A multimodal molecular attribute prediction method and system based on hierarchical divergence perception uncertainty fusion

Abstract

This invention discloses a multimodal molecular property prediction method and system based on hierarchical divergence-aware uncertainty fusion, relating to the fields of artificial intelligence, drug discovery, and molecular property prediction. The invention includes data acquisition, sequence text encoding, molecular fingerprint construction, molecular structure preprocessing, 3D conformation generation, and divergence-aware uncertainty fusion. By representing the same molecule as SMILES sequences, molecular fingerprints, two-dimensional molecular graphs, and three-dimensional geometric structures, this invention extracts molecular features from multiple perspectives, including sequence semantics, structural descriptors, topological connectivity, and spatial geometric constraints. Furthermore, based on a hierarchical fusion mechanism, it models the degree of divergence between different modal features and the inherent uncertainty of each modality, dynamically adjusting the contribution weight of each modality in the fusion process to reduce the impact of low-quality modalities on the prediction results. This invention improves the accuracy, stability, and generalization ability of predicting molecular toxicity, solubility, bioactivity, and physicochemical properties, and is applicable to drug screening, molecular property evaluation, candidate compound discovery, and AI-assisted drug development.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence, bioinformatics, cheminformatics and drug discovery, and in particular to a multimodal molecular property prediction method and system based on hierarchical divergence perception uncertainty fusion. Background Technology

[0002] Molecular property prediction is a crucial task in computational drug discovery, materials design, and cheminformatics. Its goal is to predict the biological activity, toxicity, solubility, lipophilicity, free energy, and other properties of a given molecule based on its structural information. Accurate prediction of molecular properties can assist in the discovery of potential candidate compounds in the early screening stages, reduce experimental costs, and improve the efficiency of drug development.

[0003] Existing methods for predicting molecular properties typically rely on a specific molecular representation for modeling, such as SMILES sequences, molecular fingerprints, two-dimensional molecular graphs, or three-dimensional geometric conformations. SMILES sequences can represent molecular structures as strings, facilitating the extraction of local fragment patterns using natural language processing or sequence modeling methods. Molecular fingerprints can characterize statistical features such as functional groups, substructures, and pharmacophores through predefined rules or hash encoding. Two-dimensional molecular graphs can describe the topological structure of molecules with atoms as nodes and chemical bonds as edges. Three-dimensional geometric conformations can further provide information such as bond lengths, spatial arrangement, and local conformational constraints. While these representations characterize molecular structures from different perspectives, a single modality is insufficient to fully express the complex structure and multi-level properties of molecules.

[0004] To enhance the completeness of molecular characterization, multimodal molecular attribute prediction methods have gained increasing attention. These methods typically fuse information such as smils, fingerprints, graph structures, and 3D geometry to obtain richer molecular representations. However, existing fusion methods often employ fixed rules such as simple summation, averaging, max pooling, or concatenation, assuming that different modalities contribute similarly to the final prediction, making it difficult to automatically adjust modal importance based on different molecular samples and tasks. While some attention- or gating-based methods can learn modal weights, they usually focus on feature weighting and lack explicit modeling of the reliability of different modalities themselves and the degree of cross-modal conflict.

[0005] In practical applications, the information quality of different modalities is not always consistent. For example, SMILES sequences are affected by linearized writing methods, molecular fingerprints rely on predefined rules, two-dimensional graphs focus on topological structure but lack spatial conformation, and three-dimensional geometric conformations may be affected by conformation generation algorithms and energy optimization processes. When different modalities describe the same molecule differently or conflict, using a fixed ratio for fusion can easily introduce noise and weaken prediction performance. Therefore, how to explicitly measure modal divergence and modal uncertainty in multimodal molecular property prediction, and adaptively adjust the contributions of different modalities based on this, is a pressing problem in the field of molecular machine learning.

