Intelligent drug research and development molecular screening method
By constructing a molecule-target binary interaction graph and iteratively coupling it with a graph neural network, the multi-scale features of molecules and targets are calculated in real time. This solves the problems of excessive dependence on target structure and insufficient modeling of dynamic interactions in existing technologies, and achieves more efficient and accurate drug screening.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- AFFILIATED HOSPITAL OF NANTONG UNIV
- Filing Date
- 2026-01-17
- Publication Date
- 2026-04-28
AI Technical Summary
Existing computer-aided drug screening methods rely excessively on the structure of target proteins, making it difficult to accurately simulate the dynamic interactions between molecules and targets. This results in models having insufficient understanding of novel molecule-target pair binding modes and poor generalization ability.
A molecule-target binary interaction graph is constructed. Through dynamic interaction feature generation and iterative coupling with graph neural networks, multi-scale features of candidate molecules and target proteins are calculated in real time to generate exclusive dynamic interaction features for scoring and screening.
It enhances the ability to generalize the understanding of novel molecule-target pairs, improves the specificity and accuracy of predictions, and is applicable to scenarios where the target structure is incomplete or where there is an induced fit effect.
Smart Images
Figure CN121938499A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of interdisciplinary technology of drug development and artificial intelligence, and more specifically, to an intelligent molecular screening method for drug development. Background Technology
[0002] In the field of drug development, efficiently and accurately screening potential active candidate molecules from massive compound libraries targeting specific diseases is a crucial initial step in shortening the development cycle, reducing costs, and lowering failure rates. Traditional experimental screening methods have limited throughput and are costly; therefore, developing efficient and reliable computer-aided intelligent screening methods is of significant practical importance for accelerating the drug discovery process.
[0003] Existing computer-aided screening methods can be mainly divided into two categories. The first category is simulation methods based on molecular docking. This method heavily relies on the precise three-dimensional structure of the target protein and has poor applicability to targets with unknown structures or high flexibility. Its calculation process usually treats molecules as rigid or semi-rigid bodies, making it difficult to accurately simulate real conformational changes and induced fitting effects. The second category is methods based on machine learning or deep learning, such as using molecular fingerprints or graph neural networks to directly predict activity. Although these methods can utilize large amounts of data, they often simply splice or prematurely fuse the features of molecules and targets, failing to explicitly model the complex and dynamic interaction process between them. Many graph neural network-based models use static, pre-calculated features as edge attributes when constructing the molecule-target association graph. They cannot generate specific interaction representations in real time based on the specific pairing of molecules and targets during model inference, limiting the model's ability to understand and generalize the binding patterns of novel molecule-target pairs.
[0004] Therefore, this paper proposes an intelligent molecular screening method for drug development to address the above problems, which solves the problems of over-reliance on target protein structure, insufficient modeling of dynamic interactions between molecules and targets, and static characterization of interaction features in existing technologies. Summary of the Invention
[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide an intelligent drug development molecular screening method to address the problems raised in the background section.
[0006] To achieve the above objectives, the present invention provides the following technical solution: an intelligent drug development molecular screening method, comprising the following steps: S1: Obtain the candidate molecule set and the target protein set, encode the structural information of each candidate molecule and each target protein, and obtain the initial molecular feature representation of each candidate molecule and the initial protein feature representation of each target protein respectively. S2: Based on the candidate molecule set and target protein set, a bipartite interaction graph is constructed, wherein the candidate molecule node and the target protein node are initialized by the initial molecular feature representation and the initial protein feature representation, respectively, and an interaction edge to be predicted is established between each candidate molecule node and each target protein node. S3: For each interaction edge in the bipartite interaction graph, perform a dynamic interaction feature generation operation, which includes: extracting multi-scale molecular features of candidate molecular nodes and multi-scale protein features of target protein nodes connected by the edge; inputting the multi-scale molecular features and multi-scale protein features into an interaction attention module, which generates a dynamic interaction feature that characterizes the potential binding mode of the specific molecule-target pair by calculating the cross-scale attention weight of molecular features to protein features, and assigning the dynamic interaction feature to the corresponding interaction edge as its edge feature; S4: Input the bipartite interaction graph with edge features into a graph neural network; the graph neural network performs at least one round of message passing. In each round of message passing, a node receives information from its neighboring nodes. The weight of the information aggregation is determined by the dynamic interaction features of the edges connecting the node and its neighboring nodes, and the feature representation of each node is updated accordingly. S5: After the graph neural network completes message passing, for each interaction edge, based on its updated edge features, the updated features of the candidate molecule nodes it connects to, and the updated features of the target protein nodes, the interaction prediction score of the molecule-target pair is calculated through a scoring function. S6: Sort and filter all candidate molecule-target pairs based on the predicted scores of all interaction edges, and output the filtering results.
