the training process of the conditional diffusion model. By using affinity-guided loss, it achieves real-time optimization guidance for molecule generation, ensuring that the generated molecules have both chemical efficacy and high theoretical binding affinity. This overcomes the inefficiency and distribution shift problems caused by the separation of prediction and generation tasks in traditional drug design, and can rapidly generate novel molecular structures with high binding potential. Attached Figure Description

[0034] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0035] Figure 1 This is a flowchart of a molecular generation method for drug discovery according to Embodiment 1 of the present invention;

[0036] Figure 2 This is a schematic diagram illustrating the correlation between the affinity values ​​predicted by the model and the actual affinity values ​​on the Davis dataset in Embodiment 1 of the present invention.

[0037] Figure 3 This is a schematic diagram illustrating the correlation between the affinity values ​​predicted by the model and the actual affinity values ​​on the KIBA dataset in Embodiment 1 of the present invention.

[0038] Figure 4 This is a quality assessment index diagram of the generated molecules in Embodiment 1 of the present invention;

[0039] Figure 5 This is a comparison diagram of the distribution of molecules generated on the Davis dataset and molecules in the training set in terms of key physicochemical properties in Embodiment 1 of the present invention;

[0040] Figure 6 This is a comparison diagram of the distribution of molecules generated on the KIBA dataset and molecules in the training set in terms of key physicochemical properties in Embodiment 1 of the present invention;

[0041] Figure 7 This is a clustering analysis diagram of the generated molecule and the original molecule in the latent space of Embodiment 1 of the present invention;

[0042] Figure 8 This is a predicted affinity distribution map of molecules generated on the Davis dataset in Embodiment 1 of the present invention;

[0043] Figure 9 This is a predicted affinity distribution map of molecules generated on the KIBA dataset in Embodiment 1 of the present invention;

[0044] Figure 10 This is a comparison diagram of the first binding mode obtained by molecular docking of the generated molecule and the original molecule in Embodiment 1 of the present invention;

[0045] Figure 11 This is a comparison diagram of the second binding mode obtained by molecular docking of the generated molecule and the original molecule in Embodiment 1 of the present invention;

[0046] Figure 12 This is a comparison diagram of the third binding mode obtained by molecular docking of the generated molecule and the original molecule in Embodiment 1 of the present invention;

[0047] Figure 13 This is a comparison diagram of the fourth binding mode obtained by molecular docking of the generated molecule and the original molecule in Embodiment 1 of the present invention;

[0048] Figure 14 This is a schematic diagram of the structure of a computer device according to Embodiment 4 of the present invention. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0050] Example 1

[0051] This embodiment provides a molecular generation method for drug discovery.

[0052] This embodiment provides a molecular generation method for drug discovery that achieves deep synergy between the two tasks of unified affinity prediction and molecular generation at the gradient level, thereby breaking through the bottlenecks of existing technologies and providing a truly efficient and targeted design tool for drug discovery.

[0053] This embodiment provides a molecular generation method for drug discovery, the core of which lies in constructing a multi-task diffusion model (MT-DiffGen) that integrates affinity prediction and molecular generation into an end-to-end differentiable pipeline.

[0054] This embodiment provides a molecular generation method for drug discovery, constructing an end-to-end deep learning framework. A dual-branch feature encoding network processes molecular graph structures and protein sequences separately, extracting unified feature representations. Based on this, it simultaneously performs two tasks: accurate prediction of drug target affinity and target-aware molecular generation. Innovatively, the gradient signal of affinity prediction is introduced as a guiding term into the training process of the conditional diffusion model. Affinity-guided loss provides real-time optimization guidance for molecular generation, ensuring that the generated molecules possess both chemical efficacy and high theoretical binding affinity. This overcomes the inefficiency and distribution shift problems caused by the separation of prediction and generation tasks in traditional drug design, enabling the rapid generation of novel molecular structures with high binding potential, and providing an effective computational platform for accelerating the discovery of drug lead compounds.

[0055] This embodiment provides a molecular generation method for drug discovery, such as... Figure 1 As shown, it includes the following steps:

[0056] Step 1: Input processing steps.

