Molecular property prediction method and device based on molecular fragmentation and multi-modal fusion

Abstract

The application provides a molecular property prediction method and device based on molecular fragmentation and multi-modal fusion, relates to the technical field of chemical informatics and artificial intelligence, and comprises the following steps: performing molecular fragmentation processing on a target molecule to obtain molecular fragments; extracting global molecular structure features of the target molecule and multi-modal features of each molecular fragment; constructing a graph structure representation of the target molecule; and inputting the graph structure representation into a molecular-level graph neural network to obtain a property prediction result of the target molecule output by the molecular-level graph neural network. The method and device provided by the application decompose a complex molecular property prediction problem to a fragment level with more definite chemical meanings, improve the interpretability of molecular property prediction, fuse multi-modal features from multiple information sources such as structures, texts and atomic levels, improve the semantic integrity and structural rationality of molecular representation, and significantly improve the accuracy of molecular property prediction.

Description

Technical Field

[0001] This application relates to the interdisciplinary fields of cheminformatics and artificial intelligence, and in particular to a method and apparatus for predicting molecular properties based on molecular fragmentation and multimodal fusion. Background Technology

[0002] Accurate and efficient prediction of molecular properties is a core task in many applications such as drug discovery, materials science, and chemical reaction design.

[0003] The molecular modeling methods provided by related technologies mostly rely on artificially designed features (such as molecular fingerprints and atomic pair statistics). Although these methods have achieved some success in specific tasks, they have poor versatility, limited feature representation capabilities, and are difficult to capture complete information about complex molecular systems or adapt to the task requirements in novel chemical spaces. Furthermore, their construction process requires a great deal of expert knowledge.

[0004] Therefore, improving the accuracy and interpretability of molecular property predictions has become a pressing technical problem for the industry. Summary of the Invention

[0005] This application provides a method and apparatus for predicting molecular properties based on molecular fragmentation and multimodal fusion, which addresses the technical problem of how to improve the accuracy and interpretability of molecular property prediction.

[0006] This application provides a method for predicting molecular properties based on molecular fragmentation and multimodal fusion, including: The target molecule is subjected to molecular fragmentation to obtain at least one molecular fragment; Extract the global molecular structure features of the target molecule and the multimodal features of each molecular fragment; the multimodal features include at least two of the following: molecular fragment structure features, text description features, and atomic-level structure features; A graph structure representation of the target molecule is constructed using each molecular fragment as a node; the node features of each node in the graph structure representation are obtained by fusing the multimodal features of each node and the global molecular structure features. The graph structure representation is input into a molecular-level graph neural network to obtain the property prediction results of the target molecule output by the molecular-level graph neural network.

[0007] In some embodiments, the multimodal features are extracted based on the following steps: Molecular fragments are input into a structural semantic coding model to obtain the molecular fragment structural features of the molecular fragments output by the structural semantic coding model. Molecular fragments are input into a preset large language model to obtain the descriptive text of the molecular fragments output by the preset large language model. The descriptive text is then encoded into the textual descriptive features of the molecular fragments based on a pre-trained text embedding model in the field of chemistry. An atomic-level graph is constructed based on the atoms and chemical bonds of the molecular fragments. The atomic-level graph is then input into an atomic-level graph neural network to obtain the atomic-level structural features of the molecular fragments output by the atomic-level graph neural network.

[0008] In some embodiments, the node features of each node in the graph structure representation are obtained based on the following steps: Attention alignment and fusion are performed on at least two of the molecular fragment structure features, text description features and atomic-level structure features of each molecular fragment to obtain the fused structure features of each molecular fragment. The fusion structural features of each molecular fragment are fused with the global molecular structural features to obtain the node features of each molecular fragment in the graph structural representation.

[0009] In some embodiments, the attention alignment and fusion of at least two of the molecular fragment structural features, textual description features, and atomic-level structural features of each molecular fragment to obtain the fused structural features of each molecular fragment includes: Molecular attribute information of each molecular fragment is obtained based on a chemical information database, and the molecular attribute information is encoded into molecular attribute features. Attention alignment and fusion are performed on the molecular fragment structure features, text description features, atomic-level structure features and molecular attribute features of each molecular fragment to obtain the fused structure features of each molecular fragment.

[0010] In some embodiments, the method further includes: Based on the distance between the atomic-level structural features and the molecular fragment structural features, and the distance between the atomic-level structural features and the text description features, an alignment loss function is determined; Based on the alignment loss function and the main task loss function used for molecular property prediction, the total loss function for training the atomic-level graph neural network and the molecular-level graph neural network is determined.

[0011] In some embodiments, the atomic-level graph neural network is a directed message-passing neural network; the molecular-level graph neural network is a graph attention neural network.

[0012] This application provides a molecular property prediction device based on molecular fragmentation and multimodal fusion, comprising: The molecular fragmentation module is used to perform molecular fragmentation processing on the target molecule to obtain at least one molecular fragment; The feature extraction module is used to extract the global molecular structure features of the target molecule and the multimodal features of each molecular fragment; the multimodal features include at least two of the following: molecular fragment structure features, text description features, and atomic-level structure features; The graph structure representation module is used to construct a graph structure representation of the target molecule with each molecular fragment as a node; the node features of each node in the graph structure representation are obtained by fusing the multimodal features of each node and the global molecular structure features; The property prediction module is used to input the graph structure representation into a molecular-level graph neural network to obtain the property prediction results of the target molecule output by the molecular-level graph neural network.

[0013] This application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the aforementioned molecular property prediction method based on molecular fragmentation and multimodal fusion.

[0014] This application provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the aforementioned molecular property prediction method based on molecular fragmentation and multimodal fusion.

[0015] This application provides a computer program product, including a computer program that, when executed by a processor, implements the aforementioned molecular property prediction method based on molecular fragmentation and multimodal fusion.