[0007] Preferably, in step S1, the structural information encoding of the candidate molecule specifically includes: if the candidate molecule has a three-dimensional structure, a geometric neural network is used to encode its atomic coordinates and types to obtain an initial molecular feature representation; if it does not have a three-dimensional structure, its molecular graph is input into a molecular graph neural network for encoding to obtain an initial molecular feature representation; the structural information encoding of the target protein specifically includes: if the target protein has a three-dimensional structure, a geometric neural network is used to encode its amino acid residue coordinates and types to obtain an initial protein feature representation; if it does not have a three-dimensional structure, its amino acid sequence is input into a pre-trained protein language model for encoding to obtain an initial protein feature representation.
[0008] Preferably, in step S3, the extraction of multi-scale molecular features includes: extracting feature vectors corresponding to the atomic scale, chemical bond scale, and substructure or functional group scale from the initial molecular feature representation of the candidate molecular nodes; the extraction of multi-scale protein features includes: extracting feature vectors corresponding to the amino acid residue scale, secondary structure element scale, and spatial binding pocket scale from the initial protein feature representation of the target protein nodes.
[0009] Preferably, the interactive attention module is a multi-head attention network, which concatenates features from various molecular scales as a query vector, concatenates features from various protein scales as a bond vector and a value vector, calculates the similarity between the query vector and the bond vector to obtain cross-scale attention weights, and uses these weights to perform a weighted summation of the value vectors, outputting the result as the dynamic interactive feature.
[0010] Preferably, in step S4, the graph neural network performs multiple rounds of message passing, and before performing the next round of message passing, the dynamic interaction feature generation operation described in step S3 is repeated, wherein the node features used to extract multi-scale features are updated to the latest node features obtained after the previous round of message passing, thereby generating updated dynamic interaction features and replacing the edge features of the corresponding edges, and then a new round of message passing is performed based on the updated edge features.
[0011] Preferably, the graph neural network is an edge-based graph attention network. In the message passing process of each layer, for the central node, the attention coefficient between it and each neighbor node is calculated. The calculation function of the attention coefficient takes the dynamic interaction features of the edge connecting the central node and the neighbor node as one of the key inputs.
[0012] Preferably, in step S5, the scoring function is specifically a multilayer perceptron, whose input vector is composed of three parts: the first part is the dynamic interaction feature after the interaction edge is updated, the second part is the feature after the candidate molecule node is updated, and the third part is the feature after the target protein node is updated.
[0013] Preferably, the method further includes a model training phase before screening, in which a labeled dataset containing known active molecule-target pairs and inactive molecule-target pairs is used to jointly optimize the parameters of the interactive attention module, the graph neural network, and the scoring function in an end-to-end manner, using the difference between the predicted score and the true active label as the loss function.
[0014] Preferably, after the method is executed for the first time and the screening results are output, it further includes an iterative optimization screening step: the molecular-target pairs with the top 5% to 20% of the predicted scores in the initial screening results are experimentally verified; the obtained experimental verification data is used as newly added labeled data to incrementally fine-tune the trained model; and the fine-tuned model is used to re-execute steps S1 to S6 on the remaining candidate molecules or the newly added candidate molecule set for a new round of screening.