[0057] It accepts two types of one-dimensional sequences as input: the SMILES string of the ligand and the amino acid sequence of the target protein; three-dimensional structural information without any experimental resolution; and input data is preprocessed into a standard format.

[0058] Step 2: Feature extraction.

[0059] Heterogeneous input data is mapped to a unified latent space through a two-branch feature encoding network.

[0060] Step 201, Ligand coding branch.

[0061] (1) Use cheminformatics tools (e.g., the Rational Discovery Toolkit RDKit) to parse the SMILES string into a molecular diagram. ,in, V A collection of atoms E Let X be the key set. atom Let A be the atomic characteristic matrix and A be the adjacency matrix.

[0062] (2) For the molecular graph, perform initial node encoding: Among them, W node It is a learnable projection matrix. b represents the original feature vector (such as atom type, charge, etc.) of the i-th node (atom) in the molecular diagram. node The learnable bias term is used to adjust the baseline of the node embedding; the initial node encoding is followed by layer-norm and ReLU activation; to inject topological context information, the normalized graph Laplacian matrix is ​​calculated. , ,in, Represents the normalized Laplace matrix. It is the identity matrix. The degree matrix of the molecular graph. Let be the adjacency matrix of the molecular graph. This represents the value in the i-th row and j-th column of the adjacency matrix of the molecular graph. The value of the i-th row in the degree matrix of the molecular diagram is represented; and the eigenvectors corresponding to the k=8 smallest eigenvalues ​​are selected as structural position codes: , ,in, For structural position encoding matrix, This is the initial encoding vector for the node. This is the feature vector of node i in the structural position encoding matrix. It is a learnable projection matrix.

[0063] (3) Parallel graph attention network (GAT) and graph convolutional network (GCN) are used to encode the molecular graph.

[0064] Among them, GAT captures dynamic, local interactions between atoms (such as pharmacophores) through a multi-head attention mechanism, namely:

[0065] , ;

[0066] in, It is the attention coefficient between nodes i and j in the h-th attention head; It is the feature vector of node i in the l-th layer; It is the feature vector of node j in the l-th layer; It is the feature vector of node k in the l-th layer; It is the learnable weight matrix of the h-th attention head; It is the set of first-order neighbors of node i; σ is the activation function (ELU); concatenation operation Aggregate the outputs of H=8 attention heads; LeakyReLU represents the leakage rectified linear unit; a represents the attention coefficients of nodes i and j in the h-th attention head; The GAT output feature vector of node i in layer l+1 is obtained by combining neighbor features with attention weighting, processing with activation functions, and then concatenating the outputs of 8 attention heads.

[0067] GCN integrates the global topology of molecules through message passing, namely:

[0068] ;

[0069] in, It is a layer-specific weight matrix; This indicates that the average aggregation operation is performed on node i and its neighboring nodes; residual connection Gradient flow used to stabilize deep graph networks.

[0070] After L=4 layers of calculation, the GAT outputs are obtained respectively. and GCN output .

[0071] (4) Global average pooling and global max pooling are performed on the outputs of GAT and GCN respectively. The four feature vectors obtained are concatenated and then dimensionality reduced and fused through a fully connected layer (MLP) to finally output a fixed-dimensional ligand feature vector. This dual-pooling strategy preserves both the average and extreme features of the substructures, ultimately generating a 1024-dimensional ligand vector.

[0072] Step 202, Protein coding branch: A hierarchical hybrid architecture is used to process the amino acid sequence.

[0073] First, each amino acid residue is mapped to a dense vector through a learnable embedding layer, i.e.: ; Sigma=20 corresponds to the encoding mapping of 20 standard amino acids, that is, mapping each amino acid to a numerical vector through one-hot encoding and other methods; where s t It is the unique heat vector of the t-th amino acid label; It is a learnable residue embedding matrix; It is the bias vector; The input will be a multi-scale CNN-BiLSTM-Transformer stack. This embodiment does not use frozen ESM-2 (Evolutionary Scale Modeling v2) weights, but instead... End-to-end updates are performed so that gradients from the affinity predictor and the diffusion decoder can be jointly optimized to shape a task-specific residue representation.