[0006] To address the aforementioned issues, this invention proposes a multimodal molecular attribute prediction method and system based on hierarchical divergence-aware uncertainty fusion. By constructing four modal representations—SMILES sequence, molecular fingerprint, two-dimensional molecular graph, and three-dimensional geometry—and introducing a divergence-aware uncertainty fusion mechanism during the hierarchical fusion process, modal weights are dynamically allocated based on the modality's own uncertainty and the degree of cross-modal divergence, thereby improving the accuracy and stability of molecular attribute prediction. Summary of the Invention

[0007] In view of this, the present invention provides a multimodal molecular attribute prediction method and system based on hierarchical divergence-aware uncertainty fusion. The method first constructs a sequence text modality, a molecular fingerprint modality, a two-dimensional molecular graph modality, and a three-dimensional geometric modality from the molecular SMILES string. Then, it performs hierarchical fusion step-by-step according to a hierarchical path: priority fusion of molecular fingerprint and two-dimensional graph, supplementation of three-dimensional geometric information, and final supplementation of SMILES sequence semantics. At each fusion stage, local fusion weights are calculated through an uncertainty estimation network and a divergence mapping network, and these local fusion weights are progressively expanded into the contribution weights of the four original modalities to the final prediction. Finally, the fused representation is input into the prediction module, which outputs classification or regression results.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] A multimodal molecular property prediction method based on hierarchical divergence-perceived uncertainty fusion includes the following steps:

[0010] S1 Data Acquisition Steps: Acquire a molecular dataset containing molecular structure information and its corresponding attribute labels, which must include at least the string "SMILES" and attribute labels including classification labels or regression numerical labels;

[0011] S2 Sequence Text Feature Encoding Steps: Convert the SMILES string of each molecule into a fixed-length sequence index and input it into a multi-scale one-dimensional convolutional neural network to extract the sequence semantic feature representation of the molecule;

[0012] S3 Molecular fingerprint feature construction steps: Extract multiple molecular fingerprint features based on the preprocessed molecular object, and concatenate different types of fingerprint vectors to form a molecular structure descriptor feature representation;

[0013] S4 Two-Dimensional Molecular Graph Construction Steps: Represent each molecule as a two-dimensional graph structure, where atoms are nodes and chemical bonds are edges. Construct node features, edge connection relationships, and edge features, and input them into a graph neural network to extract two-dimensional topological structure feature representations.

[0014] S5 Three-dimensional conformation generation steps: Generate a three-dimensional conformation based on the molecular structure, extract the three-dimensional coordinates of heavy atoms and calculate the edge distance features to obtain a three-dimensional geometric feature representation;

[0015] S6 Disagreement-aware uncertainty fusion step: Based on the degree of disagreement between different modal features and the uncertainty of each modality itself, the different modal features are dynamically weighted and fused to generate fused feature representations step by step;

[0016] S7 Molecular Attribute Prediction Step: Input the final fused molecular characterization into the prediction module and output the molecular attribute prediction results.

[0017] As can be seen from the above technical solution, compared with the prior art, the present invention has the following beneficial effects:

[0018] 1) By simultaneously constructing four modalities—SMILES sequence, molecular fingerprint, two-dimensional molecular graph, and three-dimensional geometric conformation—the molecular structure can be characterized from multiple perspectives, including sequence semantics, structural fragments, topological connections, and spatial geometry, thus alleviating the problem of incomplete information in single-molecule representation.

[0019] 2) By adopting a hierarchical fusion strategy, molecular fingerprint modes and two-dimensional graph modes with similar information sources and strong complementarity are prioritized for fusion, and then three-dimensional geometric modes and SMILES sequence modes are introduced to make the fusion process have clear structural logic and interpretability.

[0020] 3) By using the divergence-aware uncertainty fusion mechanism, the uncertainty of the mode itself and the degree of cross-modal divergence are explicitly estimated at each fusion stage, and the mode contribution weight is dynamically adjusted accordingly, which can reduce the interference of low reliability or high conflict modes on the final prediction.

[0021] 4) The system supports both classification and regression molecular attribute prediction tasks and is applicable to molecular attribute prediction datasets such as BBBP, BACE, ClinTox, Tox21, SIDER, FreeSolv, ESOL, and Lipo. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments are briefly introduced below. Obviously, the drawings described below are merely embodiments of the present invention, and those skilled in the art can obtain other drawings based on the provided drawings without any creative effort.

[0023] Figure 1 This is a flowchart of a multimodal molecular attribute prediction method based on hierarchical divergence perception uncertainty fusion disclosed in this invention;

[0024] Figure 2 This is a schematic diagram of the principle framework of a multimodal molecular attribute prediction method based on hierarchical divergence perception uncertainty fusion disclosed in this invention.