[0016] The molecular property prediction method and apparatus based on molecular fragmentation and multimodal fusion provided in this application obtain molecular fragments by performing molecular fragmentation on the target molecule, decomposing the complex molecular property prediction problem into a fragment level with clearer chemical meaning, thereby improving the interpretability of molecular property prediction; a graph structure with molecular fragments as nodes is constructed, and multimodal features from multiple information sources such as structure, text, and atomic level are fused into each node, which greatly enriches the input information of the model and improves the semantic integrity and structural rationality of molecular representation; by learning the interaction between fragments through molecular-level graph neural network, the key factors that determine molecular properties can be captured more accurately, thereby significantly improving the accuracy of molecular property prediction. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0018] To more clearly illustrate the technical solutions in 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 some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is one of the flowcharts of the molecular property prediction method based on molecular fragmentation and multimodal fusion provided in this application.

[0020] Figure 2 This is a flowchart of the molecular property prediction method provided in this application, which is based on molecular fragmentation technology and combines multimodal information with graph neural networks.

[0021] Figure 3 This is the second flowchart of the molecular property prediction method based on molecular fragmentation and multimodal fusion provided in this application.

[0022] Figure 4 This is a schematic diagram of the molecular property prediction device based on molecular fragmentation and multimodal fusion provided in this application.

[0023] Figure 5 This is a schematic diagram of the structure of the electronic device provided in this application. Detailed Implementation

[0024] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0025] It should be noted that the terms "first," "second," etc., used in this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that comprises a series of steps, units, or modules is not necessarily limited to those explicitly listed, but may include other steps, units, or modules not explicitly listed or inherent to such processes, methods, products, or devices.

[0026] In recent years, Graph Neural Networks (GNNs) have attracted widespread attention in molecular characterization and property prediction due to their inherent suitability for modeling molecular graph structures. GNNs can perform local dependency modeling on graph structures and learn the interactions between atoms and bonds. However, GNNs themselves rely on structural input and have limited capabilities for semantic understanding and chemical knowledge reasoning. Especially when dealing with complex functional groups, inter-segment relationships, or implicit knowledge that is not explicitly described, GNNs tend to overlook potential semantic connections, affecting prediction performance and model interpretability.

[0027] The development of large language models has brought new opportunities to chemical language representation. By pre-training on large-scale chemical literature or corpora, large language models have the ability to generate molecular descriptive language, extract semantic information from fragments, and answer chemical questions. However, these models usually only process language-level representations and are difficult to directly align with molecular structural representations (as shown in the figure), resulting in a structural-semantic disconnect in downstream tasks.

[0028] Furthermore, most existing molecular representation methods model molecular information based on overall molecular data, neglecting important structural and functional modules at the fragment level within the molecule. However, in actual chemical reactions and property performance, key fragments (such as carboxyl groups, benzene rings, and ester groups) often dominate molecular behavior. Therefore, how to construct more detailed representation frameworks based on molecular fragments and combine them with graph structure modeling and semantic information has become an important research direction for improving the accuracy and interpretability of molecular predictions.

[0029] Currently, there is a lack of a unified modeling framework that can effectively integrate multi-source information such as molecular structure representation, linguistic semantic description, and chemical properties to support high-precision, interpretable molecular property prediction tasks. Furthermore, current research on joint structural and semantic modeling methods for molecular fragments is still immature, lacking a collaborative representation mechanism among the internal structure of fragments (atomic level), connections between fragments (structural diagrams), and semantic descriptions of fragments (linguistic embeddings).

[0030] Therefore, this application provides a novel molecular characterization framework that can unify and integrate structure, semantics, and knowledge based on fragmentation technology, and efficiently integrate multimodal information using graph neural networks and attention mechanisms, thereby improving the expressive power, generalization ability, and interpretability of molecular prediction models and serving complex chemical task scenarios.

[0031] Figure 1 This is one of the flowcharts illustrating the molecular property prediction method based on molecular fragmentation and multimodal fusion provided in this application, such as... Figure 1 As shown, the method includes steps 110, 120, 130 and 140.

[0032] Step 110: Perform molecular fragmentation on the target molecule to obtain at least one molecular fragment.

[0033] Specifically, the execution entity of the molecular property prediction method based on molecular fragmentation and multimodal fusion provided in this application embodiment is a molecular property prediction device or system. This device can be implemented in software, such as a molecular property prediction program based on molecular fragmentation and multimodal fusion; or it can be a device that executes the molecular property prediction method based on molecular fragmentation and multimodal fusion, such as a terminal, computer, or server.

[0034] The target molecule is the molecule whose chemical properties need to be predicted. The target molecule can be characterized using a Simplified Molecular Input Line Entry System (SMILES) string or any other chemical structure representation. A structurally complex target molecule is decomposed into multiple relatively simple, chemically defined sub-structural units, i.e., molecular fragments.

[0035] Molecular fragmentation can be performed using any technique known in the art. For example, fragmentation methods based on predefined chemical reaction rules can be employed, such as Breaking of Retrosynthetically Interesting Chemical Substructures (BRICS) or Retrosynthetic Combinatorial Analysis Procedure (RECAP). These methods break down molecules according to a set of chemically sound fragmentation rules (e.g., cleaving amide bonds, ester bonds, etc.).

[0036] This application employs the MacFrag molecular fragmentation method. MacFrag is an adaptive molecular fragmentation method based on chemical structure semantics and bond energy distribution. Its core idea is to decompose complex molecules into recombinable fragments with clear chemical meaning, while ensuring the rationality of chemical bonds and the integrity of functional groups, thereby enhancing the model's learnability and generalization ability regarding structure-property relationships. Unlike BRICS or RECAP, MacFrag introduces an adaptive decision-making mechanism at the molecular graph level. It comprehensively evaluates interatomic chemical bonds, electronic effects, and local environments (such as heteroatoms, aromaticity, and substituent types), selecting the bond sites with the least change in information entropy to break, thus obtaining a more stable and representative set of fragments.

[0037] This method can be expressed as: In the formula, The SMILES string for the target molecule. It consists of molecular fragments.

[0038] Step 120: Extract the global molecular structure features of the target molecule and the multimodal features of each molecular fragment; the multimodal features include at least two of the following: molecular fragment structure features, text description features, and atomic-level structure features.

[0039] Specifically, molecules and their fragments can be characterized from different granularities and different information dimensions.

