A drug molecule property prediction method, system, device and computer medium

Abstract

The application discloses a drug molecule property prediction method, system, device and computer medium, relates to the technical field of artificial intelligence assisted drug discovery and molecular graph neural network modeling, and constructs a virtual interaction edge between each functional group pair in a functional group set of a target drug molecule according to a preset strong rule interaction knowledge base and a weak rule scoring function. The virtual interaction edge is used for representing non-covalent interaction between physically unconnected functional groups; the virtual interaction edge is added to an original molecular graph to generate an enhanced molecular graph; a dual-channel encoder containing a physical bond channel and a virtual interaction channel is applied to perform encoding processing on the enhanced molecular graph to obtain a graph representation; a SMILES sequence is encoded to obtain a sequence representation; cross attention fusion is performed on the graph representation and the sequence representation to obtain a fusion feature vector; and a parameter set is applied to perform property prediction on the fusion feature vector to obtain a property prediction result. The prediction accuracy and generalization are high.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence-assisted drug discovery and molecular graph neural network modeling technology, and more specifically, to a method, system, device and computer medium for predicting the properties of drug molecules. Background Technology

[0002] In drug discovery, accurately predicting the physicochemical properties and biological activities of candidate molecules is a core step in achieving high-throughput virtual screening. Experimental determination is time-consuming and labor-intensive, making it difficult to support rapid evaluation of large-scale compound libraries. Therefore, efficient and accurate computational prediction methods are urgently needed.

[0003] Currently, molecular property prediction can be performed using Graph Neural Networks (GNNs). However, the molecular graphs generated by these methods only encode covalent bond connections between atoms. For example, the electron push-pull effect between an amino group and a nitro group on a para-substituted benzene ring requires at least four hops of propagation to achieve information exchange in the molecular graph. Due to its over-smoothing properties, GNNs typically only stack three to four layers effectively, leading to severe attenuation of long-range chemical effect signals during propagation. Although some studies have attempted to incorporate chemical knowledge, they all employ late-stage fusion strategies, meaning that chemical knowledge does not participate in the underlying message passing and fails to fundamentally change the receptive field and information propagation path of GNNs. This results in poor accuracy and generalization ability in predicting drug molecule properties, hindering subsequent drug discovery.

[0004] In summary, improving the accuracy and generalization ability of drug analysis property prediction is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] The purpose of this application is to provide a method for predicting the properties of drug molecules, which can, to some extent, address the technical problem of how to improve the accuracy and generalization ability of predicting drug properties. This application also provides a drug molecule property prediction system, electronic equipment, and a computer-readable storage medium.

[0006] To achieve the above objectives, this application provides the following technical solution: A method for predicting the molecular properties of a drug includes: Obtain the original molecular map and SMILES sequence of the target drug molecule; Identify the set of functional groups in the target drug molecule; Based on a preset strong rule interaction knowledge base and weak rule scoring function, virtual interaction edges are constructed between each functional group pair in the functional group set. The virtual interaction edges are used to characterize the non-covalent interaction between physically unconnected functional groups. The virtual interactive edges are added to the original molecular graph to generate an enhanced molecular graph; A dual-channel encoder, including physical bond channels and virtual interaction channels, is used to encode the enhanced molecular map to obtain a graphical representation; The SMILES sequence is encoded to obtain a sequence representation; Cross-attention fusion is performed on the graph representation and the sequence representation to obtain a fused representation vector; The properties of the fused representation vector are predicted by applying the pre-trained parameter set, thereby obtaining the property prediction results of the target drug molecule.

[0007] Preferably, the step of constructing virtual interaction edges between pairs of functional groups in the functional group set based on a preset strong-rule interaction knowledge base and a weak-rule scoring function includes: Iterate through all functional group pairs in the functional group set and determine whether the category combination of the current functional group pair matches the non-covalent interaction in the strong rule interaction knowledge base. If a hit occurs, a typed virtual interaction edge is created between the current functional group pairs in the set of functional groups. If no match is found, an interaction potential score is generated based on the physicochemical properties of the current functional group pair, and it is determined whether the interaction potential score is greater than or equal to a preset weak rule threshold. If the interaction potential score is greater than or equal to the weak rule threshold, then an implicit virtual interaction edge is created between the current functional group pairs in the functional group set. The non-covalent interactions include: electron donation and acceptance synergy, hydrogen bond network enhancement, hydrophobic cluster synergy, metal chelation synergy, steric hindrance antagonism, and acid-base neutralization.

[0008] Preferably, the generation of interaction potential scores based on the physicochemical properties of the current functional group pairs includes: The interaction potential score is obtained by weighting and summing the electronegativity difference, polarizability difference, hydrogen bond donor-acceptor complementarity, spatial matching degree, and distance decay factor of the current functional group pair.

[0009] Preferably, the virtual interaction channel is used for message passing on the virtual interaction edges in the enhanced molecular graph. The application includes a dual-channel encoder with a physical bond channel and a virtual interaction channel. The encoding process of the enhanced molecular graph includes: Configure an initial confidence level and a layer decay coefficient for each virtual interaction edge, wherein the layer decay coefficient is distinguished according to the interaction type; In the execution of the virtual interaction channel l During layer message passing, the current confidence of the virtual interaction edge is updated based on the product of the initial confidence, the layer decay coefficient, and the task adaptive attention weight. Determine if the current confidence level is lower than the preset confidence pruning threshold. If so, then at the... l Skip virtual interaction edges in message passing within the layer.

[0010] Preferably, the physical bond channel is used for message passing of covalent bonds in the enhanced molecular graph. The application includes a dual-channel encoder with a physical bond channel and a virtual interaction channel. The encoding process of the enhanced molecular graph includes: Obtain the physical key channel output, virtual interaction channel output, and the representation of the previous node of the current node; Generate the gating value for the current node based on the physical key channel output, virtual interaction channel output, and the representation of the previous node. The physical key channel output and the virtual interaction channel output are weighted and summed according to the gating value to obtain the weighted summation result; Perform residual connection and layer normalization processing on the weighted summation result to obtain the fusion node representation of the current layer; In the application of the dual-channel encoder, the upper bound of the mutual information between the output of the physical key channel and the output of the virtual interaction channel is minimized by mutual information regularization constraints.

[0011] Preferably, the step of performing cross-attention fusion processing on the graph representation and the sequence representation to obtain a fused representation vector includes: The graph representation and the sequence representation are respectively projected onto a shared fusion space; The graph representation is queried in the sequence representation, and the sequence representation is queried in the graph representation to obtain the graph representation after cross-enhancement and the sequence representation after cross-enhancement, respectively. Based on the cross-enhanced graph representation and the cross-enhanced sequence representation, a fusion gate value is generated; Based on the fusion gate value, the cross-enhanced graph representation and the cross-enhanced sequence representation are weighted to obtain the weighted processing result; The weighted result is concatenated with the unweighted graph representation after cross-enhancement and the sequence representation after cross-enhancement to obtain the concatenated result; The splicing result is mapped to obtain the fused representation vector.