[0074] Next, a multi-scale one-dimensional convolutional neural network (CNN) (e.g., kernel sizes of 3, 5, and 7) is used to extract local short peptide motifs from the sequence, namely: ;in, Starting from position t, with a width of Amino acid insertion fragments; It is the weight tensor of the k-th one-dimensional convolution kernel; This represents a one-dimensional convolution operation; It is the bias vector; ReLU is a non-linear activation function; the output is... This refers to the local motif feature extracted by the k-th convolutional kernel at position t. The local motif feature will be input into the downstream BiLSTM for long-range dependency modeling.

[0075] Then, long-range dependencies in the sequence are captured using a bidirectional long short-term memory (BiLSTM) network, specifically by updating the hidden state using a two-layer bidirectional LSTM. ;in, This represents the negative hidden state at position t. This represents the positive hidden state at position t. This represents the input feature at position t, and the final bidirectional outputs are concatenated to obtain a concatenated vector of total length L: L is the total length of the assembled features.

[0076] Next, a Transformer encoder is used to model global residue-residue interactions through a self-attention mechanism: a two-module Transformer encoder with a 4-head self-attention mechanism is employed for feature refinement. Self-attention weights can explicitly model long-range interactions (such as allosteric rings) without requiring a structural template; the output H of the Transformer... trans Perform sequence-level average pooling (mean) t ), to obtain the protein feature vector ,in, This indicates sequence-level average pooling.

[0077] Step 203: Transfer the ligand feature vector Protein feature vectors They are spliced ​​together and input into the downstream affinity predictor and conditional diffusion decoder.

[0078] All embedding matrices employ learnable parameters and are optimized end-to-end to ensure that chemical and biological features can synergistically adapt to the dual tasks of affinity regression and de novo generation.

[0079] Step 3, Affinity Prediction Steps.

[0080] First, the ligand feature vector and protein feature vectors By splicing them together, a fused interaction representation is formed. D p This indicates a splicing operation.

[0081] Then, the fusion representation z (0) Inputting a multilayer perceptron (MLP) regressor, it predicts the binding affinity value (e.g., pIC). 50 That is, through a three-layer fully connected module, the representation is refined and integrated step by step: ,in, and The parameters are learnable; ELU (Exponential Linear Unit) is chosen as the smooth, non-saturating activation function. L1 represents the total number of layers in the multilayer perceptron. This represents the fusion representation of the l-th layer.

[0082] Finally, a Pre-LN (pre-layer normalization) structure with layer normalization and a dropout strategy (p=0.1 as the dropout probability) is adopted before the affine transformation to enhance the stability during training with the diffusion decoder. The scalar output is finally obtained through affine mapping. ,in, To incorporate affinity predictions, w represents a learnable weight. Let be the output feature vector, and b be the learnable bias term.

[0083] Affinity prediction models lose L through mean squared error (MSE). aff To conduct training, that is Where B represents batch data. To predict the difference between affinity and true affinity.

[0084] This embodiment applies weight decay. Gradient clipping (threshold 1.0) is used to ensure the stability of the diffusion generation path during end-to-end training; Maintaining complete differentiability throughout the graph structure and sequence encoder, making pIC 50 The gradient can flow to both affinity regression and conditional SMILES generation paths simultaneously in a single backpropagation.

[0085] Step 4: Molecular Generation. A conditional diffusion model is used to generate new molecules. The conditional diffusion model uses protein feature vectors... As a condition.

[0086] (1) Forward process: representing the real ligands from the training data (i.e., ligand feature vector) (This can be obtained from ligand branching) Gaussian noise is gradually added during the T-step until it becomes pure noise. ,Right now: ,in, Indicates the forward diffusion distribution probability. Represents the noise figure. The covariance matrix of the distribution is denoted by ; its conditional posterior distribution is: , As a posterior constraint, this means that protein conditions do not participate in the noise addition operation during forward diffusion, thus ensuring that the binding signal is injected into the conditional diffusion model only during the generation phase.