[0040] First, extract the global molecular structure features of the target molecule. This feature is a holistic vector representation of the complete target molecule as input. It can be generated by a pre-trained molecular encoder that encodes the SMILES string or graph structure of the molecule into a fixed-dimensional vector. For example, models such as Molecular Representation Learning (Mol2Vec) can be used, which borrows the idea of ​​Word2Vec in natural language processing, analogizing substructures in the molecule to words and the entire molecule to sentences, learning the contextual relationships between substructures to obtain the vector representation of the molecule. This global feature can capture the overall topological and chemical information of the molecule.

[0041] Secondly, multimodal features are extracted from each molecular fragment. Here, multimodality refers to features originating from different information sources or representation methods, aiming to characterize the same molecular fragment from multiple perspectives. Multimodal features include molecular fragment structural features, textual description features, and atomic-level structural features.

[0042] Molecular fragment structure characteristics This is a structured vector representation of the molecular fragment itself. Similar to extracting global molecular structure features, the SMILES string of each molecular fragment can be input into the same or another molecular encoder (such as the Mol2Vec model) to obtain the molecular fragment structure features. These features mainly reflect the fragment's own topological structure. This method can be expressed as: .

[0043] Text description features Chemical semantic knowledge is used to capture molecular fragments. The extraction process typically involves two steps: (1) Input the structural information of the molecular fragment (such as SMILES) into a pre-defined Large Language Model (LLM). By designing specific prompts, guide the LLM to generate a natural language text describing the chemical properties, functions, or structural characteristics of the fragment. For example, for the molecular fragment C(=O)O, you can input the prompt to the LLM: "C(=O)O is a molecular fragment. Please give a descriptive description of its properties in a concise and refined manner." The LLM may output: "It is a carboxylic acid group, has weak acidity, easily forms hydrogen bonds, is highly water-soluble, and often participates in esterification and amidation reactions."

[0044] (2) Text encoding: The descriptive text generated by the large language model is input into a text embedding model (such as ChemBERTa) pre-trained in the field of chemistry or biomedicine, and the natural language text is converted into a fixed-dimensional semantic feature vector, that is, text description features. This feature contains rich chemical knowledge and experience learned from a large number of documents.

[0045] This method can be expressed as: In the formula, Prompt words for inputting a large language model.

[0046] Atomic-level structural features The aim is to capture the internal structural information of fragments at a more refined atomic level. The extraction process is as follows: First, an atomic-level graph is constructed separately for each molecular fragment, with atoms as nodes and chemical bonds as edges. Then, this atomic-level graph is input into an atomic-level graph neural network. This network learns the interactions and fine-grained structural conformations of atoms within the fragment through message passing and aggregation between atomic nodes. Finally, through a readout or pooling operation, the information from the entire fragment graph is aggregated into a single vector as the atomic-level structural features.

[0047] This method can be expressed as: In the formula, It is an atomic-level graph neural network.

[0048] Step 130: Construct a graph structure representation of the target molecule using each molecular fragment as a node; the node features of each node in the graph structure representation are obtained by fusing the multimodal features of each node and the global molecular structure features.

[0049] Specifically, a graph structure representation (also known as a molecular-level graph) is constructed, which is structured as follows: (1) Node: Each molecular fragment obtained after the target molecule is fragmented is a node in the graph.

[0050] (2) Edge: If two molecular fragments are directly connected in the original complete molecule by one or more chemical bonds, then an edge is established between the nodes corresponding to the two fragments. The characteristics of the edge can be further defined, such as the type of connecting bond (single bond, double bond, etc.) and the type of connecting atoms.

[0051] Using the above method, a complex macromolecule is abstracted into a simpler graph structure representation composed of several functional nodes (fragments).

[0052] The node features of each node in this graph representation are obtained by fusing the multimodal features of the node (fragment) itself with the global molecular structure features of the target molecule extracted earlier. The fusion process aims to align and integrate information from different sources and dimensions to generate a unified and more expressive node representation. Fusion can be achieved using various techniques, such as simple vector concatenation, element-wise addition or averaging, as well as more complex attention mechanisms.

[0053] Attention mechanisms can adaptively learn the importance of different modal features and global features for the current prediction task and assign them different weights, thereby achieving weighted fusion of information.

[0054] For example, for a given fragment node, its molecular fragment structure features, textual description features, and atomic-level structure features can be fused using an attention mechanism to obtain a fused vector representation. Then, the fused vector representation is combined with the global molecular structure features. The nodes are then fused to obtain the final node features. This method can be expressed as: ; .

[0055] Step 140: Input the graph structure representation into the molecular-level graph neural network to obtain the property prediction results of the target molecule output by the molecular-level graph neural network.

[0056] Specifically, the role of molecular-level graph neural networks is to facilitate information transfer and aggregation between fragment nodes, thereby learning the interactions between fragments and their synergistic effects on the overall properties of the molecule.

[0057] Molecular-level graph neural networks can be any GNN model commonly used in this field, such as graph convolutional networks, graph attention networks (GAT), message passing neural networks (MPNN), etc. The network updates the representation of each fragment node through computation over several layers, so that it contains not only its own information but also information about its neighboring fragments (i.e., structurally connected fragments).

[0058] After multiple rounds of information propagation, a graph-level readout operation (such as summing or averaging the feature vectors of all nodes, or using more complex pooling methods) yields a final vector representation of the entire molecular-level graph. Finally, this final vector is input into one or more fully connected layers (also known as prediction heads) to output predictions of the target molecule's properties. The property prediction result can be a classification result (such as the probability of being toxic / non-toxic) or a regression result (such as a specific value of solubility). This method can be expressed as: . For molecular-level graph neural networks, This refers to the edge relationships in a graph neural network.

[0059] The molecular property prediction method based on molecular fragmentation and multimodal fusion provided in this application decomposes the complex molecular property prediction problem into a fragment level with clearer chemical meaning by processing the target molecule into molecular fragments, thereby improving the interpretability of molecular property prediction. A graph structure with molecular fragments as nodes is constructed, and multimodal features from multiple information sources such as structure, text, and atomic level are fused into each node, greatly enriching the model's input information and improving the semantic integrity and structural rationality of molecular representation. By learning the interactions between fragments through a molecular-level graph neural network, the key factors determining molecular properties can be captured more accurately, thus significantly improving the accuracy of molecular property prediction.

[0060] It should be noted that each implementation method of this application can be freely combined, rearranged, or executed individually, and does not need to rely on or depend on a fixed execution order.