[0012] Preferably, before applying the pre-trained parameter set to predict the properties of the fused representation vector, the method further includes: Obtain the training sample set; Random deletion perturbation is performed on the virtual interaction edges of the training sample set, and the molecular property prediction model is trained using the binary classification of virtual edge existence as the training task. Random deletion perturbation and interaction type tampering perturbation are performed on the virtual interaction edges of the training sample set, and the molecular property prediction model is trained with multi-class classification of interaction types as the training task. Random deletion perturbation, interaction type tampering perturbation, and erroneous routing perturbation are performed on the virtual interaction edges of the training sample set, and functional group routing consistency contrastive learning is used as the training task to train the molecular property prediction model. Obtain the parameter set of the trained molecular property prediction model.

[0013] A drug molecule property prediction system, comprising: The data acquisition module is used to acquire the original molecular map and SMILES sequence of the target drug molecule; A functional group recognition module is used to identify the set of functional groups in the target drug molecule; A creation module is used to construct virtual interaction edges between pairs of functional groups in the functional group set based on a preset strong rule interaction knowledge base and weak rule scoring function. The virtual interaction edges are used to characterize the non-covalent interactions between physically unconnected functional groups. An addition module is used to add the virtual interactive edges to the original molecular graph to generate an enhanced molecular graph; The graph encoding module is used to encode the enhanced molecular graph using a dual-channel encoder that includes physical bond channels and virtual interaction channels to obtain a graph representation; A sequence encoding module is used to encode the SMILES sequence to obtain a sequence representation; The fusion module is used to perform cross-attention fusion on the graph representation and the sequence representation to obtain a fused representation vector; The prediction module is used to apply a pre-trained parameter set to predict the properties of the fused representation vector, thereby obtaining the property prediction results of the target drug molecule.

[0014] An electronic device, comprising: Memory, used to store computer programs; A processor for executing the computer program to implement the steps of the drug molecule property prediction method as described above.

[0015] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of any of the above-described methods for predicting the properties of drug molecules.

[0016] This application provides a method for predicting the properties of drug molecules, which involves: acquiring the original molecular graph and SMILES sequence of the target drug molecule; identifying the set of functional groups in the target drug molecule; constructing virtual interaction edges between pairs of functional groups in the functional group set based on a pre-set strong-rule interaction knowledge base and weak-rule scoring function, whereby the virtual interaction edges are used to characterize the non-covalent interactions between physically unconnected functional groups; adding the virtual interaction edges to the original molecular graph to generate an enhanced molecular graph; applying a dual-channel encoder containing physical bond channels and virtual interaction channels to encode the enhanced molecular graph to obtain a graph representation; encoding the SMILES sequence to obtain a sequence representation; performing cross-attention fusion on the graph representation and the sequence representation to obtain a fused representation vector; and applying a pre-trained parameter set to predict the properties of the fused representation vector to obtain the property prediction results of the target drug molecule. In this application, by constructing virtual interaction edges to characterize non-covalent interactions between physically disconnected functional groups and adding them to the original molecular graph to generate an enhanced molecular graph, the graph neural network can directly perceive the effects between functional groups during message passing. This overcomes the inherent limitation of traditional GNNs, which cannot effectively stack deep networks due to oversmoothing, and expands the model's receptive field to long-range chemical effects. Furthermore, by using physical bond channels and virtual interaction channels with independent parameters to process covalent and non-covalent interaction information respectively, feature confusion caused by mixing and encoding two types of chemical information with different physical natures is avoided. By fusing graph representation and sequence representation through cross-attention, deep interaction and complementary enhancement of local topological information and global sequence semantic information are achieved. Combined with the general molecular representation prior knowledge provided by the pre-trained parameter set, better generalization performance can be obtained in tasks with limited data, improving the prediction accuracy and generalization of drug molecule properties. The drug molecule property prediction system, electronic device, and computer-readable storage medium provided in this application also solve the corresponding technical problems. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0018] Figure 1 A flowchart of a method for predicting the properties of drug molecules provided in this application embodiment; Figure 2 This is an architecture diagram of the dual-channel GNN encoder in this application; Figure 3 This is a schematic diagram of confidence decay and soft pruning; Figure 4A schematic diagram illustrating the cross-attention fusion of graph representation and sequence representation provided in an embodiment of this application; Figure 5 A graph showing the loss function values ​​for pre-training in a three-stage progressive course. Figure 6 This is a schematic diagram of the structure of a drug molecule property prediction system provided in an embodiment of this application; Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application; Figure 8 This is another structural schematic diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] Please see Figure 1 , Figure 1 A flowchart illustrating a method for predicting the molecular properties of a drug provided in this application embodiment.

[0021] This application provides a method for predicting the molecular properties of drugs, which may include the following steps: Step S101: Obtain the original molecular map and SMILES sequence of the target drug molecule.

[0022] In practical applications, the original molecular graph and SMILES (Simplified Molecular Input Line Entry System) sequence of the target drug molecule can be obtained first. This allows the chemical structure of the drug molecule to be transformed into numerical graph structural data that can be processed by a graph neural network, while preserving the chemical properties of atoms and the connection information of chemical bonds. This provides complete and accurate structural information input for subsequent molecular property prediction. Here, "target drug" refers to the drug to be evaluated and can be flexibly set according to the application scenario.

[0023] In an exemplary embodiment, the SMILES string representation of the target drug molecule can be obtained first. The SMILES string is a linear encoding standard that uses ASCII characters to represent molecular structures. For example, the SMILES representation of a benzene molecule is c1ccccc1. Different molecules have unique SMILES representations. Therefore, SMILES can be used as a universal identifier and input format for molecules. Then, the SMILES string is parsed using cheminformatics toolkits such as RDKit to construct the corresponding original molecular diagram. In other words, after obtaining the SMILES string, it is parsed using the RDKit toolkit. RDKit is an open-source cheminformatics toolkit that provides functions such as creating molecular objects from SMILES strings, extracting atomic and chemical bond information, and calculating molecular descriptors. The parsing process is as follows: First, RDKit functions are called to decompose the SMILES string into tokens such as atomic symbols, chemical bond symbols, branch symbols, and ring symbols, and a molecular graph structure is constructed according to the SMILES syntax rules. Then, an atomic list and a chemical bond list are extracted from the RDKit molecular object. Each atom corresponds to a node (Vertex) in the graph, and each chemical bond corresponds to an edge (Edge) in the graph. Next, a 35-dimensional initial feature vector is extracted for each atomic node. Atomic features include atom type, atomicity, formal charge, number of free radical electrons, hybridization mode, whether it is in a ring, and chirality type. Finally, a 6-dimensional edge feature vector is extracted for each chemical bond. Chemical bond features include bond type, whether it is conjugated, and whether it is in a ring. Through the above processing, the original molecular graph is obtained. G o ( V , E phys It is constructed as an attribute graph with node features and edge features, where V It is a set of 35-dimensional atomic nodes. E phys It is a 6-dimensional chemical bond feature vector.