[0087] (2) Reverse process (training): Based on the ELBO (lower bound of evidence) derivation, the complete diffusion loss includes the noise prediction loss and the KL divergence loss. Among them, the KL divergence (relative entropy) loss constrains the consistency between the reverse distribution and the forward and posterior distributions. ,in, This represents the mathematical expectation of the time step, molecular characterization, and prediction noise. KL divergence (relative entropy) is used to measure the difference between two probability distributions. For the posterior distribution of the forward process, This represents the potential molecular characterization at step t-1. Represents the mathematical expectation. Describe the model probability distribution of the reverse process; train a noise prediction network. ,in, This represents the potential molecular characterization after adding noise, where t represents the time step information. For protein constraint information; learning from pure noise The added Gaussian noise is predicted in the middle.

[0088] The core innovation of this embodiment lies in the fact that, when training the noise prediction network, its loss function, in addition to the standard noise prediction loss... In addition, an affinity-guided loss was added, namely: ;in, This represents the mathematical expectation of the time step, molecular characterization, and prediction noise. This represents the actual noise added during forward diffusion. This indicates the noise predicted by the noise prediction network; Used to balance noise reduction fidelity and drug efficacy; This is a differentiable mean pooling operation; Used as a reward mechanism, it is the frozen affinity MLP; y represents the true affinity value.

[0089] Specifically, in each denoising step, the intermediate latent representations are... Perform a differentiable graph-level ReadOut operation, and input the readout vector into the affinity prediction module to obtain information about the corresponding target protein. The predicted value is calculated, and the mean square error between the predicted value and the true affinity value is calculated; the gradient of the mean square error is backpropagated to the noise prediction network, so that each step of the denoising process is directly guided by the gradient of the goal of "generating molecules with higher affinity".

[0090] (3) Generation process (reasoning): Given a target protein feature vector Pure noise sequences randomly sampled from a Gaussian distribution Initially, a noise prediction network predicts the noise level for the current step, and the noise level is subtracted from the pure noise level. A total of T steps of conditional denoising are performed. Each step is influenced by both protein conditions and affinity (which can be adjusted by changing the predicted values ​​after training). Ultimately, a novel, target-optimized potential molecular characterization is generated. .

[0091] Step 5: Structure Decoding Step.

[0092] The generated low-noise potential characterization By directly inputting a pre-trained and finely tuned molecular language decoder (such as the molecular T5 model), which is based on the Transformer architecture, continuous latent representations can be mapped to syntactically correct and chemically valid canonical SMILES strings without any post-processing graph-to-string conversion steps, near 100% chemical validity can be guaranteed. ,in, This represents the molecular characterization generated by noise diffusion. This indicates the average pooling operation. This represents the Smiles (simplified molecular linear input specification) used to generate molecular decoding. This represents the fine-tuned molecular T5 model decoder. During the diffusion training phase, the decoder parameters are frozen to avoid catastrophic forgetting, but its cross-entropy loss is... The weights are added to the overall objective function.

[0093] The entire model incorporates affinity loss. diffusion loss and KL divergence loss The multi-task objective function is used for end-to-end joint training, which ensures the collaborative optimization of all modules in a unified latent space.

[0094] Experiment 1: Affinity prediction and molecular generation based on the Davis dataset (kinase inhibitor binding affinity dataset).

[0095] Data preparation: The Davis dataset was used, which contains 30,056 binding affinity data points for 68 drugs and 442 target proteins; the dissociation constant (K) was used. d Convert to the negative logarithm of the dissociation constant pK d value: ; Filter out K d =5,000 nM of inactive data points, retaining measurable affinity data; the training and test sets are divided in an 80 / 20 ratio.

[0096] Model configuration:

[0097] Ligand branch: Atomic feature dimension: 78; GATv2: 10 attention heads, output dimension 780; GCN: 4 convolutional layers, output dimension 312; Feature fusion: global average pooling + global max pooling, output dimension 128. ;

[0098] Protein branch: Amino acid feature dimension: 25; Multi-scale CNN: Kernel size [3,7,15], number of filters 128; BiLSTM: 2 layers, 64 hidden units, output 128 dimensions; Transformer: 4-head attention, 2 encoder layers; Output: 128 dimensions ;

[0099] Diffusion module: 1000 steps, noise figure The linear schedule is [1e-4, 0.02].