[0015] Preferably, when the target protein has a three-dimensional structure, before extracting its spatial binding pocket-scale features, the algorithm first identifies potential binding pockets in its three-dimensional structure and calculates at least one of the geometric shape descriptor, electrostatic potential distribution, and hydrophobicity distribution of the inner surface of the pocket. The encoding of these descriptors is then fused with the local features obtained from the geometric neural network to form the feature vector at the spatial binding pocket scale.
[0016] The technical effects and advantages of this invention are as follows: Compared to existing technologies, this invention designs a dynamic interaction feature generation process for each interaction edge to be predicted within the constructed molecule-target bipartite graph. This process utilizes an interactive attention mechanism to calculate the cross-scale correlation between the multi-scale features of candidate molecules and the multi-scale features of target proteins in real time, and generates dynamic interaction features specific to that particular pair as edge attributes. This allows the model to explicitly infer the potential binding patterns for each specific molecule-target combination, rather than relying on fixed prior knowledge. This enhances the generalization ability for novel molecule-target pairs and improves the specificity and accuracy of predictions, especially suitable for scenarios where target structural information is incomplete or where induced fit effects exist.
[0017] Unlike existing graph neural network models that use static edge features, this invention iteratively couples dynamically generated interaction features with the message passing process of the graph neural network. Before each round of network propagation, more accurate dynamic interaction features are regenerated based on the updated node features, driving the next round of information aggregation. This design achieves the co-evolution of node and edge features, enabling the model to simulate the mutual adaptation and feature tuning process between molecules and targets as they approach their binding state. It overcomes the limitation of static features failing to reflect dynamic changes during the binding process, resulting in node and edge features that contain richer interaction context information, thus providing a more reliable and detailed basis for the final force scoring. Attached Figure Description
[0018] Figure 1 This is a flowchart illustrating the overall workflow of the method of the present invention.
[0019] Figure 2 This is a flowchart illustrating the core processing and optimization of the present invention.
[0020] Figure 3 This is a flowchart of the data encoding and feature initialization process of the present invention. Detailed Implementation
[0021] 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.
[0022] Example 1 As attached Figures 1 to 3 The method described is an intelligent drug discovery molecule screening approach. The process is as follows: First, the method structurally encodes the input candidate molecules and target proteins, transforming them into initial feature representations that can be processed by a machine. Based on these features, a bipartite interaction graph is constructed, connecting each candidate molecule with the target protein. Each edge in the graph represents a potential interaction to be evaluated.
[0023] Then, the interaction modeling stage begins: for each edge, a dynamic interaction feature is calculated in real time. This feature is not fixed, but is specifically generated in real time through an attention mechanism for the specific molecules and targets at both ends of the edge, in order to capture their unique potential binding patterns.
[0024] This dynamic feature is then injected into a graph neural network, driving the node features to be updated in a context-aware manner in multiple iterations. Each node update feeds back into the generation of more accurate dynamic interaction features, forming a closed loop of co-evolution.
[0025] Finally, using the optimized edge and node features, a scoring function is used to calculate the predicted probability of the interaction, and the sorting and filtering are completed accordingly. This achieves a computational simulation of the dynamic characteristics of mutual induction and adaptation between molecules and target sites during the binding process.
[0026] Furthermore, in the data encoding and initialization steps, a differentiated encoding strategy is adopted for candidate molecules based on data availability.
[0027] For molecules possessing a reliable three-dimensional conformation (e.g., the lowest energy conformation obtained through molecular force field optimization), the following is employed: - Encoding is performed using an equivariant graph neural network. The output of this network is equivariant to the rotation and translation of the input molecule, meaning that the learned feature representation is semantically consistent regardless of how the molecule is positioned in three-dimensional space.
[0028] Its inter-layer update rule can be formally expressed as: for atoms In its first Layer coordinates and characteristics The update incorporates data from neighboring atoms. Information: here, It is the generated edge message vector, which contains not only atomic features It also includes the square of the relative distance between atoms. and key types Information such as... The function generates a scalar weight based on the message vector, applies it to the distance vector, and thus performs an equal-variable update on the coordinates. The aggregated message is then transformed to update the atomic features. Finally, the features of all atoms are pooled (e.g., global average pooling) to form the initial feature representation of the molecule. .