[0024] Step S102: Identify the set of functional groups in the target drug molecule.

[0025] In practical applications, it is also necessary to identify the functional group set of the target drug molecule to accurately identify and locate the key structural units of the functional groups in the drug molecule. This provides precise anchor points for the subsequent construction of virtual interaction edges, ensuring that the virtual interaction edges accurately reflect the non-covalent interactions between functional groups. For example, based on a predefined functional group SMARTS pattern library, substructure matching can be performed on the target drug molecule to identify each functional group present in the molecule, and a core atom can be determined for each functional group as the anchor point of the virtual interaction edge. Functional groups are key structural units that determine the chemical properties and biological activities of molecules. For example, the carboxyl group (-COOH) imparts acidity to the molecule, the primary amino group (-NH2) imparts basicity, and the hydroxyl group (-OH) imparts the ability to form hydrogen bonds. Different functional groups influence the overall properties of the molecule through non-covalent interactions such as hydrogen bonds, hydrophobic interactions, and electrostatic interactions. Therefore, accurately identifying the functional groups and their spatial distribution in the molecule is a prerequisite for constructing virtual interaction edges.

[0026] In an exemplary embodiment, during the functional group identification process, a SMARTS (SMILES Arbitrary Target Specification) functional group pattern library can be constructed first. SMARTS is a specification language for describing molecular substructure patterns, extended from the SMILES syntax, and can cover nine categories of functional groups: acidic groups, basic groups, hydrogen bond donors, hydrogen bond acceptors, aromatic systems, electron-withdrawing groups, hydrophobic groups, metal coordinating groups, and halogen substituents. Some examples are shown in Table 1. Next, the core atom of each functional group is determined. The atom with the strongest chemical reactivity or best representing the characteristics of the functional group can be selected as the core atom. When a functional group contains multiple equivalent atoms, the first matching atom can be selected as the core atom. For aromatic ring systems, the carbon atom with the lowest ring index can be selected as the core atom. Through the above processing, the functional group set of the target drug molecule is obtained. F ={ f 1, f 2,..., f n}, each functional group f i It includes type labels, core atom indexes, and chemical category information.

[0027] Table 1. Definition of SMARTS patterns of functional groups

[0028] Step S103: Based on the preset strong rule interaction knowledge base and weak rule scoring function, construct virtual interaction edges between each functional group pair in the functional group set. The virtual interaction edges are used to represent the non-covalent interaction between physically unconnected functional groups.

[0029] Step S104: Add virtual interactive edges to the original molecular graph to generate an enhanced molecular graph.

[0030] In practical applications, based on the functional group identification results, virtual interaction edges need to be constructed between each functional group pair in the functional group set. These virtual interaction edges are then added to the original molecular graph to generate an enhanced molecular graph. This transforms the domain knowledge of non-covalent cooperation and antagonism between functional groups, which cannot be expressed through physical covalent bonds, into virtual interaction edges in the graph topology. Furthermore, a five-dimensional physicochemical scoring mechanism with a weak rule layer is introduced to cover more potential functional group combinations. Virtual interaction edges are a special type of graph edge used to characterize the interaction relationship between two physically unconnected but chemically non-covalently interacting functional groups. Virtual interaction edges are not actual chemical bonds in the molecule, but rather explicit encodings of interaction relationships inferred from physical organic chemistry and medicinal chemistry knowledge within the molecular graph topology.

[0031] In an exemplary embodiment, during the process of constructing virtual interaction edges between functional group pairs in the functional group set based on a preset strong-rule interaction knowledge base and a weak-rule scoring function, each functional group pair in the functional group set can be traversed to determine whether the category combination of the current functional group pair matches the non-covalent interaction in the strong-rule interaction knowledge base. If it matches, a typed virtual interaction edge is created between the current functional group pair in the functional group set, and the corresponding interaction type code can be labeled. The typed virtual interaction edge has a clear chemical semantics, and its interaction type is known and determined, with a high initial confidence level. If it does not match, an interaction potential score is generated based on the physicochemical properties of the current functional group pair, and it is determined whether the interaction potential score is greater than or equal to a preset weak-rule threshold. , The value can be flexibly set as needed; if the interaction potential score is greater than or equal to the weak rule threshold, an implicit virtual interaction edge is created between the current functional group pairs in the functional group set; among them, non-covalent interactions include: electron donation and acceptance coordination, hydrogen bond network enhancement, hydrophobic cluster coordination, metal chelation coordination, steric hindrance antagonism and acid-base neutralization.

[0032] It should be noted that the strong-rule interaction knowledge base is a structured set of rules for interactions between functional groups, which can be stored in the form of a lookup table or a rule engine. Each rule defines a combination of functional group categories and the corresponding type of non-covalent interaction. In this application, based on the principles of physical organic chemistry and medicinal chemistry, six types of non-covalent interactions between functional groups are predefined in the strong-rule interaction knowledge base, as shown in Table 2. Specifically, to avoid... Overgeneralization of rules based on Taft steric hindrance parameters and topological distance pairs Three constraints are imposed on the edge: the first is that the E of at least one functional group sThe conditions for triggering the event are: 1) the steric hindrance of the functional group is significantly greater than that of the methyl group; 2) the shortest path distance between the two functional groups is ≤4 hops to ensure that a physical collision is possible; and 3) at least one functional group contains a hydrophobic group. All three conditions must be met simultaneously to trigger the event. .

[0033] Table 2. Functional Group Interaction Knowledge Base

[0034] It should also be noted that if a functional group pair does not match any type of non-covalent interaction in the strong-rule interaction knowledge base, it does not mean that there is no meaningful non-covalent interaction between the functional group pair; it may simply mean that this type of combination has not yet been included in the existing knowledge base. Therefore, a weak-rule layer is introduced to calculate continuous interaction potential scores based on the physicochemical properties of the functional group pair, in order to cover the long-tail functional group combination space that the knowledge base cannot cover. The formula for generating the interaction potential score is: ;in w k The weights for each feature component are given, and the sum of the five weights is 1.0. The variables in the weak rule scoring formula are shown in Table 3; only when... When creating Implicit virtual interaction edges, with a confidence level of In this way, the knowledge of non-covalent interactions between functional groups in medicinal chemistry can be explicitly encoded as virtual interaction edges in a graph topology, enabling graph neural networks to directly perceive long-range chemical effects that originally required multiple layers of message passing to capture, fundamentally expanding the receptive field of GNNs for long-range interactions. Furthermore, the strong rule layer ensures accurate identification and high-confidence encoding of known interaction types, while the weak rule layer covers potential interaction combinations not included in the knowledge base through continuous scoring. The combination of the two achieves full coverage of non-covalent interaction signals in molecules.