[0100] Training details: Batch size: 512; Initial learning rate: 1e-3; Number of training epochs: 1000; Loss function: Gradient clipping: L2 norm = 1.0.

[0101] result:

[0102] Affinity prediction performance (test set): CI (Conformity Index): 0.895±0.001; MSE (Mean Squared Error): 0.209±0.002; (Modified coefficient of determination): 0.683 ± 0.004;

[0103] Molecular formation quality: Chemical validity: 99.9%; Novelty: 30%; Uniqueness: 26%.

[0104] Experiment 2: Model validation based on the KIBA (Kinetic Inhibitor Bioactivity Dataset) dataset.

[0105] Data preparation: The KIBA dataset was used, containing 117,954 data points for 2,111 drugs and 229 target proteins; weak binders with KIBA scores below 10 were excluded; the same 80 / 20 training / test split as Davis was used.

[0106] Model configuration: The architecture and hyperparameters remained the same as in Experiment 1. MoIT5 decoder fine-tuning: Base model: molt5-small (small molecular T5 model); learning rate: 1e-3; batch size: 8; number of training epochs: 500; using a cosine annealing learning rate scheduler.

[0107] result:

[0108] Affinity prediction performance (test set): CI (Conformity Index): 0.891±0.001; MSE (Mean Squared Error): 0.142±0.002; (Modified coefficient of determination): 0.762 ± 0.004;

[0109] Molecular generation quality: Chemoavailability: 99.9%; Diversity score: 0.827; Recovery rate: 0.447.

[0110] Experiment 3: Analysis of the physicochemical properties of the generated molecules.

[0111] Analytical methods: 10,000 molecules were generated from both the Davis and KIBA test sets; RDKit was used to calculate eight key physicochemical properties: molecular weight (MW); lipid-water partition coefficient (LogP); number of hydrogen bond donors (HBD); number of hydrogen bond acceptors (HBA); topological polar surface area (TPSA); number of rotatable bonds; number of aromatic rings; and quantitative evaluation of drug-likeness (QED).

[0112] Results: The physicochemical properties of all generated molecules were highly consistent with those of the training set molecules. Molecular weight differences were less than 50 Da, and the LogP distribution patterns were almost identical. The average difference in the number of hydrogen bond donors / acceptors was within ±0.1. The difference in TPSA distribution patterns was less than 5 Ų. The QED distribution peak was approximately 0.75, consistent with the training set.

[0113] Experiment 4: Verification of chemical spatial distribution.

[0114] Analysis methods: Principal component analysis (PCA) and t-distributed random neighborhood embedding (t-SNE) were used for dimensionality reduction visualization. K-means clustering profile coefficients were calculated. The nearest neighbor Euclidean distance distribution was analyzed.

[0115] Results: In both PCA and t-SNE plots, the generated molecules and the original training molecules highly overlap in the latent space. K-means silhouette coefficients: Davis 0.122, KIBA 0.077. The nearest neighbor distance histogram shows a similar distance distribution between the generated molecules and the training set molecules. There is no evidence of mode collapse; the generated molecules are uniformly distributed in the training set chemical space.

[0116] Experiment 5: Validation of the effectiveness of the affinity guidance mechanism.

[0117] Experimental design: Compare the performance of models with and without affinity-guided loss. Use the same training data and hyperparameters. Evaluate the docking score and synthetic accessibility of the generated molecules.

[0118] result:

[0119] Affinity-guided model: docking score improved by 18%; synthesis accessibility score (SA Score) decreased by 24%; generated molecule prediction pIC... 50 The distribution is closer to the highly active areas;

[0120] The model without affinity guidance: the affinity distribution of generated molecules is not significantly different from that of the training set; the docking score is comparable to that of randomly generated molecules.