[0025] Figure 3 This is a block diagram of a multimodal molecular property prediction system based on hierarchical divergence perception uncertainty fusion disclosed in this invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. The described embodiments are only some embodiments of the present invention, and not all embodiments. Other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are all within the protection scope of the present invention.

[0027] In this application, 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 an actual sequential relationship between these entities or operations. The terms "comprising," "including," or 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 the listed elements but may also include other elements not expressly listed.

[0028] See Figure 1 and Figure 2 As shown in the figure, this embodiment discloses a multimodal molecular property prediction method based on hierarchical divergence-aware uncertainty fusion. Its overall process includes data acquisition, sequence text feature encoding, molecular fingerprint feature construction, two-dimensional molecular graph construction, three-dimensional conformation generation, divergence-aware uncertainty fusion, and molecular property prediction.

[0029] S1 Data Acquisition Steps: Acquire a molecular dataset containing molecular structure information and its corresponding attribute labels, which must include at least the string SMILES, and the attribute labels must include classification labels or regression numerical labels;

[0030] S2 Sequence Text Feature Encoding Steps: Convert the SMILES string of each molecule into a fixed-length sequence index and input it into a multi-scale one-dimensional convolutional neural network to extract the sequence semantic feature representation of the molecule;

[0031] S3 Molecular fingerprint feature construction steps: Extract multiple molecular fingerprint features based on the preprocessed molecular object, and concatenate different types of fingerprint vectors to form a molecular structure descriptor feature representation;

[0032] S4 2D molecular graph construction steps: Represent each molecule as a 2D graph structure, where atoms are nodes and chemical bonds are edges. Construct node features, edge connection relationships, and edge features, and input them into a graph neural network to extract 2D topological structure feature representation;

[0033] S5 3D conformation generation steps: Generate a 3D conformation based on the molecular structure, extract the 3D coordinates of heavy atoms, and obtain the 3D geometric feature representation of the molecule;

[0034] S6 Disagreement-aware uncertainty fusion step: Based on the degree of disagreement between different modal features and the uncertainty of each modality itself, the different modal features are dynamically weighted and fused, and then the fused feature representation is generated step by step;

[0035] S7 Molecular Attribute Prediction Steps: Input the final fused molecular characterization into the prediction module, and output the molecular attribute prediction results. The evaluation metric for the classification task is ROC-AUC, and the evaluation metric for the regression task is RMSE.

[0036] Furthermore, S1 acquires the molecular dataset. In this embodiment, the molecular dataset is represented as: ,in, Indicates the first The SMILES string for each molecule, This represents the attribute label corresponding to the molecule, and N represents the number of molecule samples.

[0037] This implementation is applicable to several commonly used molecular attribute prediction datasets in MoleculeNet, including BBBP, BACE, ClinTox, Tox21, SIDER, FreeSolv, ESOL, and Lipo. ROC-AUC is primarily used as the performance evaluation metric for classification tasks, while RMSE is primarily used for regression tasks.

[0038] Table 1 shows the typical molecular property prediction datasets that can be supported in this embodiment.

[0039] BBBP Classification 1 ROC-AUC BACE Classification 1 ROC-AUC ClinTox Classification 2 ROC-AUC Tox21 Classification 12 ROC-AUC SIDER Classification 27 ROC-AUC FreeSolv return 1 RMSE ESOL return 1 RMSE Lipo return 1 RMSE

[0040] Furthermore, SMILES sequence text features are constructed. The SMILES string of each molecule is character-level encoded, and the characters in the SMILES string are mapped to the corresponding index numbers. The string is then truncated or padded into a fixed-length sequence to obtain a fixed-length SMILES sequence vector.

[0041] Then, the fixed-length sequence is input into the embedding layer to obtain character embedding representations, and local sequence patterns at different scales are extracted through one-dimensional convolutional layers with kernel sizes of 2, 4, and 8, respectively. Different convolutional scales can capture the sequence semantic information of short, medium-scale, and long segments, respectively. Subsequently, channel enhancement, non-linear activation, and max pooling are applied to the convolutional results at each scale. The resulting multi-scale representations are concatenated and mapped to a unified dimension through a fully connected layer to obtain the SMILES sequence modal feature representation. .