[0035] Table 3. Explanation of Variables in Weak Rule Scoring Formulas

[0036] In an exemplary embodiment, the edge features of the virtual interaction edges can be designed as 18-dimensional vectors, including 7-dimensional one-hot encoding of interaction type, 8-dimensional one-hot encoding of topological distance, 1-dimensional estimated spatial distance, 1-dimensional electronegativity difference, and 1-dimensional polarizability difference. Thus, through the above processing, the original molecular graph... G o ( V , E phys ) was enhanced into an enhanced molecular map G aug ( V , E phys ∪ Evirt ),in E virt For all newly created sets of virtual interaction edges, each virtual interaction edge is bound to an 18-dimensional feature vector and an initial confidence score.

[0037] Step S105: Apply a dual-channel encoder containing physical bond channels and virtual interaction channels to encode the enhanced molecular map and obtain a graphical representation.

[0038] In practical applications, physical covalent bonds (short-range, localized, electron-sharing) and virtual interaction edges (long-range, non-localized, non-covalently cooperative) have distinctly different chemical and physical properties. Mixing them into a single message-passing channel would lead to feature confusion and degree distribution distortion. Therefore, a dual-channel encoder containing both physical bond and virtual interaction channels is used to encode the enhanced molecular graph, resulting in a graph representation.

[0039] In an exemplary embodiment, a virtual interaction channel is used for message passing between virtual interaction edges in the enhanced molecular graph. A dual-channel encoder, including a physical bond channel and a virtual interaction channel, is applied to encode the enhanced molecular graph. During this process, an initial confidence level and a layer attenuation coefficient can be configured for each virtual interaction edge, with the layer attenuation coefficient differentiated according to the interaction type. During the execution of the virtual interaction channel... l During layer message passing, the current confidence of the virtual interaction edge is updated based on the product of the initial confidence, the layer decay coefficient, and the task adaptive attention weight; it is then determined whether the current confidence is lower than a preset confidence pruning threshold. If so, then at the 1st... lVirtual interaction edges are skipped during message passing within the layer. The physical bond channel is used for message passing between covalent bonds in the enhanced molecular graph. When using a dual-channel encoder containing both physical and virtual interaction channels to encode the enhanced molecular graph, the physical bond channel output, virtual interaction channel output, and the previous layer node representation of the current node can be obtained. Based on these outputs, a gating value is generated for the current node. The physical bond channel output and virtual interaction channel output are then weighted and summed according to the gating value to obtain a weighted sum. Residual connections and layer normalization are performed on the weighted sum to obtain the fused node representation of the current layer. In the application of the dual-channel encoder, mutual information regularization constraints are used to minimize the upper bound of the mutual information between the physical bond channel output and the virtual interaction channel output. Specifically, CLUB (Contrastive Log-ratio Upper Bound) is used. The mutual information upper bound estimation method (Bound) forces the two channels to learn complementary rather than redundant molecular feature representations by minimizing the mutual information between their outputs. The physical bond channel focuses on local electronic effects related to covalent bonds, such as bond type, hybridization, and aromaticity, while the virtual interaction channel focuses on long-range cooperative effects of non-covalent bonds, such as electron transfer between functional groups, hydrogen bond networks, and hydrophobic collapse. Minimizing the mutual information between the two channels ensures information decoupling between them.

[0040] To better understand the process of the dual-channel encoder encoding the enhanced molecular map, please refer to [link to relevant documentation]. Figure 2 , Figure 2 This is a diagram of the architecture of the dual-channel GNN encoder in this application. The dual-channel encoder contains two channels with independent parameters. The physical bond channel is used to handle covalent bond message passing in the enhanced molecular graph. This channel can use a Graph Isomorphism Network (GIN) as its backbone architecture, with a total of 3 layers and a hidden dimension of 256. l Layer nodes v Physical channel representation The updated formula is ;in, For the first l Multilayer perceptron, These are the learnable parameters of the physical channel. For nodes v The set of physical covalent bond neighbors, For nodes With nodes v Features of covalent bond edges between them. Indicates the first l Nodes at level -1 v The physical channel representation, Indicates the first 1Nodes at level -1 u The physical channel representation, The network is used for edge feature encoding; the virtual interaction channel is used to process remote collaborative messages, and includes adaptive confidence decay. It can also adopt the GIN architecture, but its parameters are completely independent of the physical key channel. The virtual interaction channel introduces an adaptive confidence decay mechanism during message transmission, and each virtual edge is bound to a learnable composite confidence weight. for ,in, The virtual edge base confidence is learned by the confidence scoring network. The layer attenuation coefficient is determined by edge type; different edge types use different coefficients. Value, for example 1. 2. 3. 4. 6 uses =0.85 standard attenuation, 5 uses =0.75 accelerates decay, 0 use =0.70 accelerates decay, For task-adaptive attention weights, For message passing layer index; when Below the threshold In this case, message passing for the virtual edge at that layer is automatically skipped (soft pruning), which reduces the computational overhead of the virtual channel while maintaining prediction accuracy. A diagram illustrating confidence decay and soft pruning is shown below. Figure 3 As shown in the figure, curve 1 is R 1- R 4 and R The curve corresponding to 6, curve 2 is R The curve corresponding to 5, curve 3 is R The curve corresponding to 0.

[0041] Understandably, after completing independent message passing between the physical key channel and the virtual interaction channel, the outputs of the two channels are merged through a gating mechanism. Specifically, this involves obtaining the current node... v Physical bond channel output Virtual interactive channel output and the representation of the node at the next higher level Generate the current node based on the three factors. v gating value , , ,in, σ It is the Sigmoid activation function. For the learnable weight matrix of the gated network, For the bias term, [ The symbol ] represents a vector concatenation operation; residual join and layer normalization are performed on the weighted summation result to obtain the fused node representation of the current layer. , After message passing is completed across all layers, the node-level representations are aggregated into a graph-level representation vector through the graph readout layer. The graph readout layer can employ three pooling strategies for concatenation, resulting in a graph-level representation. hgnn The generating formula is ,in, L For network layers, For the first L Layer nodes v The fusion representation, For the learnable weight matrix of the readout layer, This represents mean pooling, used to capture global average features. This represents max pooling, used to capture signals from prominent functional groups. This indicates attention pooling based on LSTM. In this way, separate processing through two channels ensures that message passing between physical covalent bonds and virtual interaction edges does not interfere with each other, avoiding feature confusion; adaptive fusion of the two channels can be achieved through gating mechanisms, allowing the model to dynamically adjust its dependence on the two types of information according to molecular structure and task requirements; mutual information regularization forces the learning of complementary information between channels, avoiding information redundancy; and confidence decay and soft pruning mechanisms improve inference efficiency while maintaining performance.