[0029] For molecules with only a two-dimensional topology (SMILES strings or molecular graphs), a standard molecular graph neural network is used. Its message passing does not involve coordinate updates; it focuses only on feature propagation of atoms and chemical bonds, ultimately obtaining the desired result through pooling. .
[0030] For target proteins, if there is an experimentally resolved crystal structure or a structure output by a high-precision prediction model (such as AlphaFold2), a protein-optimized geometric neural network (such as GVP-GNN) is used for processing. The principle is similar, but it is specially designed for the complex side chain and main chain geometry of amino acid residues.
[0031] If only amino acid sequences are available, a language model pre-trained on a large-scale protein sequence library (such as ESM-2) is used. This model, through self-supervised learning on billions of sequences, can extract implicit structural, evolutionary, and functional information from the sequences, outputting the contextual embedding of each residue. Then, the initial feature representation of the target is obtained through pooling or focusing on preset active site regions. This differentiated encoding scheme ensures that regardless of the completeness of the input data, it can be transformed into high-quality, information-rich feature vectors.
[0032] Furthermore, after constructing the bipartite graph and obtaining the initial node features, in order to perform fine-grained interactive analysis, it is necessary to parse sub-information at different scales from these overall features, which is achieved through multiple parallel, lightweight feature mapping modules.
[0033] For each molecular node, its initial characteristics It is fed into three independent small feedforward neural networks (typically - (layers), each mapping atomic-scale features. Chemical bond scale characteristics Synthesis / functional group scale features .
[0034] Atomic-scale features aim to preserve the abstract expression of each atom's own chemical properties (such as element type, hybridization state, and formal charge); chemical bond-scale features focus on the connection relationships and bonding properties between atoms; and substructure-scale features attempt to capture fragment information with specific chemical or pharmacological significance, such as benzene rings, carboxyl groups, and amide bonds.
[0035] Similarly, for each target protein node, its initial features It is also mapped to three scales: amino acid residue scale features (Reflecting the identity, hydrophilicity / hydrophobicity, charge, etc. of individual residues), secondary structure scale characteristics (reflect spiral, (Regular local compositional patterns such as folding), and spatial combination with pocket-scale features. (Reflecting the overall environment such as the spatial shape of the binding site, electrostatic distribution, and hydrophobic patches). This explicit multi-scale decoupling is a prerequisite for the subsequent generation of dynamic interaction features, enabling the model to learn cross-scale association rules such as "a specific functional group of the molecule (substructure scale) tends to electrostatically interact with a charged residue in the target pocket (residue scale)," rather than performing fuzzy overall matching.
[0036] Furthermore, in the dynamic interaction feature generation step, the goal is to generate features for each connecting molecule in the graph. and target For each edge, calculate a unique feature vector. .
[0037] This feature is generated through a multi-head attention mechanism, where molecules are analogous to query roles and target points are analogous to the key-value pairs being retrieved. Specifically, the multi-scale features of molecules are concatenated into a query vector. The multi-scale features of the target points are concatenated into a key vector. Sum value vector .
[0038] The computation of the attention mechanism reveals information about a specific molecule, its target scale, and the most relevant features. This relevance is calculated using dot product similarity and normalized to obtain the attention weights. In a multi-head setup, the... The process of calculating the size is as follows: In the formula, These are learnable linear projection matrices that map input features to different representation subspaces; It is the dimension of the key vector. It is a scaling factor used to stabilize the gradient. The function ensures that the sum of all weights is .
[0039] The calculated attention weight matrix, where each element essentially represents the degree of attention paid by a "feature at a certain scale of the molecule" to a "feature at a certain scale of the target." Weighting the target's value vector with this weight matrix and summing the weights means extracting and combining the most relevant summary from the target's features based on the molecule's query. The outputs of multiple heads are concatenated and finally passed through an output linear layer. Fusion: This is how it was generated. It possesses a high degree of specificity and dynamism: specificity because it is entirely composed of molecules. and target The characteristics of each node determine its own characteristics; the dynamic nature arises because if the node characteristics are updated in a subsequent process, re-executing the process will yield different results. .