[0061] In some embodiments, multimodal features are extracted based on the following steps: Molecular fragments are input into the structural semantic coding model to obtain the molecular fragment structural features of the molecular fragments output by the structural semantic coding model. Molecular fragments are input into a pre-defined large language model to obtain descriptive text of the molecular fragments output by the pre-defined large language model. Based on a pre-trained text embedding model in the field of chemistry, the descriptive text is encoded into textual descriptive features of the molecular fragments. Atomic-level graphs are constructed based on the atoms and chemical bonds of molecular fragments. These graphs are then input into an atomic-level graph neural network to obtain the atomic-level structural features of the molecular fragments output by the neural network.

[0062] Specifically, (1) in terms of molecular fragment structure feature extraction: Structural semantic encoding models are models that can map the structural information of molecules or molecular fragments to a high-dimensional vector space. These models are typically obtained through self-supervised pre-training on large, unlabeled molecular databases, enabling them to learn the intrinsic relationships and semantic similarities between chemical substructures.

[0063] In one specific embodiment, the structural semantic encoding model can be the Mol2Vec model.

[0064] The construction of the Mol2Vec model can be divided into three stages: The first stage is molecular fragmentation.

[0065] Using tools such as RDKit, molecules are decomposed into a series of substructures (e.g., atomic environments in Morgan fingerprints), often using atomic neighborhoods with radii of 0 to 2 as molecular fragments. Each fragment is represented as a string (e.g., "[C:1]-[O:2]"), forming a vocabulary similar to that in natural language.

[0066] The second stage is contextual embedding.

[0067] By inputting a sequence of molecular fragments into a shallow neural network, and maximizing the probability of predicting neighboring fragments from the central fragment, semantically similar molecular fragment vectors are learned. If two fragments frequently appear together in similar molecular structures, their vectors will also be close in high-dimensional space.

[0068] The third stage is molecular-level representation generation (Molecular Embedding Aggregation).

[0069] By averaging or weighting the structure vectors of all segments in the same molecule, the global molecular structure characteristics of the entire molecule can be obtained.

[0070] This representation preserves local structural information and has strong global generalization ability.

[0071] The structural features of molecular fragments obtained through the structural semantic coding model not only encode the topological structure of the fragments, but also preserve their proximity relationship with other chemical structures in the semantic space. That is, fragments with similar structures or functions are closer in space in terms of feature vectors.

[0072] (2) In terms of text description feature extraction: The powerful knowledge reasoning and text generation capabilities of large language models can be leveraged to imbue each molecular fragment with rich chemical semantic information. This process can be specifically divided into two sub-steps: First, generate the descriptive text. Choose a pre-defined large language model, which can be a general-purpose language model or a model fine-tuned specifically for the scientific domain. To enable the model to accurately understand the task and generate high-quality descriptions, a standardized prompt template needs to be designed. For example, the following template can be constructed: System Prompt: "You are an expert in organic chemistry and molecular structure analysis, skilled at describing the chemical properties and structural characteristics of molecular fragments in concise and accurate language." User Prompt: "{fragment_smiles} is a molecular fragment. Please provide a descriptive description of its properties. Requirements: 1. Do not repeat the question. 2. Be brief and concise. 3. Begin your description directly." Here, {fragment_smiles} is a placeholder string for the SMILES of the molecular fragments to be processed. After filling the template with the specific fragment SMILES, it is input into the large language model. The model will generate corresponding descriptive text based on its internally stored vast chemical knowledge base, such as: "It is a benzoic acid group, with weak acidity, both aromatic and polar, can form hydrogen bonds, and easily participates in esterification, amidation and other reactions." Secondly, the descriptive text is encoded. The generated text is input into a pre-trained text embedding model in the field of chemistry. Such models (e.g., ChemBERTa) are specifically trained on massive amounts of text such as chemical literature and patents, enabling them to deeply understand specialized semantics such as chemical terminology, reaction types, and structure-activity relationships. This model transforms the input descriptive text into a high-dimensional feature vector, i.e., textual descriptive features. This feature vector can be viewed as the coordinates of the molecular fragment in the chemical knowledge space, capturing implicit knowledge described in books and literature that is difficult to express using traditional structural features.

[0073] (3) In terms of atomic-level structural feature extraction: First, construct an atomic-level graph. For each molecular fragment, represent it as a graph structure. The set of nodes represents all the atoms in the fragment, and the set of edges represents the chemical bonds connecting these atoms.

[0074] Initial node features can consist of a series of physicochemical properties, such as atomic number, atomicity (number of bonds connected), formal charge, hybridization type, and whether it is part of an aromatic ring. This information can be easily obtained using standard cheminformatics toolkits. Initial edge features can include: bond type (single, double, triple, aromatic), whether it is conjugated, whether it is within a ring, and stereochemical information (such as cis-trans isomerism).

[0075] Then, the constructed atomic-level graph and its initial features are input into an atomic-level graph neural network. This network updates the representation of nodes (atoms) through an iterative message-passing mechanism. In each layer, each atom aggregates information passed from its neighboring atoms through chemical bonds and updates its own feature vector based on its current state.

[0076] Finally, a readout operation aggregates the final feature vectors of all atoms in the fragment into a single vector, which serves as the atomic-level structural feature of the fragment. The pooling operation can be a simple mean pooling or summation pooling. This feature vector highly condenses the fine, atomic-level structural information within the fragment.

[0077] The molecular property prediction method based on molecular fragmentation and multimodal fusion provided in this application ensures that the subsequent fusion steps can obtain the most comprehensive and richest information input through a multi-angle and multi-level feature extraction strategy, thus providing a solid data foundation for achieving high-precision molecular property prediction.

[0078] In some embodiments, the atomic-level graph neural network is a Directed Message Passing Neural Network (D-MPNN).

[0079] Specifically, for atomic-level graphs of molecular fragments, this application employs D-MPNN. D-MPNN is a graph neural network specifically designed for chemical molecules. During information transfer, the information flow is not from node to node, but rather from directed edges (bonds) to the target atom. The specific steps for atomic-level structure feature extraction are as follows: Step 1: Prepare data input.

[0080] The data input required for D-MPNN includes: atomic feature matrix, edge index, and edge feature.