[0042] Step S106: Encode the SMILES sequence to obtain a sequence representation.

[0043] In practical applications, the SMILES sequence also needs to be encoded to obtain a sequence representation. For example, the SMILES string of the target drug molecule can be input into the Transformer encoder for encoding to obtain a sequence representation. This allows the global self-attention mechanism of the Transformer to capture long-distance dependencies in the SMILES sequence, providing global sequence semantic supplementation to the topological local information of the GNN graph and enriching the information dimension of molecular representation.

[0044] In an exemplary embodiment, during the processing of the SMILES sequence using the Transformer encoder, the SMILES string can be segmented at the character / atom level. The vocabulary contains approximately 50 tokens, such as C, N, O, S, F, Cl, Br, =, #, (,), [, @, 1, 2, etc., and two character atoms are automatically merged into a single token, such as Cl, Br, Si, Se. The maximum sequence length is 256. A standard 4-layer Transformer encoder is used, with a model dimension of 256, 8-head attention, and a feedforward network dimension of 512. A sinusoidal positional encoding scheme is used to add positional information to each position in the sequence. A special [CLS] token is added to the beginning of the input sequence. After Transformer encoding, the output vector corresponding to the [CLS] token is used as the sequence-level representation. .

[0045] Step S107: Perform cross-attention fusion on the graph representation and the sequence representation to obtain the fused representation vector.

[0046] In practical applications, the SMILES sequence encoder and the GNN graph encoder form a complementary information acquisition path. GNN excels at capturing local topology, such as chemical bond connection patterns, the direct neighborhood of functional group core atoms, and specific cooperative relationships encoded by virtual interaction edges. Transformer, on the other hand, excels at capturing global sequence dependencies, such as the positional relationships of functional groups in the SMILES sequence and the overall hydrophobic / hydrophilic patterns of substituents. The fusion of the two modal information helps to understand molecular structure more comprehensively. To this end, cross-attention fusion is performed on graph representation and sequence representation to obtain a fused representation vector.

[0047] In an exemplary embodiment, during the process of performing cross-attention fusion processing on graph representation and sequence representation to obtain a fused representation vector, the graph representation and sequence representation can be projected onto a shared fusion space respectively; the sequence representation is queried using the graph representation, and the graph representation is queried using the sequence representation to obtain the cross-enhanced graph representation and the cross-enhanced sequence representation respectively; a fusion gate value is generated based on the cross-enhanced graph representation and the cross-enhanced sequence representation; the cross-enhanced graph representation and the cross-enhanced sequence representation are weighted based on the fusion gate value to obtain a weighted processing result; the weighted processing result is concatenated with the unweighted cross-enhanced graph representation and the cross-enhanced sequence representation to obtain a concatenated result; and the concatenated result is mapped to obtain the fused representation vector.

[0048] For a better understanding of this process, please refer to [link / reference]. Figure 4 , Figure 4 This is a schematic diagram illustrating the process of cross-attention fusion of graph representation and sequence representation provided in an embodiment of this application. Specifically, the GNN graph representation... hgnn ∈1024 and Transformer sequence representation h trans Each of the ∈256 is mapped to a 512-dimensional shared fusion space via a linear projection layer. The query sequence is represented by a graph, and the query graph is represented by a sequence. After residual connections and layer normalization, the cross-enhanced graph representation is obtained. and the sequence representation after cross-enhancement Perform gated adaptive weighted and quaternion concatenation output. ， ,in, W f To integrate the learnable weight matrix of the gating network, b f For the bias term, the fusion gate value ∈[0,1] reflects the degree of dependence of the current prediction task on information from both modalities. As the value approaches 1, the prediction task relies more on the topological information of the GNN graph. When the value is close to 0, it relies more on global information of the Transformer sequence. This represents the fused molecular representation vector. In this way, a deep interaction and fusion of graph topological information and sequence semantic information can be achieved through a cross-attention mechanism, making the information from the two modalities complementary. Furthermore, a task-adaptive fusion weight allocation can be achieved through a gating mechanism, enabling the model to dynamically adjust its dependence on the two modalities based on the characteristics of different prediction tasks, thus improving the model's flexibility and generalization ability.

[0049] Step S108: Apply the pre-trained parameter set to predict the properties of the fused representation vector to obtain the property prediction results of the target drug molecule.

[0050] In practical applications, the obtained fused representation vector needs to be input into a pre-trained parameter set for processing, and the prediction head outputs the property prediction results of the target drug molecule. Furthermore, molecular descriptors can be concatenated to output the prediction results; for example, eight key molecular descriptor features calculated by RDKit can be additionally concatenated to the fused representation vector to further improve the prediction accuracy for downstream tasks sensitive to global properties.

[0051] In an exemplary embodiment, to ensure the effectiveness of property prediction, the pre-training process can be set according to chemical cognitive logic, such as being divided into three stages of training. The first stage answers whether there is an interaction between the two functional groups. This is the most basic judgment, which only requires binary classification and converges the fastest. The second stage, after confirming the existence of an interaction, answers what type of interaction it is, and then introduces contrastive loss to require the gating mechanism to use virtual channels to distinguish the semantic differences of different interaction types. The third stage answers the most complex question: which core atom should be the starting point and the ending point of this interaction? The introduction of error routing forces the model to understand the spatial directionality of the interaction relationship, as shown in Table 4. Based on this, assuming that pre-training adopts a self-supervised learning method, before applying the pre-trained parameter set to predict the properties of the fused representation vector, a training sample set can be obtained; random deletion perturbation is performed on the virtual interaction edges of the training sample set, and virtual edge existence binary classification is used as the training task to train the molecular property prediction model; random deletion perturbation and interaction type tampering perturbation are performed on the virtual interaction edges of the training sample set, and interaction type multi-classification is used as the training task to train the molecular property prediction model; random deletion perturbation, interaction type tampering perturbation, and error routing perturbation are performed on the virtual interaction edges of the training sample set, and functional group routing consistency contrastive learning is used as the training task to train the molecular property prediction model. Here, error routing perturbation refers to randomly selecting an existing virtual interaction edge, randomly replacing the endpoint functional group of the virtual interaction edge with another functional group in the target drug molecule that has no virtual interaction relationship with the starting functional group of the virtual interaction edge, and assigning an error type label different from the original interaction type to the replaced error interaction edge; the parameter set of the trained molecular property prediction model is obtained; where the pre-training loss function can be... ,in, It is the cross-entropy of binary classification based on edge existence and the cross-entropy of multi-class classification based on interaction type. It is the InfoNCE contrast loss. It is a CLUB mutual information decoupling term.