[0040] Furthermore, the optimization process of the graph neural network is tightly coupled with dynamic interaction features, forming an iterative feedback loop. Assume the preset number of iterations is... (usually) - At the initial moment ( The first edge feature is generated using the initial node features. In the first In rounds of iteration ( First, message passing in the graph attention network is performed.
[0041] For a molecular node (Its neighbors are only target nodes) It receives information from the target node. Aggregate weights. The calculation is not simply based on node feature similarity, but explicitly incorporates edge features. : here, It is a weight matrix. It is the attention vector, and || represents vector concatenation. Nonlinearity is introduced. This design ensures that the strength of message passing depends not only on the states of the source node (target) and the target node (molecule) themselves, but also directly on the dynamic interaction pattern between them inferred in the previous round. The stronger the interaction pattern (potentially indicating a higher tendency to combine), the greater the weight of information transmission. The updated characteristics of the molecular node are: in It is a non-linear activation function. For the target node, the update process is similar, but it aggregates information from all connected molecular nodes.
[0042] Once all nodes have completed this round of updates, the system will not immediately proceed to the next round of graph propagation, but will instead restart the dynamic interactive feature generation process.
[0043] At this point, use the newly updated node features. and (These already contain contextual information from the first round of graph propagation), and multi-scale feature extraction and attention computation are performed again to generate updated, theoretically more accurate dynamic interaction features. This new feature will replace the old feature and will be used in the next round. Weight calculation in graph attention networks.
[0044] This iterative process of "node feature update → edge feature refresh → node feature update again" simulates the dynamic process of molecules and targets probing and adjusting each other until a stable state is reached during virtual binding. Through such iterations, the final node features are obtained. Sum of edge features It is the result of multiple collaborative optimizations and contains rich interaction context information.
[0045] Furthermore, the specific architecture of the graph neural network is implemented using an edge-based graph attention network. Its core lies in the design of the message passing function. As mentioned above, using the dynamic features of the edges as the key input for calculating the attention coefficient helps handle highly imbalanced graph structures like virtual screening: a target node may be connected to hundreds of thousands or even millions of molecular nodes, introducing dynamic interaction features. Then, attention weights The calculation yielded additional, highly relevant discriminative information. Because This is already a preliminary, local encoding of the binding probability of a specific molecule-target pair. Using it to guide attention is equivalent to enabling the graph network to prioritize connections with stronger local interaction signals when aggregating global information. This allows the feature updates of target nodes to more effectively focus on potential molecular subsets, improving the efficiency of information aggregation and the signal-to-noise ratio.
[0046] Furthermore, after completing the iterative optimization of the graph network, a final interaction score needs to be performed on each edge. The scoring function is a multilayer perceptron, and its input is a joint vector that integrates the final node state and edge state. : This splicing operation integrates three types of information: (1) A refined dynamic interaction mode that has undergone multiple rounds of collaborative optimization; (2) Molecular features taking into account the target context; (3) Target features that take into account the context of all candidate molecules.
[0047] Then, a typical MLP is used: in For learnable parameters, For activation function, Compress the output to Intervals, as predicted probabilities of interaction. This ensures that the scoring incorporates evidence from both the direct interaction model and the global graph environment, making the predictions more robust.
[0048] Furthermore, the entire system is trained as a differentiable computational graph. The training data consists of known active molecule-target pairs (positive samples) and inactive pairs (negative samples) constructed through random sampling or based on similarity strategies. Binary cross-entropy is used as the loss function. For a batch of training data, the loss is... The calculation is as follows: in It is a training batch, containing both positive and negative samples. It is a real label ( Indicates activity, (Representing inactivity). Through backpropagation, gradients are backpropagated layer by layer from the loss function, simultaneously updating the parameters of the scoring function MLP, graph attention network, dynamic interactive feature generation module (attention head parameters), and the initial encoder. Joint optimization forces all modules to work together, jointly learning the entire mapping from the original structured data to the final activity prediction, ensuring the consistency of internal representations.