[0042] Furthermore, molecular fingerprint features are constructed. In this embodiment, an RDKit molecular object is first constructed based on the molecular SMILES string, and MACCS fingerprint, PubChem fingerprint, ErG fingerprint, Morgan fingerprint, and AtomPairs fingerprint are extracted respectively. The MACCS fingerprint is used to describe predefined chemical bonds, the PubChem fingerprint is used to characterize molecular structural fragments, the ErG fingerprint is used to characterize pharmacophore-related structural information, the Morgan fingerprint is used to describe the atomic neighborhood environment, and the AtomPairs fingerprint is used to represent the topological relationships between atomic pairs. These multiple molecular fingerprint vectors are concatenated to form a hybrid molecular fingerprint representation.

[0043]

[0044] In one specific embodiment, the MACCS fingerprint dimension is 166, the PubChem fingerprint dimension is 881, the ErG fingerprint dimension is 441, the Morgan fingerprint dimension is 1024, and the AtomPairs fingerprint dimension is 1024, resulting in a combined fingerprint dimension of 3536. The combined fingerprint is then input into a feedforward neural network, processed through a fully connected layer, a batch normalization layer, an activation function, and a random deactivation layer to obtain the molecular fingerprint modal feature representation. .

[0045] Furthermore, a two-dimensional molecular graph feature is constructed, representing each molecule as a two-dimensional graph structure: ,in, Represents a set of atomic nodes. The graph neural network represents a set of chemical bond edges. For each atomic node, node features are constructed, including atom type, atomicity, formal charge, number of unpaired electrons, hybridization mode, aromaticity, number of hydrogen atoms, and chirality information. For each chemical bond, edge features are constructed, including bond type, conjugation, whether it is located in a ring, and stereochemical information. In one embodiment, the graph neural network employs a multi-layer GraphConv structure. Each layer of graph convolution is followed by batch normalization, ReLU activation, and random deactivation operations. Global average pooling aggregates node-level features into a graph-level representation. Two-dimensional topological features are extracted through multi-layer graph convolution and graph-level pooling operations to obtain a two-dimensional molecular graph modal feature representation.

[0046] Furthermore, a three-dimensional conformation is generated and geometric features are extracted. To further characterize the spatial geometric information that is difficult to express in two-dimensional topological structures, this implementation generates a three-dimensional conformation based on molecular objects. Molecular objects are constructed based on the SMILES string, and hydrogen atoms are temporarily added to improve the stability of the three-dimensional conformation generation. The initial three-dimensional conformation of the molecule is generated using the ETKDGv3 method, and geometric optimization is performed using the MMFF94s force field. After optimization, only the three-dimensional coordinates of heavy atoms are retained, ensuring that the three-dimensional coordinate matrix maintains the same atomic node order as in the two-dimensional molecular diagram, thus obtaining the three-dimensional coordinate matrix.

[0047]

[0048] in, Indicates the first The number of heavy atoms in a molecule;

[0049] The edge distance characteristics are obtained by calculating the Euclidean distance between atoms at both ends of a chemical bond using a three-dimensional coordinate matrix:

[0050]

[0051] in, and These represent the three-dimensional coordinates of the atoms at both ends of a chemical bond;

[0052] Based on atomic adjacency relationships, bond angle indices and bond angle features are further constructed to describe the local spatial conformational constraints of molecules. In one embodiment, edge distance features, edge connectivity relationships, and node affiliation information are input into a 3D geometric encoder. First, the edge distances are nonlinearly transformed through an edge distance mapping network. Then, the edge features are aggregated to the target nodes. Finally, the 3D geometric modal feature representation is obtained through node mapping and global average pooling. .

[0053] Furthermore, the divergence-perceived uncertainty fusion step is to map the SMILES sequence modal feature representation, molecular fingerprint modal feature representation, two-dimensional molecular graph modal feature representation, and three-dimensional geometric modal feature representation to the same dimension and then perform hierarchical fusion.