[0052] Table 4 Overview of the Three-Stage Progressive Pre-training Design

[0053] It is understandable that after obtaining the property prediction results of the target drug molecule, this application can flexibly apply these results as needed. For example, based on the property prediction results, candidate molecules with target activity and good pharmacokinetic properties can be quickly screened from a large-scale compound library, reducing the number of compounds screened and lowering R&D costs. For example, during the iterative optimization of medicinal chemistry, the property prediction results can be used to quickly predict the impact of structural modifications on molecular properties, guiding the direction of chemical synthesis. For example, based on the property prediction results, the absorption, distribution, metabolism, excretion, and toxicity properties of candidate molecules can be evaluated in the early stages of drug development, eliminating candidate molecules with undesirable properties in advance. For example, based on the property prediction results, new indications can be predicted and screened for existing drug molecules, accelerating drug repositioning research, etc.

[0054] This application provides a method for predicting the properties of drug molecules, which involves: acquiring the original molecular graph and SMILES sequence of the target drug molecule; identifying the set of functional groups in the target drug molecule; constructing virtual interaction edges between pairs of functional groups in the functional group set based on a pre-set strong-rule interaction knowledge base and weak-rule scoring function, whereby the virtual interaction edges are used to characterize the non-covalent interactions between physically unconnected functional groups; adding the virtual interaction edges to the original molecular graph to generate an enhanced molecular graph; applying a dual-channel encoder containing physical bond channels and virtual interaction channels to encode the enhanced molecular graph to obtain a graph representation; encoding the SMILES sequence to obtain a sequence representation; performing cross-attention fusion on the graph representation and the sequence representation to obtain a fused representation vector; and applying a pre-trained parameter set to predict the properties of the fused representation vector to obtain the property prediction results of the target drug molecule. In this application, by constructing virtual interaction edges to characterize non-covalent interactions between physically disconnected functional groups and adding them to the original molecular graph to generate an enhanced molecular graph, the graph neural network can directly perceive the effects between functional groups during message passing. This overcomes the inherent limitation of traditional GNNs, which cannot effectively stack deep networks due to oversmoothing, and expands the model's receptive field to long-range chemical effects. Furthermore, by using physical bond channels and virtual interaction channels with independent parameters to process covalent and non-covalent interaction information respectively, feature confusion caused by mixing and encoding two types of chemical information with different physical natures is avoided. By fusing graph representation and sequence representation through cross-attention, deep interaction and complementary enhancement of local topological information and global sequence semantic information are achieved. Combined with the general molecular representation prior knowledge provided by the pre-trained parameter set, better generalization performance can be obtained in tasks with limited data, improving the prediction accuracy and generalization of drug molecule properties.

[0055] To facilitate understanding of the three-stage pre-training process provided in this application, an experimental scheme was used to validate it. Pre-training utilized the ZINC250k dataset, selecting a subset of 65,000 molecules with maximized chemical diversity using a stratified sampling strategy. Self-supervised learning was employed, and the subset was randomly divided into a pre-training set and a validation set at a 9:1 ratio. Downstream fine-tuning used five datasets from the MoleculeNet benchmark: the BACE dataset, which includes 1513 drug molecules with significant structural diversity and their corresponding inhibitory activities, aiming to build a computational model predicting the binding ability of small molecule compounds to β-secretase; the BBBP dataset, which assesses blood-brain barrier permeability through a binary classification task and contains 2039 experimentally validated small molecule compounds, contributing to understanding molecule-brain interactions; and the ClinTox dataset, which integrates the toxicity characteristics of 1478 FDA-approved drugs and failed clinical trial compounds, comparing the structure-activity relationship between therapeutic doses and toxic responses. The dataset provides insights into drug safety assessment; the FreeSolv dataset records experimental solvation free energy data for 642 molecular complexes, covering a wide chemical space from simple ions to complex organic molecules; the ESOL dataset includes measured solubility values ​​of 1128 small organic molecules in water and their corresponding molecular descriptors, playing a role in drug formulation research by associating key physicochemical parameters such as molecular weight, polar surface area, and number of hydrogen bond donors and acceptors with solubility performance; all datasets are divided into training, validation, and test sets in an 8:1:1 ratio using the Bemis-Murcko scaffold splitting method, with a fixed random seed of 42.

[0056] Pre-training uses SMILES strings of 65,000 molecules as input, batch size=256, employs the AdamW optimizer, learning rate 1e-4, and weight decay. Cosine annealing learning rate scheduling was used. Pre-training consisted of 80 epochs, divided into three phases: (S1) epochs 1-24 trained only edge existence binary classification, with a contrastive loss weight β=0; (S2) epochs 25-56 introduced type tampering perturbations and interaction type six-class classification, with β=0.3; (S3) epochs 57-80 introduced erroneous routing perturbations and functional group routing consistency contrastive learning, with β=0.5. The mutual information regularization weight α was 0.01, 0.02, and 0.02 in S1, S2, and S3, respectively. A checkpoint was saved every epoch to support breakpoint continuation training. The pre-trained graph structure was completely consistent with the downstream augmented graph, both being augmented molecular graphs constructed using hierarchical virtual edges. The loss function value curve for the three-stage progressive course pre-training is shown below. Figure 5 As shown.

[0057] Each downstream dataset was independently fine-tuned, and Table 5 reports the overall performance of the proposed method compared to six existing methods. Experimental results show that the proposed method achieves optimal values ​​on the evaluation metrics for three classification tasks and two regression tasks compared to other state-of-the-art baseline models, demonstrating the effectiveness of the proposed method in molecular property prediction tasks.

[0058] Table 5. Comparison results on 3 classification tasks and 2 regression tasks.

[0059] Furthermore, it has been verified that, using the processing flow designed in this application and based on the functional group interaction knowledge base and pre-trained model, the property prediction and inference of single-molecule compounds can be completed in milliseconds. The confidence soft pruning mechanism automatically skips approximately 30-47% of low-reliability virtual edges in deep message passing, reducing the computational overhead of virtual channels while maintaining prediction accuracy. The virtual edge confidence distribution map can show the model's confidence level for each interaction rule, and the task fusion gating value can show the differences in the dependence of different downstream tasks on GNN graph topology and Transformer sequence information. These provide actionable analytical basis for medicinal chemists' molecular optimization decisions and have strong interpretability. Pre-training can be completed in approximately 52 hours with a single NVIDIA RTX 4060 card running 80 epochs, and fine-tuning and evaluation of all 5 downstream datasets can be completed within 3 hours, demonstrating low cost and high universality.

[0060] Please see Figure 6 , Figure 6 This is a schematic diagram of the structure of a drug molecule property prediction system provided in an embodiment of this application.