[0049] Furthermore, this method can be deployed in an iterative active learning loop to optimize screening efficiency.
[0050] After the initial full database screening, the predicted scores are selected. highest molecule (e.g.) Further in vitro experiments were conducted to verify this. After obtaining the actual activity / inactivity data of each molecule, these data become new labeled samples for that specific target.
[0051] Subsequently, the entire model is incrementally fine-tuned in a few epochs using these new samples (often combined with some of the original data to prevent catastrophic forgetting) at a small learning rate. This process allows the model to quickly absorb experimental feedback and adjust its internal parameters, potentially enabling it to better capture the target-specific chemical spatial preferences or subtle interaction patterns.
[0052] The fine-tuned model was then used to perform a new round of prediction screening on the remaining compound library or a completely new molecular library. Because the model had been specifically trained, its prediction accuracy in the new round of screening was improved, especially in identifying molecules with similar structures or mechanisms of action to those in the first round of matching.
[0053] Furthermore, when the target protein has a three-dimensional structure, its "binding pocket-scale characteristics" are also observed. Extraction requires specific computational chemistry methods.
[0054] The process begins by scanning the surface of the target protein using pocket detection algorithms (such as FPocket or DoGSiteScorer) to identify potential ligand-binding cavities. For the primary pocket of interest, the physicochemical fields of its surface need to be calculated. For example, calculating the electrostatic potential typically involves solving the Poisson-Boltzmann equation. Under a simplified model, the pocket's spatial grid points can be approximated using Coulomb's law. electrostatic potential at : in, It is a protein atom Partial charge (unit: elementary charge unit) ), It is an atom To grid point Distance (usually in units of) ), It is a distance-dependent function of the dielectric constant, used to simulate solvent effects.
[0055] Simultaneously, surface curvature, hydrophobicity index, and other parameters can be calculated, and the resulting field data are converted into grid features or statistical descriptors. On the other hand, from the atomic / residue-level features encoded by the protein geometric neural network, a subset of features corresponding to residues within the pocket can be extracted and aggregated (e.g., average pooling).
[0056] final, The pocket features are formed by concatenating and fusing the calculated materialized descriptor vector with the local structural feature vector extracted by the neural network. This fusion ensures that the pocket features contain both physically based, interpretable global properties and data-driven local information that can capture subtle structural patterns.
[0057] To more clearly illustrate the complete workflow of this technical solution, an end-to-end explanation is provided below using a hypothetical application scenario. Assume the task is to analyze a database containing approximately 100,000 molecules, targeting a substance called "kinase". Seed compounds were screened for targets (known amino acid sequences, but no high-resolution three-dimensional structures).
[0058] Step 1: Data Encoding and Graph Construction. The system reads in the SMILES strings and kinases of 100,000 molecules. The FASTA sequence.
[0059] For all molecules, since no 3D structure was provided, a pre-trained molecular graph neural network (MPNN) was used for encoding, resulting in 100,000 initial molecular feature vectors. .
[0060] For kinases Its sequence is obtained through ESM- Model encoding yields the initial target feature vector. The system then constructs a bipartite graph containing 100,000 molecular nodes. and a target node .
[0061] At each molecular node With target node Create an edge between them There are a total of 100,000 edges, and each node is initialized with its initial feature vector.
[0062] Step 2: Initial generation of dynamic interaction features. For each edge... The system executes: (1) From Extracting multi-scale features ; (2) From Extracting multi-scale features (in Based on ESM- (Predicted residue conservation or simple sequence template estimation). (3) Input these features into the interaction attention module. (Using the edge...) (corresponding molecule) Taking (e.g.) as an example, calculate its initial dynamic interaction characteristics. : This process is executed in parallel 100,000 times to generate initial values for all edges. .
[0063] Step 3: Iterative graph optimization (with...) (Taking wheels as an example).