[0054] The hierarchical fusion sequence is as follows: First, the molecular fingerprint mode and the two-dimensional molecular graph mode are fused to obtain the first-stage fusion representation; then, the first-stage fusion representation is fused with the three-dimensional geometric mode to obtain the second-stage fusion representation; finally, the second-stage fusion representation is fused with the SMILES sequence mode to obtain the final fused molecular characterization; the process is represented as follows:

[0055]

[0056]

[0057]

[0058] in, This represents a fusion function for discrepancy-perceived uncertainty. This represents the molecular characterization after final fusion.

[0059] For any two input modal features x and y, their log-variance is calculated using an uncertainty estimation network:

[0060]

[0061]

[0062] in, and This represents the uncertainty estimation network for the corresponding modality; the degree of divergence between two modal features is calculated using Euclidean distance:

[0063]

[0064] The uncertainty of the two modes is corrected by analyzing the divergence in Chengdu:

[0065]

[0066]

[0067] in, and The adjustable coefficients are learnable; the accuracy weights for the two modes are calculated based on the corrected uncertainty.

[0068]

[0069]

[0070] The fusion weights of the two modes are then obtained by normalization:

[0071]

[0072]

[0073] in, To prevent extremely small constants with a denominator of zero;

[0074] The final fusion is represented as:

[0075]

[0076] Therefore, the fusion weights are not fixed parameters, but are determined by the modal uncertainty and cross-modal divergence of the current sample. The fusion order is used to ensure the hierarchical structure of information integration, and the fusion weights are used to reflect the differences in the dependence of different samples on each modality in the final prediction.

[0077] Step 7: Output the molecular property prediction results.

[0078] Finally, the molecular representation input is fused into the linear prediction layer to obtain the molecular attribute prediction output. When the task is a classification task, the prediction layer outputs the class logits corresponding to each task, and the class probability is obtained through the Softmax function. The model is trained using the cross-entropy loss function, and the prediction performance is evaluated using ROC-AUC. When the task is a regression task, the prediction layer outputs the continuous attribute prediction values, and the model is trained using the mean squared error loss function. The prediction performance is evaluated using RMSE.

Claims

1. A multimodal molecular attribute prediction method based on hierarchical divergence-perceived uncertainty fusion, comprising the following steps: S1 Data Acquisition Steps: Acquire a molecular dataset containing molecular structure information and its corresponding attribute labels, which must include at least the string SMILES, and the attribute labels must include classification labels or regression numerical labels; S2 Sequence Text Feature Encoding Steps: Convert the SMILES string of each molecule into a fixed-length sequence index and input it into a multi-scale one-dimensional convolutional neural network to extract the sequence semantic feature representation of the molecule; S3 Molecular fingerprint feature construction steps: Extract multiple molecular fingerprint features based on the preprocessed molecular object, and concatenate different types of fingerprint vectors to form a molecular structure descriptor feature representation; S4 2D molecular graph construction steps: Represent each molecule as a 2D graph structure, where atoms are nodes and chemical bonds are edges. Construct node features, edge connection relationships, and edge features, and input them into a graph neural network to extract 2D topological structure feature representation; S5 3D conformation generation steps: Generate a 3D conformation based on the molecular structure, extract the 3D coordinates of heavy atoms, and obtain the 3D geometric feature representation of the molecule; S6 Disagreement-aware uncertainty fusion step: Based on the degree of disagreement between different modal features and the uncertainty of each modality itself, the different modal features are dynamically weighted and fused, and then the fused feature representation is generated step by step; S7 Molecular Attribute Prediction Steps: Input the final fused molecular characterization into the prediction module and output the molecular attribute prediction results. The evaluation index for classification tasks is ROC-AUC, and the evaluation index for regression tasks is RMSE.

2. The multimodal molecular attribute prediction method based on hierarchical divergence-perceived uncertainty fusion according to claim 1, characterized in that, The molecular dataset in S1 is represented as: wherein, denotes the SMILES string of the th molecule, denotes the attribute label corresponding to the molecule, and N denotes the number of molecular samples.

3. The multimodal molecular attribute prediction method based on hierarchical divergence-perceived uncertainty fusion according to claim 2, characterized in that, The sequence text feature encoding step described in S2 includes: character-level encoding of the SMILES string for each molecule, mapping the characters in the SMILES string to their corresponding index numbers, and truncating or padding them into a fixed-length sequence to obtain a fixed-length SMILES sequence vector. .