[0061] This application provides a drug molecule property prediction system, which may include: The data acquisition module 101 is used to acquire the original molecular map and SMILES sequence of the target drug molecule; Functional group recognition module 102 is used to identify the set of functional groups in the target drug molecule; Module 103 is created to construct virtual interaction edges between pairs of functional groups in the functional group set based on a preset strong rule interaction knowledge base and weak rule scoring function. The virtual interaction edges are used to characterize the non-covalent interactions between physically unconnected functional groups. Add module 104 to add virtual interactive edges to the original molecular graph to generate an enhanced molecular graph; The graph encoding module 105 is used to encode the enhanced molecular graph using a dual-channel encoder that includes physical bond channels and virtual interaction channels to obtain a graph representation. The sequence encoding module 106 is used to encode the SMILES sequence to obtain a sequence representation; The fusion module 107 is used to perform cross-attention fusion on graph representation and sequence representation to obtain a fused representation vector. The prediction module 108 is used to apply the pre-trained parameter set to predict the properties of the fused representation vector and obtain the property prediction results of the target drug molecule.

[0062] This application provides a drug molecule property prediction system, the creation module of which includes: The traversal unit is used to traverse each pair of functional groups in the set of functional groups; A creation unit is used to determine whether the category combination of the current functional group pair matches the non-covalent interaction in the strong rule interaction knowledge base. If it matches, a typed virtual interaction edge is created between the current functional group pairs in the functional group set. If it does not match, an interaction potential score is generated based on the physicochemical properties of the current functional group pair, and it is determined whether the interaction potential score is greater than or equal to a preset weak rule threshold. If the interaction potential score is greater than or equal to the weak rule threshold, an implicit virtual interaction edge is created between the current functional group pairs in the functional group set. Non-covalent interactions include: electron donation and acceptance coordination, hydrogen bond network enhancement, hydrophobic cluster coordination, metal chelation coordination, steric hindrance antagonism, and acid-base neutralization.

[0063] This application provides a drug molecule property prediction system, in which the creation unit is specifically used to: perform a weighted summation of the electronegativity difference, polarizability difference, hydrogen bond donor-acceptor complementarity, spatial matching degree, and distance decay factor of the current functional group pair to obtain an interaction potential score.

[0064] This application provides a drug molecule property prediction system. A virtual interaction channel is used for message passing between virtual interaction edges in an enhanced molecular graph. Specifically, the graph encoding module is used to: configure an initial confidence level and a layer decay coefficient for each virtual interaction edge, with the layer decay coefficient differentiated according to the interaction type; and during the execution of the virtual interaction channel... l During layer message passing, the current confidence of the virtual interaction edge is updated based on the product of the initial confidence, the layer decay coefficient, and the task adaptive attention weight; it is then determined whether the current confidence is lower than a preset confidence pruning threshold. If so, then at the 1st... l Skip virtual interaction edges in message passing within the layer.

[0065] This application provides a drug molecule property prediction system. The physical bond channel is used for message passing between covalent bonds in an enhanced molecular graph. Specifically, the graph encoding module is used to: acquire the physical bond channel output, virtual interaction channel output, and the previous layer node representation of the current node; generate a gating value for the current node based on the physical bond channel output, virtual interaction channel output, and the previous layer node representation; perform a weighted summation of the physical bond channel output and the virtual interaction channel output based on the gating value to obtain a weighted summation result; and perform residual connection and layer normalization processing on the weighted summation result to obtain the fused node representation of the current layer. During the training process of the dual-channel encoder, mutual information regularization constraints are used to minimize the upper bound of the mutual information between the physical bond channel output and the virtual interaction channel output.

[0066] This application provides a drug molecule property prediction system, the fusion module of which may include: The projection unit is used to project the graph representation and the sequence representation onto the shared fusion space, respectively. The query unit is used to query the sequence representation using the graph representation and query the graph representation using the sequence representation, thereby obtaining the graph representation after cross-enhancement and the sequence representation after cross-enhancement, respectively. The gating unit is used to generate a fusion gating value based on the graph representation after cross-enhancement and the sequence representation after cross-enhancement; The processing unit is used to perform weighted processing on the graph representation after cross-enhancement and the sequence representation after cross-enhancement according to the fusion gate value, so as to obtain the weighted processing result; The splicing unit is used to splice the weighted processing result with the unweighted cross-enhanced graph representation and the cross-enhanced sequence representation to obtain the spliced ​​result; The mapping unit is used to map the splicing result to obtain the fused representation vector.

[0067] The drug molecule property prediction system provided in this application embodiment may further include: The training set acquisition module is used to acquire a training sample set before the prediction module applies the pre-trained parameter set to predict the properties of the fused representation vector. The first training module is used to perform random deletion perturbation on the virtual interaction edges of the training sample set, and to train the molecular property prediction model by using the binary classification of virtual edge existence as the training task. The second training module is used to perform random deletion perturbation and interaction type tampering perturbation on the virtual interaction edges of the training sample set, and to train the molecular property prediction model with multi-class classification of interaction types as the training task. The third training module is used to perform random deletion perturbation, interaction type tampering perturbation, and error routing perturbation on the virtual interaction edges of the training sample set, and to train the molecular property prediction model with functional group routing consistency contrastive learning as the training task. The parameter set acquisition module is used to acquire the parameter set of the trained molecular property prediction model.

[0068] This application also provides an electronic device and a computer-readable storage medium, both of which have the corresponding effects of the drug molecule property prediction method provided in the embodiments of this application. Please refer to... Figure 7 , Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0069] An electronic device provided in this application includes a memory 201 and a processor 202. The memory 201 stores a computer program, and the processor 202 executes the computer program to implement the steps of the drug molecule property prediction method described in any of the above embodiments.

[0070] Please see Figure 8 Another electronic device provided in this application embodiment may further include: an input port 203 connected to the processor 202 for transmitting commands input from the outside to the processor 202; a display unit 204 connected to the processor 202 for displaying the processing results of the processor 202 to the outside; and a communication module 205 connected to the processor 202 for enabling communication between the electronic device and the outside. The display unit 204 may be a display panel, a laser scanning display, etc.; the communication method adopted by the communication module 205 includes, but is not limited to, Mobile High-Definition Link (MHL), Universal Serial Bus (USB), High-Definition Multimedia Interface (HDMI), wireless connection: Wireless Fidelity (WiFi), Bluetooth communication technology, Bluetooth Low Energy communication technology, and communication technology based on IEEE 802.11s.

[0071] This application provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the steps of the drug molecule property prediction method described in any of the above embodiments.

[0072] The computer-readable storage media involved in this application include random access memory (RAM), memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disks, removable disks, CD-ROMs (compact disc read-only memory), or any other form of storage media known in the art.

[0073] This application provides a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the drug molecule property prediction method described in any of the above embodiments.