[0064] Iteration rounds ( ): Graph propagation: For molecular nodes It starts from the unique neighbor target node. Receive information. Calculate attention weights: ,because With only one neighbor, weight However, the calculations in the formula are still meaningful because the parameters are shared, and The value of this value will affect subsequent losses. Node update: ; Meanwhile, target nodes Information from 100,000 molecular nodes needs to be aggregated. From molecules The weight of the received information is calculated as follows: ; ; The updated features are: This step enables target characterization. It became a weighted summary of all molecular features, with weights determined by the initial interaction features. guide.
[0065] Edge feature refresh: Update all node features to and Then, the system immediately regenerates dynamic interaction features for each edge. For each edge... : in Depend on The newly decoupled multi-scale features constitute Depend on The newly decoupled multi-scale features constitute the structure.
[0066] Iteration rounds ( ): Graph propagation: using refreshed edge features Proceed to the second round of message passing.
[0067] Molecular nodes Update: Target node Update: ; Obtaining final features: After the second round, obtain the final node features. and final edge features .
[0068] Step 4: Scoring and Filtering. For each edge Construct a joint vector and score it. For example: The system performs parallel computation to obtain 100,000 prediction scores. Subsequently, according to All molecules are sorted from highest to lowest, and the IDs and scores of the top 1000 molecules are output as the preliminary results of this virtual screening.
[0069] Step 5: Iterative Optimization (Optional). Researchers select the top 100 molecules for experimental testing. Assuming feedback is received after several weeks: 12 molecules show activity, and 88 are inactive. The system will then analyze these 100 new molecules (kinases). (Label) data pairs are added to the training set, and based on the original pre-trained model, [the following is performed / processed]. - The model undergoes several rounds of fine-tuning training. After fine-tuning, researchers can apply the adjusted model to a dataset containing... An expanded library of 10,000 new molecules is created by executing the complete process from step one to step four above, with the aim of using the learned knowledge to discover a new batch of active molecules more efficiently.
[0070] Finally, the following points should be noted: First, in the description of this application, it should be noted that, unless otherwise specified and limited, the terms "installation", "connection", and "linkage" should be interpreted broadly, and can be mechanical or electrical connections, or internal connections between two components, or direct connections. "Up", "down", "left", "right", etc. are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may change. Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other. In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A smart molecular screening method for drug development, characterized in that, Includes the following steps: S1: Obtain the candidate molecule set and the target protein set, encode the structural information of each candidate molecule and each target protein, and obtain the initial molecular feature representation of each candidate molecule and the initial protein feature representation of each target protein respectively. S2: Based on the candidate molecule set and target protein set, a bipartite interaction graph is constructed, wherein the candidate molecule node and the target protein node are initialized by the initial molecular feature representation and the initial protein feature representation, respectively, and an interaction edge to be predicted is established between each candidate molecule node and each target protein node. S3: For each interaction edge in the bipartite interaction graph, perform a dynamic interaction feature generation operation, which includes: extracting multi-scale molecular features of candidate molecular nodes and multi-scale protein features of target protein nodes connected by the edge; inputting the multi-scale molecular features and multi-scale protein features into an interaction attention module, which generates a dynamic interaction feature that characterizes the potential binding mode of the specific molecule-target pair by calculating the cross-scale attention weight of molecular features to protein features, and assigning the dynamic interaction feature to the corresponding interaction edge as its edge feature; S4: Input the bipartite interaction graph with edge features into a graph neural network; the graph neural network performs at least one round of message passing. In each round of message passing, a node receives information from its neighboring nodes. The weight of the information aggregation is determined by the dynamic interaction features of the edges connecting the node and its neighboring nodes, and the feature representation of each node is updated accordingly. S5: After the graph neural network completes message passing, for each interaction edge, based on its updated edge features, the updated features of the candidate molecule nodes it connects to, and the updated features of the target protein nodes, the interaction prediction score of the molecule-target pair is calculated through a scoring function. S6: Sort and filter all candidate molecule-target pairs based on the predicted scores of all interaction edges, and output the filtering results.