[0081] An atomic feature matrix is ​​constructed from the atomic-level diagrams built from each molecular fragment.

[0082] Each atomic feature and edge feature is initialized using the RDKit library. Atomic features can include atomic number, number of bonds, charge, hybridization type, and aromaticity. Edge features can include bond type, whether it is conjugated, whether it is on a ring, and stereoisomer information.

[0083] Step 2: Construct directional edge pairs.

[0084] For each undirected edge (i, j), create two directed edges (i, j) and (j, i). i and j are the indices of the nodes (atoms) in the atomic-level graph.

[0085] Step 3: Initialize the hidden state of each directed edge.

[0086] The initial features are the features of the two connected atoms and the bond. A multi-layer perceptron (MLP) is used to process the atomic and bond features to generate a hidden state vector for each directed edge.

[0087] Step 4: Perform T edge-center message passing.

[0088] For each edge (i, j), collect the states of all edges that "enter i but do not come from j" (to avoid loops), sum these edge states, and feed them into the MLP to obtain the new state.

[0089] Step 5: Aggregate from edge to node.

[0090] After updating the state of all directed edges, the sum of the edge states entering a certain node is used as the final representation of that node.

[0091] This step transmits the directional information back to the atom, resulting in an atomic-level feature representation.

[0092] Step 6: The final feature vectors of all atoms are aggregated into a single vector.

[0093] By performing mean pooling or summation pooling on the atoms in the entire fragment diagram, the atomic-level structural features of the fragments can be obtained.

[0094] The molecular property prediction method based on molecular fragmentation and multimodal fusion provided in this application adopts a directed message-passing neural network as an atomic-level graph neural network, which improves the interpretability and information transmission efficiency of the model and provides higher quality atomic-level structural features with clearer chemical meaning.

[0095] In some embodiments, the node features of each node in the graph structure representation are obtained based on the following steps: Attention alignment and fusion are performed on at least two of the molecular fragment structure features, text description features and atomic-level structure features of each molecular fragment to obtain the fused structure features of each molecular fragment. By fusing the fusion structural features of each molecular fragment with the global molecular structural features, the node features of each molecular fragment in the graph structure representation are obtained.

[0096] Specifically, the core idea of ​​attention alignment is that when fusing features from different modalities, they are not simply treated equally (e.g., directly concatenated or averaged). Instead, the model autonomously learns the relative importance of different modal information for the current task or context, and assigns different weights accordingly. This can significantly improve the fusion effect and bridge the differences between features from different modalities in the semantic space.

[0097] In a specific embodiment, alignment can be achieved using a multi-head self-attention mechanism or a multi-head cross-attention mechanism. Taking the fusion of three modal features (molecular fragment structure features, text description features, and atomic-level structure features) as an example, the process can be described as follows: Step 1: Linear projection.

[0098] A learnable linear transformation matrix is ​​applied to the features (embedding vectors) of each modality, mapping them to a semantic space of uniform dimension: .

[0099] In the formula, For embedding vectors, , and These are linear transformation matrices, , and These are the query vector, key vector, and value vector, respectively.

[0100] Step 2: Construct a multi-head cross-attention mechanism.

[0101] Within each attention head, a specific modality is selected as the query, and other modalities are used as key-value pairs. Attention weights are calculated, and information is aggregated. .

[0102] In the formula, The dimension of the query vector.

[0103] Attention weights reflect the relevance between other modalities (Key) and the dominant modality (Query). Each attention head uses the modality of the graph network as a query to capture different cross-modal relationships.

[0104] Step 3: Multi-head output fusion and nonlinear mapping.

[0105] The outputs of all attention heads are concatenated and then non-linearly mapped through a fully connected layer to obtain the final fused node feature representation: .

[0106] In the formula, For the output of each attention head, For concatenation functions, This is the mapping matrix of the fully connected layer. These are the node features after fusion.

[0107] Step 4: Residual connection and normalization.

[0108] To enhance training stability, a graph neural network is used as the primary network. Residual connections are added after fusion, and layer normalization is performed, which can be represented as: .

[0109] in, Characterized by molecular fragment structure. For layer normalization function, This is the final node feature.

[0110] The fused node features are used as the initial feature input for the corresponding molecular fragment nodes in the graph neural network, supporting subsequent graph structure aggregation and molecular property prediction tasks.

[0111] Using the above method, we can first perform attention alignment and fusion on at least two of the molecular fragment structure features, text description features and atomic-level structure features of each molecular fragment to obtain the fused structure features of each molecular fragment; then we fuse the fused structure features of each molecular fragment with the global molecular structure features (attention mechanism can still be used at this stage) to obtain the node features of each molecular fragment in the graph structure representation.

[0112] Figure 2 This is a flowchart of the molecular property prediction method based on molecular fragmentation technology and combining multimodal information with graph neural networks provided in this application, such as... Figure 2 As shown, the method includes: First, the target molecule is subjected to molecular fragmentation to obtain multiple molecular fragments.

[0113] Secondly, global molecular structural features of the target molecule are extracted using a structural semantic encoding model.

[0114] Next, each molecular fragment is input into a structural semantic encoding model to obtain the structural features of each molecular fragment; each molecular fragment is input into a pre-defined large language model and a text embedding model to obtain the textual description features of each molecular fragment; and the atomic-level graphs of each molecular fragment are input into an atomic-level graph neural network to obtain the atomic-level structural features of each molecular fragment.

[0115] Finally, using each molecular fragment as a node, the node features of each node are obtained by fusing the multimodal features of each node and the global molecular structure features, and a graph structure representation of the target molecule is constructed. The graph structure representation is then input into a molecular-level graph neural network to obtain the property prediction results of the target molecule.

[0116] The molecular property prediction method based on molecular fragmentation and multimodal fusion provided in this application first performs attention alignment on molecular fragment structural features, textual description features, and atomic-level structural features, effectively bridging the semantic gap between different modal features and achieving deep fusion and complementarity of information within fragments. Then, the fused structural features of each molecular fragment are fused with global molecular structural features, so that the representation of each node has a global perspective. This greatly enhances the quality and information density of graph structure representation, providing high-quality input for subsequent molecular-level graph neural networks to learn the complex interactions between fragments, thereby significantly improving the performance and robustness of the entire prediction model.