[0074] For descriptions of relevant parts of the drug molecule property prediction system, electronic device, and computer-readable storage medium provided in the embodiments of this application, please refer to the detailed descriptions of the corresponding parts in the drug molecule property prediction method provided in the embodiments of this application, and they will not be repeated here. Furthermore, parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of corresponding technical solutions in the prior art have not been described in detail to avoid excessive elaboration.

[0075] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0076] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for predicting the molecular properties of a drug, characterized in that, include: Obtain the original molecular map and SMILES sequence of the target drug molecule; Identify the set of functional groups in the target drug molecule; Based on a preset strong rule interaction knowledge base and weak rule scoring function, virtual interaction edges are constructed between each functional group pair in the functional group set. The virtual interaction edges are used to characterize the non-covalent interaction between physically unconnected functional groups. The virtual interactive edges are added to the original molecular graph to generate an enhanced molecular graph; A dual-channel encoder, including physical bond channels and virtual interaction channels, is used to encode the enhanced molecular map to obtain a graphical representation; The SMILES sequence is encoded to obtain a sequence representation; Cross-attention fusion is performed on the graph representation and the sequence representation to obtain a fused representation vector; The properties of the fused representation vector are predicted by applying the pre-trained parameter set, thereby obtaining the property prediction results of the target drug molecule.

2. The method according to claim 1, characterized in that, The step of constructing virtual interaction edges between pairs of functional groups in the set of functional groups based on a preset strong-rule interaction knowledge base and a weak-rule scoring function includes: Iterate through all functional group pairs in the functional group set and determine whether the category combination of the current functional group pair matches the non-covalent interaction in the strong rule interaction knowledge base. If a hit occurs, a typed virtual interaction edge is created between the current functional group pairs in the set of functional groups. If no match is found, an interaction potential score is generated based on the physicochemical properties of the current functional group pair, and it is determined whether the interaction potential score is greater than or equal to a preset weak rule threshold. If the interaction potential score is greater than or equal to the weak rule threshold, then an implicit virtual interaction edge is created between the current functional group pairs in the functional group set. The non-covalent interactions include: electron donation and acceptance synergy, hydrogen bond network enhancement, hydrophobic cluster synergy, metal chelation synergy, steric hindrance antagonism, and acid-base neutralization.

3. The method according to claim 2, characterized in that, The generation of interaction potential scores based on the physicochemical properties of the current functional group pairs includes: The interaction potential score is obtained by weighting and summing the electronegativity difference, polarizability difference, hydrogen bond donor-acceptor complementarity, spatial matching degree, and distance decay factor of the current functional group pair.

4. The method according to claim 1, characterized in that, The virtual interaction channel is used for message passing between the virtual interaction edges in the enhanced molecular graph. The application includes a dual-channel encoder with a physical bond channel and a virtual interaction channel. The encoding process of the enhanced molecular graph includes: Configure an initial confidence level and a layer decay coefficient for each virtual interaction edge, wherein the layer decay coefficient is distinguished according to the interaction type; In the execution of the virtual interaction channel l During layer message passing, the current confidence of the virtual interaction edge is updated based on the product of the initial confidence, the layer decay coefficient, and the task adaptive attention weight. Determine if the current confidence level is lower than the preset confidence pruning threshold. If so, then at the... l Skip virtual interaction edges in message passing within the layer.

5. The method according to claim 4, characterized in that, The physical bond channel is used for message passing between covalent bonds in the enhanced molecular graph. The application includes a dual-channel encoder with both a physical bond channel and a virtual interaction channel. The encoding process of the enhanced molecular graph includes: Obtain the physical key channel output, virtual interaction channel output, and the representation of the previous node of the current node; Generate the gating value for the current node based on the physical key channel output, virtual interaction channel output, and the representation of the previous node. The physical key channel output and the virtual interaction channel output are weighted and summed according to the gating value to obtain the weighted summation result; Perform residual connection and layer normalization processing on the weighted summation result to obtain the fusion node representation of the current layer; In the application of the dual-channel encoder, the upper bound of the mutual information between the output of the physical key channel and the output of the virtual interaction channel is minimized by mutual information regularization constraints.

6. The method according to claim 1, characterized in that, The process of performing cross-attention fusion on the graph representation and the sequence representation to obtain a fused representation vector includes: The graph representation and the sequence representation are respectively projected onto a shared fusion space; The graph representation is queried in the sequence representation, and the sequence representation is queried in the graph representation to obtain the graph representation after cross-enhancement and the sequence representation after cross-enhancement, respectively. Based on the cross-enhanced graph representation and the cross-enhanced sequence representation, a fusion gate value is generated; Based on the fusion gate value, the cross-enhanced graph representation and the cross-enhanced sequence representation are weighted to obtain the weighted processing result; The weighted result is concatenated with the unweighted graph representation after cross-enhancement and the sequence representation after cross-enhancement to obtain the concatenated result; The splicing result is mapped to obtain the fused representation vector.

7. The method according to claim 1, characterized in that, Before applying the pre-trained parameter set to predict the properties of the fused representation vector, the method further includes: Obtain the training sample set; Random deletion perturbation is performed on the virtual interaction edges of the training sample set, and the molecular property prediction model is trained using the binary classification of virtual edge existence as the training task. Random deletion perturbation and interaction type tampering perturbation are performed on the virtual interaction edges of the training sample set, and the molecular property prediction model is trained with multi-class classification of interaction types as the training task. Random deletion perturbation, interaction type tampering perturbation, and erroneous routing perturbation are performed on the virtual interaction edges of the training sample set, and functional group routing consistency contrastive learning is used as the training task to train the molecular property prediction model. Obtain the parameter set of the trained molecular property prediction model.

8. A drug molecule property prediction system, characterized in that, include: The data acquisition module is used to acquire the original molecular map and SMILES sequence of the target drug molecule; A functional group recognition module is used to identify the set of functional groups in the target drug molecule; A creation module is used to construct virtual interaction edges between pairs of functional groups in the functional group set based on a preset strong rule interaction knowledge base and weak rule scoring function. The virtual interaction edges are used to characterize the non-covalent interactions between physically unconnected functional groups. An addition module is used to add the virtual interactive edges to the original molecular graph to generate an enhanced molecular graph; The graph encoding module is used to encode the enhanced molecular graph using a dual-channel encoder that includes physical bond channels and virtual interaction channels to obtain a graph representation; A sequence encoding module is used to encode the SMILES sequence to obtain a sequence representation; The fusion module is used to perform cross-attention fusion on the graph representation and the sequence representation to obtain a fused representation vector; The prediction module is used to apply a pre-trained parameter set to predict the properties of the fused representation vector, thereby obtaining the property prediction results of the target drug molecule.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the steps of the method for predicting the properties of drug molecules as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the drug molecule property prediction method as described in any one of claims 1 to 7.

Images