[0121] Experiment 6: Ablation Experiment Study.

[0122] On the Davis dataset, each module was removed from the system to evaluate its impact on performance. The experimental results are shown in Table 1.

[0123] Table 1. Comparison of Experimental Results

[0124]

[0125] Experiment 7: Drug design and affinity prediction platform based on MT-DiffGen model.

[0126] (1) Data preparation and preprocessing.

[0127] Datasets: The publicly available benchmark datasets Davis and KIBA were used for model training and evaluation. The training and test sets were divided according to the established 80 / 20 ratio of the research.

[0128] Ligand processing: The SMILES string for each drug was converted to canonical form using RDKit and further parsed into a molecular diagram. Atomic features include element type, degree, hybridization state, aromaticity, implicit hydrogen number, etc., totaling 78 dimensions.

[0129] Protein processing: Amino acid sequences were uniformly processed to a maximum length of 1024; excessively long sequences were truncated, and excessively short sequences were padded. A learnable embedding table was used to encode 20 natural amino acids into 128-dimensional dense vectors.

[0130] (2) Model construction and configuration

[0131] The key hyperparameter configurations for the MT-DiffGen model are shown in Table 2.

[0132] Table 2. Hyperparameter Configuration of Core Model Modules

[0133]

[0134] (3) Model training.

[0135] An end-to-end joint training strategy is adopted, and the specific configuration is shown in Table 3.

[0136] Table 3. Training and Generation Configurations

[0137]

[0138] (4) Molecular generation and evaluation.

[0139] Generation process: For any target protein sequence in the test set, its conditional vector is obtained through protein branching encoding. From standard Gaussian noise Initially, conditional denoising is performed for T=1000 steps (no affinity loss is needed during inference, but its effect is already internalized into the network parameters), resulting in... .Will Input a finely tuned MoIT5 decoder to directly generate a standard SMILES string.

[0140] Table 4. Davis Test Set Results

[0141]

[0142] Affinity prediction: Calculate MSE, CI, and [missing information] on the Davis and KIBA test sets. As shown in Tables 4 and 5, MT-DiffGen achieved performance comparable to or better than state-of-the-art models. Figure 2 and Figure 3 As shown, a scatter plot of the predicted affinity values ​​versus experimental values ​​on the Davis and KIBA datasets is presented, along with the model-predicted pIC. 50 The correlation between the predicted values ​​and the actual experimental values, with data points densely distributed along the diagonal, demonstrates the accuracy of the prediction.

[0143] Table 5. Results of the KIBA test set

[0144]

[0145] Molecular generation: The effectiveness (proportion of molecules checked by RDKit), novelty (proportion of effective molecules not in the training set), and uniqueness (proportion of effective molecules that are not repeated) of the generated molecules were evaluated. As shown in Table 6, MT-DiffGen achieved a molecular generation effectiveness of 99.9%, and its novelty and uniqueness were significantly higher than the baseline generation model.

[0146] Table 6. Results of Molecular Generation

[0147]

[0148] Figure 4 The table shows a quantitative comparison of the diversity, recovery rate, and success rate of generated molecules on the Davis and KIBA test sets, intuitively demonstrating the superior performance of generated molecules in multiple dimensions. Table 7 shows the chemical structures of randomly selected generated molecules and their corresponding original molecules in the training set, used to qualitatively evaluate the structural rationality and novelty of generated molecules.

[0149] Table 7. Comparison of generated molecules with original corresponding molecules

[0150]

[0151] (5) Results analysis.

[0152] Physicochemical property analysis: such as Figure 5 and Figure 6 As shown, the distribution of the generated molecule in several key drug-likeness indicators, such as molecular weight, LogP, hydrogen bonding characteristics, and polar surface area, is highly consistent with that of the training set molecules, proving its good drug-likeness and that no distribution drift has occurred.

[0153] Chemical spatial analysis: such as Figure 7As shown, PCA / t-SNE visualization and nearest neighbor distance analysis indicate that the generated molecules fully explored the chemical space of the training set, did not exhibit mode collapse, and possessed sufficient structural diversity.