[0117] In some embodiments, at least two of the molecular fragment structural features, textual description features, and atomic-level structural features of each molecular fragment are attention-aligned and fused to obtain the fused structural features of each molecular fragment, including: Molecular attribute information of each molecular fragment is obtained based on a chemical information database, and the molecular attribute information is encoded into molecular attribute features. Attention alignment and fusion are performed on the molecular fragment structure features, text description features, atomic-level structure features and molecular attribute features of each molecular fragment to obtain the fused structure features of each molecular fragment.

[0118] Specifically, molecular property information refers to information that can be numerically quantified to describe the basic physical, chemical, or topological properties of a molecule or molecular fragment.

[0119] Molecular property information can be obtained through standard chemical information repositories or toolkits, such as the widely used open-source cheminformatics toolkit (RDKit). For each molecular fragment, a series of property values ​​can be calculated and extracted. These properties may include, but are not limited to, physical properties, structure counts, and electronic properties.

[0120] After obtaining these numerical attributes, they need to be encoded into a fixed-dimensional vector, namely molecular attribute features. These attribute values ​​can be non-linearly mapped into a more compact embedding space using a multilayer perceptron (MLP) to learn the potential associations between attributes.

[0121] Based on the original three modal features (molecular fragment structure features, text description features, and atomic-level structure features), newly extracted molecular attribute features were added, forming a feature set containing four modalities.

[0122] Attention alignment and fusion of the feature vectors from these four modalities yield the fused structural features of each molecular fragment. This method can also be expressed as: .

[0123] In the formula, These are molecular properties.

[0124] The molecular property prediction method based on molecular fragmentation and multimodal fusion provided in this application introduces molecular attribute features in addition to multimodal feature fusion. By using an attention mechanism, these four modal features are intelligently fused, making the final generated fused structural features more comprehensive and accurate. This allows for a more precise characterization of the essential properties of molecular fragments, thereby providing stronger support for improving the final performance of the entire molecular property prediction model.

[0125] In some embodiments, the method further includes: Based on the distance between atomic-level structural features and molecular fragment structural features, and the distance between atomic-level structural features and text description features, the alignment loss function is determined; Based on the alignment loss function and the main task loss function used for molecular property prediction, the total loss function for training atomic-level graph neural networks and molecular-level graph neural networks is determined.

[0126] Specifically, the purpose of the alignment loss function is to explicitly constrain the consistency of feature vectors from different modalities in the embedding space during training. The basic idea is that for the same molecular fragment, although its features come from different modalities (e.g., one from a structural encoder and the other from a text description), they describe the same chemical entity, so in an ideal high-dimensional semantic space, their representations should be similar.

[0127] To ensure semantic consistency among atomic-level structural features, molecular fragment structural features, and textual description features, an alignment loss function can be defined as follows: .

[0128] In the formula, The alignment loss function is defined as the sum of distances between these different modal feature vectors. Distance can be measured in various ways, such as Euclidean distance (L2 norm), which calculates the square of the straight-line distance between vectors. Features of atomic-level structure (features output by atomic-level GNN). These are the structural features of molecular fragments (features output by the structural semantic encoding model). Describe textual features (features determined based on the output of a large language model).

[0129] The alignment loss function can be introduced as an auxiliary term into the total loss function and jointly optimized with the main task loss (such as the mean squared error loss of property prediction) to form the final training objective, which can be expressed as: .

[0130] In the formula, This is the total loss function; Loss to the main task; The hyperparameter used to adjust the weight of the alignment loss in the overall objective is usually set between 0.1 and 1.0. This parameter is used to balance the relationship between representation consistency and task performance.

[0131] The main task loss function is directly related to the model's final prediction task.

[0132] For regression tasks (such as predicting continuous values ​​like solubility and boiling point), the mean squared error (MSE) loss function is typically used as the primary task loss function. It calculates the mean of the squared differences between the predicted and actual values. Mean absolute error (MAE) can also be used.

[0133] For classification tasks (such as predicting whether a molecule is toxic or active), the main task loss function is usually the cross-entropy loss function, such as binary cross-entropy (BCE) or multi-class cross-entropy.

[0134] The molecular property prediction method based on molecular fragmentation and multimodal fusion provided in this application can significantly improve the fusion effect between different information sources by introducing the above-mentioned alignment loss, thereby enhancing the robustness and prediction accuracy of the model under low resource conditions.

[0135] In some embodiments, the molecular-level graph neural network is a graph attention network (GAT).

[0136] Specifically, GAT's core advantage lies in its introduction of an attention mechanism to process graph-structured data. When aggregating neighbor node information, GAT can dynamically and adaptively assign different attention weights to different neighbor nodes. This means the model can autonomously learn which adjacent molecular fragments have a greater impact on the properties of the central fragment within the current molecular context.

[0137] In the graph structure representation, each node is a molecular fragment, and edges represent bonds between fragments. A molecular-level graph neural network can be constructed using GAT. Inputting the graph structure representation into GAT yields predicted properties of the target molecule. Specific steps include: Step 1: Prepare data input.

[0138] The data required by GAT includes the node feature vector of each molecular fragment and the edge connections between the molecular fragments. Each molecular fragment node needs to have an aligned vector constructed. This embodiment employs an attention alignment mechanism to dynamically adjust the weights of features from different modalities. Finally, the feature vector of each molecular fragment node is obtained. The connections between fragments are found, and edges are added between nodes in the graph. GAT can adaptively allocate attention.

[0139] Step 2, Model Propagation.

[0140] The input features of each node are first calculated using the same weight matrix. For each pair of neighboring nodes, an attention score is calculated. For each node, the attention scores of all neighbors are processed by a normalized exponential (Softmax) function to obtain weight coefficients. These weight coefficients are then used for weighted aggregation to obtain a new node representation. Subsequent linear layers are added based on the specific task to obtain the property prediction results.

[0141] The molecular property prediction method based on molecular fragmentation and multimodal fusion provided in this application uses a graph attention neural network as a molecular-level graph neural network, which can intelligently identify and weight the interactions between different fragments, and can more deeply understand the complex structural property relationships, ultimately achieving higher accuracy in molecular property prediction.

[0142] In some embodiments, the overall architecture of this application consists of two graph neural networks: D-MPNN and GAT.