2. The intelligent drug development molecular screening method according to claim 1, characterized in that, In step S1, the structural information encoding of candidate molecules specifically includes: if a candidate molecule has a three-dimensional structure, a geometric neural network is used to encode its atomic coordinates and types to obtain an initial molecular feature representation; if it does not have a three-dimensional structure, its molecular graph is input into a molecular graph neural network for encoding to obtain an initial molecular feature representation; the structural information encoding of target proteins specifically includes: if a target protein has a three-dimensional structure, a geometric neural network is used to encode its amino acid residue coordinates and types to obtain an initial protein feature representation; if it does not have a three-dimensional structure, its amino acid sequence is input into a pre-trained protein language model for encoding to obtain an initial protein feature representation.
3. The intelligent drug development molecular screening method according to claim 2, characterized in that, In step S3, the extraction of multi-scale molecular features includes: extracting feature vectors corresponding to the atomic scale, chemical bond scale, and substructure or functional group scale from the initial molecular feature representation of the candidate molecular nodes; the extraction of multi-scale protein features includes: extracting feature vectors corresponding to the amino acid residue scale, secondary structure element scale, and spatial binding pocket scale from the initial protein feature representation of the target protein nodes.
4. The intelligent drug development molecular screening method according to claim 3, characterized in that, The interactive attention module is a multi-head attention network that concatenates features from various molecular scales as a query vector and concatenates features from various protein scales as a bond vector and a value vector. It calculates the cross-scale attention weights by calculating the similarity between the query vector and the bond vector, and uses these weights to perform a weighted summation of the value vectors. The output result is the dynamic interactive feature.
5. The intelligent drug development molecular screening method according to claim 4, characterized in that, In step S4, the graph neural network performs multiple rounds of message passing, and before performing the next round of message passing, the dynamic interaction feature generation operation described in step S3 is repeated, wherein the node features used to extract multi-scale features are updated to the latest node features obtained after the previous round of message passing, thereby generating updated dynamic interaction features and replacing the edge features of the corresponding edges, and then a new round of message passing is performed based on the updated edge features.
6. The intelligent drug development molecular screening method according to claim 1 or 5, characterized in that, The graph neural network is an edge-based graph attention network. In the message passing process of each layer, for the central node, the attention coefficient between it and each neighbor node is calculated. The calculation function of the attention coefficient takes the dynamic interaction features of the edge connecting the central node and the neighbor node as one of the key inputs.
7. The intelligent drug development molecular screening method according to claim 6, characterized in that, In step S5, the scoring function is specifically a multilayer perceptron, whose input vector is composed of three parts: the first part is the dynamic interaction feature after the interaction edge is updated, the second part is the feature after the candidate molecule node is updated, and the third part is the feature after the target protein node is updated.
8. The intelligent drug development molecular screening method according to claim 1, characterized in that, Before being used for screening, the method further includes a model training phase, which uses a labeled dataset containing known active molecule-target pairs and inactive molecule-target pairs, and uses the difference between the predicted score and the true active label as a loss function to jointly optimize the parameters of the interactive attention module, the graph neural network, and the scoring function.
9. The intelligent drug development molecular screening method according to claim 1, characterized in that, After the method is executed for the first time and the screening results are output, it also includes an iterative optimization screening step: the molecular-target pairs with the top 5% to 20% of the predicted scores in the initial screening results are experimentally verified; the obtained experimental verification data is used as newly added labeled data to incrementally fine-tune the trained model. The fine-tuned model is used to repeat steps S1 to S6 on the remaining candidate molecules or the newly added candidate molecule set for a new round of screening.
10. The intelligent drug development molecular screening method according to claim 2, characterized in that, When the target protein has a three-dimensional structure, before extracting its spatial binding pocket-scale features, the algorithm first identifies potential binding pockets in its three-dimensional structure and calculates at least one of the geometric shape descriptor, electrostatic potential distribution, and hydrophobicity distribution of the inner surface of the pocket. The encoding of these descriptors is then fused with the local features obtained from the geometric neural network to form the feature vector at the spatial binding pocket scale.