[0154] Functional potential analysis: such as Figure 8 , Figure 9 , Figure 10 , Figure 11 , Figure 12 , Figure 13 As shown, the predicted affinity distribution of the generated molecule is reasonable, and molecular docking simulations show that it can form reasonable interactions with the target protein, which supports its potential as a lead compound from a computational perspective. Among them, the SwissParam Score is the molecular parameter score generated by the parameterization tool (SwissParam).

[0155] This embodiment provides a molecular generation method for drug discovery that achieves true task unification and gradient-level synergy: through a shared dual-branch encoder and an affinity-guided diffusion mechanism, it achieves deep integration of prediction and generation tasks at the gradient level for the first time, breaking the traditional fragmented process and realizing "evaluation at design time".

[0156] This embodiment provides a molecular generation method for drug discovery, which generates high-quality, high-potential molecules. Because each step of the generation process is guided in real time by the affinity gradient, the generated molecules not only have extremely high chemical efficacy (experimentally verified to be up to 99.9%), but also have a significantly higher predicted binding affinity than target-independent generation methods, with a docking score improvement of about 18% and better synthetic accessibility.

[0157] This embodiment provides a molecular generation method for drug discovery that improves the generalization ability of prediction models: the generation task provides rich "hard samples" for the prediction model, which helps to optimize its decision boundary and improve the prediction accuracy (CI, MSE, ...) on benchmark datasets such as Davis and KIBA. It has reached or surpassed the most advanced models dedicated to prediction.

[0158] This embodiment provides a molecular generation method for drug discovery that possesses strong practicality and versatility: the model relies only on one-dimensional sequence information, eliminating the dependence on the three-dimensional structure of proteins, making it widely applicable to targets lacking crystal structures, and greatly expanding its application scope in early drug discovery.

[0159] This embodiment provides a molecular generation method for drug discovery that greatly improves R&D efficiency: it compresses the traditional multi-step iterative cycle into a single forward reasoning process, significantly reducing the computational and human resource costs of virtual screening and accelerating the process from target identification to lead compound discovery.

[0160] This embodiment provides a molecular generation method for drug discovery that improves the generalization ability of affinity prediction through a shared latent space and joint training, while avoiding distribution shifts during the generation process.

[0161] Example 2

[0162] This embodiment provides a molecular generation system for drug discovery, comprising:

[0163] The feature encoding module is configured to: acquire the SMILES string of the ligand and the amino acid sequence of the target protein, and extract the ligand feature vector and the protein feature vector respectively through a two-branch feature encoding network;

[0164] The affinity prediction module is configured to: concatenate and fuse the ligand feature vector and the protein feature vector, predict the binding affinity value through a multilayer perceptron, calculate the mean squared error loss based on the binding affinity value, and train and optimize the multilayer perceptron.

[0165] The molecule generation module is configured to generate new potential molecular representations using protein feature vectors as conditions through a reverse denoising process of a conditional diffusion model. In each step of the reverse denoising process, a graph-level readout operation is performed on the pure noise, the readout vector is input into a multilayer perceptron to predict the affinity value, the mean square error between the predicted affinity value and the true value is calculated as the affinity guidance loss, and the gradient of the affinity guidance loss is backpropagated to the noise prediction network to guide the generation of new molecules toward the high affinity region.

[0166] The decoding module is configured to decode new molecular latent representations into the string SMILES using a pre-trained molecular language decoder.

[0167] It should be noted that each module in this embodiment corresponds one-to-one with each step in Embodiment 1, and their specific implementation processes are the same, so they will not be repeated here.

[0168] Example 3

[0169] This embodiment provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in a molecular generation method for drug discovery as described in Embodiment 1 above.

[0170] Example 4

[0171] This embodiment provides a computer device, such as... Figure 14 As shown, the system includes a computer-readable storage medium 1003, a processor 1001, a communication interface 1002, and a computer program stored on the computer-readable storage medium 1003 and executable on the processor 1001. The processor 1001, communication interface 1002, and computer-readable storage medium 1003 can be connected via a bus or other means. The communication interface 1002 is used to receive and transmit data. When the processor 1001 executes the program, it implements the steps in the molecular generation method for drug discovery as described in Embodiment 1 above.