[0143] The target molecule's SMILES string is input into the model. The model uses a graph neural network to aggregate and predict the results based on the molecular fragments and the structural features of the molecule. The model parameters are optimized by calculating the difference between the predicted reaction type and the actual reaction type output by the model.

[0144] The task of predicting the molecular chemical properties of the model is divided into two parts: regression and classification.

[0145] For classification tasks, the cross-entropy loss function is used as the optimization objective to calculate the difference between the predicted property type and the true property type output by the model. Loss function It can be represented as: .

[0146] In the formula, K represents the total number of reaction types. Indicates the first A true label for each reaction type The model predicts the first... The probability that a reaction type is a true reaction type.

[0147] For regression tasks, the mean squared error loss function is used as the optimization objective to calculate the difference between the model's predicted values ​​and the actual values. Loss function It can be represented as: .

[0148] In the formula, It is the sample size. It is the first The true value of each sample It is the first The predicted value for each sample.

[0149] The Adam optimizer is used in conjunction with a learning rate scheduling strategy for model training.

[0150] Once the model has been trained, it can be tested on both classification and regression tasks.

[0151] For classification tasks, the Receiver Operating Characteristic-Area Under Curve (ROC-AUC) can be used as a test metric to measure the model's ability to distinguish between positive and negative samples.

[0152] .

[0153] TPR is the true positive rate, calculated using the following formula: Where TP is the true positive count, FN is the false negative count, and FPR is the false positive rate, calculated using the following formula: Where FP is the number of false positives and TN is the number of true negatives.

[0154] For regression tasks, RMSE and MAE are used to measure prediction error. RMSE is a performance metric used to evaluate regression models. It calculates the square root of the average of the squares of the differences between predicted and true values.

[0155] In the formula, It is the sample size. It is the first The true value of each sample It is the first The predicted value for each sample.

[0156] MAE is a performance metric used to evaluate regression models. It calculates the average of the absolute differences between predicted and actual values.

[0157] .

[0158] During model training, the Moleculenet molecular property prediction benchmark dataset was used. This dataset includes both regression and prediction tasks, as well as datasets such as Quantum Mechanics (QM), Physical Chemistry (PC), and Drug Discovery.

[0159] Figure 3 This is the second flowchart of the molecular property prediction method based on molecular fragmentation and multimodal fusion provided in this application, as shown below. Figure 3 As shown, the method includes: Step 310: Obtain the predicted target molecule SMILES dataset and use the molecular fragmentation method (MacFrag) to cut the SMILES string into molecular fragments. Simultaneously, input the molecular SMILES string into the Mol2Vec model to obtain the embedding vector representation of the entire molecule.

[0160] Step 320: Input the obtained molecular fragments into the large language model using the specified prompt template to obtain descriptive language for the molecular fragments. Generally, the large language model will answer the main properties of the molecular fragments, store the descriptive language obtained for all molecular fragments in a JSON file, and select an embedding model to convert its content into vectors.

[0161] Step 330: Construct an atomic-level graph neural network. For an input molecule, its molecular fragments are used as nodes in the graph, and the bonds connecting the fragments are used as edges. The D-MPNN network is used to aggregate the nodes and edges, and finally the embedding vector representation for each molecular fragment is obtained.

[0162] Step 340: Preprocess the dataset by using the RDKit library to filter out invalid SMILES data and mask the missing parts of the dataset.

[0163] Step 350: Encode each molecular fragment using the same Mol2Vec model to obtain the embedding vector representation of the molecular fragment.

[0164] Step 360: Align the molecular fragment embedding vector representation output by the Mol2Vec model, the vector representation converted from the content description of the large language model, and the molecular fragment embedding vector representation output by the D-MPNN model using an attention mechanism, and fuse them to obtain the molecular fragment embedding vector representation.

[0165] Step 370: Construct a molecular-level graph neural network using GAT, with molecular fragments as nodes. Align the molecular fragment embedding vector representations obtained in the previous step with the embedding vector representations of the entire molecule using an attention mechanism, and fuse them to obtain the final vector representation of the molecular fragments. This vector serves as the node feature of the molecular-level graph neural network, is aggregated, and finally, the prediction result is output according to the specific prediction task.

[0166] During model training parameter optimization, systematic experiments were conducted on different combinations of loss functions and hyperparameters. The model used a hidden layer dimension (hidden_dim) of 256 and a learning rate of [missing value]. Based on this, the effects of different numbers of attention heads (num_heads) and loss function designs on model performance (AUC) were compared.

[0167] Table 1. Prediction results of each parameter model

[0168] As shown in Table 1, the experimental results show that the average AUC is 0.839 under the basic binary cross-entropy loss (BCE); when 8 attention heads are used and a weighted binary cross-entropy loss (weight=9) is introduced, the average AUC increases to 0.857; when a hybrid loss combining binary cross-entropy loss and focal loss is further combined, the average AUC is 0.850; when the class weight parameter in the loss function is set to a learnable weight, the average AUC reaches 0.860, achieving the best performance.

[0169] The results show that using learnable weighted BCE loss can effectively improve prediction accuracy under class imbalance conditions while maintaining training stability. The model uses only one layer of graph neural network, and subsequent layers can be added to observe the improvement in prediction performance.

[0170] The apparatus provided in the embodiments of this application is described below. The apparatus described below can be referred to in correspondence with the method described above.

[0171] Figure 4This is a schematic diagram of the molecular property prediction device based on molecular fragmentation and multimodal fusion provided in this application, as shown below. Figure 4 As shown, the device includes: The molecular fragmentation module 410 is used to perform molecular fragmentation processing on the target molecule to obtain at least one molecular fragment. The feature extraction module 420 is used to extract the global molecular structure features of the target molecule and the multimodal features of each molecular fragment; the multimodal features include at least two of the following: molecular fragment structure features, text description features, and atomic-level structure features; Graph structure representation module 430 is used to construct a graph structure representation of the target molecule with each molecular fragment as a node; the node features of each node in the graph structure representation are obtained by fusing the multimodal features of each node and the global molecular structure features; Property prediction module 440 is used to input the graph structure representation into the molecular-level graph neural network and obtain the property prediction results of the target molecule output by the molecular-level graph neural network.