[0172] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for molecular generation for drug discovery, characterized by, The method comprises the following steps: obtaining a SMILES string of a ligand and an amino acid sequence of a target protein, extracting a ligand feature vector and a protein feature vector through a double-branch feature encoding network respectively; after the ligand feature vector and the protein feature vector are spliced and fused, a binding affinity value is predicted through a multilayer perceptron, a mean square error loss is calculated based on the binding affinity value, and the multilayer perceptron is trained and optimized; under the condition of the protein feature vector, a new molecular latent representation is generated through a reverse denoising process of a conditional diffusion model; wherein at each step of the reverse denoising process, a graph-level reading operation is performed on pure noise, a reading vector is input into the multilayer perceptron to predict an affinity value, a mean square error between the predicted affinity value and an actual value is calculated as an affinity guide loss, a gradient of the affinity guide loss is fed back to a noise prediction network to guide a new molecule to generate towards a high-affinity region; the new molecular latent representation is decoded into a SMILES string through a pre-trained molecular language decoder.

2. A method for molecular generation for drug discovery as claimed in claim 1 wherein, The double-branch feature encoding network comprises a parallel graph attention network and a graph convolution network for processing a molecular graph, and the molecular graph is obtained by parsing the SMILES string of the ligand using a chemical informatics tool.

3. A method for molecular generation for drug discovery as claimed in claim 1 wherein, The double-branch feature encoding network comprises a hierarchical structure comprising a convolutional neural network, a bidirectional long short-term memory network and a Transformer encoder for processing the amino acid sequence of the target protein.

4. The method of claim 1, wherein the method is used for drug discovery. The training of the noise prediction network adopts a weighted sum of the affinity guide loss and a noise prediction loss as a loss function.

5. A molecular generation system for drug discovery, characterized by, The method comprises the following steps: a feature encoding module configured to: obtain a SMILES string of a ligand and an amino acid sequence of a target protein, and extract a ligand feature vector and a protein feature vector through a double-branch feature encoding network respectively; an affinity prediction module configured to: after the ligand feature vector and the protein feature vector are spliced and fused, predict a binding affinity value through a multilayer perceptron, calculate a mean square error loss based on the binding affinity value, and train and optimize the multilayer perceptron; a molecular generation module configured to: under the condition of the protein feature vector, generate a new molecular latent representation through a reverse denoising process of a conditional diffusion model; wherein at each step of the reverse denoising process, a graph-level reading operation is performed on pure noise, a reading vector is input into the multilayer perceptron to predict an affinity value, a mean square error between the predicted affinity value and an actual value is calculated as an affinity guide loss, a gradient of the affinity guide loss is fed back to a noise prediction network to guide a new molecule to generate towards a high-affinity region; a decoding module configured to: decode the new molecular latent representation into a SMILES string through a pre-trained molecular language decoder.

6. A molecular generation system for drug discovery as claimed in claim 5, wherein, The double-branch feature encoding network comprises a parallel graph attention network and a graph convolution network for processing a molecular graph, and the molecular graph is obtained by parsing the SMILES string of the ligand using a chemical informatics tool.

7. A molecular generation system for drug discovery as claimed in claim 5 wherein, The double-branch feature encoding network comprises a hierarchical structure comprising a convolutional neural network, a bidirectional long short-term memory network and a Transformer encoder for processing the amino acid sequence of the target protein.

8. A molecular generation system for drug discovery as claimed in claim 5 wherein, The training of the noise prediction network employs a weighted sum of an affinity-guided loss and a noise prediction loss as a loss function.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The program, when executed by a processor, implements the steps in a method for molecule generation for drug discovery as claimed in any of claims 1-4.

10. A computer device, comprising a computer readable storage medium, a processor, and a computer program stored on the computer readable storage medium and executable on the processor, wherein, The program, when executed by a processor, implements the steps in a method for molecule generation for drug discovery as claimed in any of claims 1-4.

Images