[0172] The molecular property prediction device based on molecular fragmentation and multimodal fusion provided in this application decomposes the complex molecular property prediction problem into a fragment level with clearer chemical meaning by processing the target molecule into molecular fragments, thereby improving the interpretability of molecular property prediction. It constructs a graph structure with molecular fragments as nodes and integrates multimodal features from multiple information sources such as structure, text, and atomic level for each node, which greatly enriches the input information of the model and improves the semantic integrity and structural rationality of molecular representation. By learning the interaction between fragments through molecular-level graph neural networks, it can more accurately capture the key factors that determine molecular properties, thereby significantly improving the accuracy of molecular property prediction.

[0173] Figure 5 This is a schematic diagram of the structure of the electronic device provided in this application, such as... Figure 5 As shown, the electronic device may include: a processor 510, a communications interface 520, a memory 530, and a communications bus 540, wherein the processor, communications interface, and memory communicate with each other via the communications bus. The processor can invoke logical commands stored in the memory to execute the methods described in the above embodiments, for example: The target molecule is fragmented to obtain at least one molecular fragment; the global molecular structure features of the target molecule and the multimodal features of each molecular fragment are extracted; the multimodal features include at least two of the following: molecular fragment structure features, textual description features, and atomic-level structure features; a graph structure representation of the target molecule is constructed using each molecular fragment as a node; the node features of each node in the graph structure representation are obtained by fusing the multimodal features of each node and the global molecular structure features; the graph structure representation is input into a molecular-level graph neural network to obtain the property prediction results of the target molecule output by the molecular-level graph neural network.

[0174] Furthermore, the logical commands in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several commands to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0175] The processor in the electronic device provided in this application embodiment can call logical instructions in the memory to implement the above method. Its specific implementation method is the same as the aforementioned method implementation method and can achieve the same beneficial effect, which will not be repeated here.

[0176] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the methods provided in the above embodiments.

[0177] The specific implementation method is the same as the aforementioned method implementation method and can achieve the same beneficial effects, so it will not be repeated here.

[0178] This application provides a computer program product, including a computer program that, when executed by a processor, implements the method described above.

[0179] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0180] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0181] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for predicting molecular properties based on molecular fragmentation and multimodal fusion, characterized in that, include: The target molecule is subjected to molecular fragmentation to obtain at least one molecular fragment; Extract the global molecular structure features of the target molecule and the multimodal features of each molecular fragment; the multimodal features include at least two of the following: molecular fragment structure features, text description features, and atomic-level structure features; A graph structure representation of the target molecule is constructed using each molecular fragment as a node; the node features of each node in the graph structure representation are obtained by fusing the multimodal features of each node and the global molecular structure features. The graph structure representation is input into a molecular-level graph neural network to obtain the property prediction results of the target molecule output by the molecular-level graph neural network.

2. The molecular property prediction method based on molecular fragmentation and multimodal fusion according to claim 1, characterized in that, The multimodal features are extracted based on the following steps: Molecular fragments are input into a structural semantic coding model to obtain the molecular fragment structural features of the molecular fragments output by the structural semantic coding model. Molecular fragments are input into a preset large language model to obtain the descriptive text of the molecular fragments output by the preset large language model. The descriptive text is then encoded into the textual descriptive features of the molecular fragments based on a pre-trained text embedding model in the field of chemistry. An atomic-level graph is constructed based on the atoms and chemical bonds of the molecular fragments. The atomic-level graph is then input into an atomic-level graph neural network to obtain the atomic-level structural features of the molecular fragments output by the atomic-level graph neural network.

3. The molecular property prediction method based on molecular fragmentation and multimodal fusion according to claim 1, characterized in that, The node features of each node in the graph structure representation are obtained based on the following steps: Attention alignment and fusion are performed on at least two of the molecular fragment structure features, text description features and atomic-level structure features of each molecular fragment to obtain the fused structure features of each molecular fragment. The fusion structural features of each molecular fragment are fused with the global molecular structural features to obtain the node features of each molecular fragment in the graph structural representation.

4. The molecular property prediction method based on molecular fragmentation and multimodal fusion according to claim 3, characterized in that, The attention alignment and fusion of at least two of the molecular fragment structural features, textual description features, and atomic-level structural features of each molecular fragment to obtain the fused structural features of each molecular fragment includes: Molecular attribute information of each molecular fragment is obtained based on a chemical information database, and the molecular attribute information is encoded into molecular attribute features. Attention alignment and fusion are performed on the molecular fragment structure features, text description features, atomic-level structure features and molecular attribute features of each molecular fragment to obtain the fused structure features of each molecular fragment.

5. The molecular property prediction method based on molecular fragmentation and multimodal fusion according to claim 2, characterized in that, The method further includes: Based on the distance between the atomic-level structural features and the molecular fragment structural features, and the distance between the atomic-level structural features and the text description features, an alignment loss function is determined; Based on the alignment loss function and the main task loss function used for molecular property prediction, the total loss function for training the atomic-level graph neural network and the molecular-level graph neural network is determined.

6. The molecular property prediction method based on molecular fragmentation and multimodal fusion according to claim 5, characterized in that, The atomic-level graph neural network is a directed message passing neural network; the molecular-level graph neural network is a graph attention neural network.

7. A molecular property prediction device based on molecular fragmentation and multimodal fusion, characterized in that, include: The molecular fragmentation module is used to perform molecular fragmentation processing on the target molecule to obtain at least one molecular fragment; The feature extraction module is used to extract the global molecular structure features of the target molecule and the multimodal features of each molecular fragment; the multimodal features include at least two of the following: molecular fragment structure features, text description features, and atomic-level structure features; The graph structure representation module is used to construct a graph structure representation of the target molecule with each molecular fragment as a node; the node features of each node in the graph structure representation are obtained by fusing the multimodal features of each node and the global molecular structure features; The property prediction module is used to input the graph structure representation into a molecular-level graph neural network to obtain the property prediction results of the target molecule output by the molecular-level graph neural network.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the molecular property prediction method based on molecular fragmentation and multimodal fusion as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the molecular property prediction method based on molecular fragmentation and multimodal fusion as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the molecular property prediction method based on molecular fragmentation and multimodal fusion as described in any one of claims 1 to 6.

Images