4. The multimodal molecular attribute prediction method based on hierarchical divergence-perceived uncertainty fusion according to claim 3, characterized in that, The molecular fingerprint feature construction steps described in S3 include: RDKit molecular objects are constructed based on the molecular SMILES string, and MACCS, PubChem, ErG, Morgan, and AtomPairs fingerprints are extracted respectively. The MACCS fingerprint describes predefined chemical bonds, the PubChem fingerprint characterizes molecular structural fragments, the ErG fingerprint delineates pharmacophore-related structural information, the Morgan fingerprint describes the atomic neighborhood environment, and the AtomPairs fingerprint represents the topological relationships between atomic pairs. These multiple molecular fingerprint vectors are concatenated to form a hybrid molecular fingerprint representation.

5. The multimodal molecular attribute prediction method based on hierarchical divergence-perceived uncertainty fusion according to claim 4, characterized in that, The two-dimensional molecular graph construction steps described in S4 include: representing each molecule as a two-dimensional graph structure: ,in, Represents a set of atomic nodes. The set of chemical bond edges is represented; for each atomic node, node features are constructed, including atom type, atomicity, formal charge, number of unpaired electrons, hybridization mode, aromaticity, number of hydrogen atoms, and chirality information. Two-dimensional topological features are extracted through multi-level graph convolution and graph-level pooling operations to obtain a two-dimensional molecular graph modal feature representation. .

6. The multimodal molecular attribute prediction method based on hierarchical divergence-perceived uncertainty fusion according to claim 5, characterized in that, The three-dimensional conformation generation steps described in S5 include: constructing a molecular object based on the SMILES string and temporarily adding hydrogen atoms to improve the stability of the three-dimensional conformation generation; generating the initial three-dimensional conformation of the molecule using the ETKDGv3 method and performing geometric optimization using the MMFF94s force field; after optimization, only the three-dimensional coordinates of heavy atoms are retained to ensure that the three-dimensional coordinate matrix is ​​consistent with the atomic node order in the two-dimensional molecular diagram, thus obtaining the three-dimensional coordinate matrix: in, Indicates the first The number of heavy atoms in a molecule; The edge distance characteristics are obtained by calculating the Euclidean distance between atoms at both ends of a chemical bond using a three-dimensional coordinate matrix: in, and These represent the three-dimensional coordinates of the atoms at both ends of a chemical bond; Based on atomic adjacency relationships, bond angle indices and bond angle features are further constructed to describe the local spatial conformational constraints of molecules. Edge distance features, edge connectivity relationships, and node affiliation information are input into a 3D geometric encoder, and 3D geometric modal feature representations are obtained through edge distance mapping, node aggregation, and graph-level pooling operations. .

7. The multimodal molecular attribute prediction method based on hierarchical divergence-perceived uncertainty fusion according to claim 6, characterized in that, The divergence-aware uncertainty fusion step described in S6 includes: mapping the SMILES sequence modal feature representation, molecular fingerprint modal feature representation, two-dimensional molecular graph modal feature representation, and three-dimensional geometric modal feature representation to the same dimension and then performing hierarchical fusion; The hierarchical fusion sequence is as follows: First, the molecular fingerprint mode and the two-dimensional molecular graph mode are fused to obtain the first-stage fusion representation; then, the first-stage fusion representation is fused with the three-dimensional geometric mode to obtain the second-stage fusion representation; finally, the second-stage fusion representation is fused with the SMILES sequence mode to obtain the final fused molecular characterization; the process is represented as follows: in, This represents a fusion function for discrepancy-perceived uncertainty. This represents the molecular characterization after final fusion. For any two input modal features x and y, their log-variance is calculated using an uncertainty estimation network: in, and This represents the uncertainty estimation network for the corresponding modality; the degree of divergence between two modal features is calculated using Euclidean distance: The uncertainty of the two modes is corrected by analyzing the divergence in Chengdu: in, and The adjustable coefficients are learnable; the accuracy weights for the two modes are calculated based on the corrected uncertainty. The fusion weights of the two modes are then obtained by normalization: in, To prevent extremely small constants with a denominator of zero; The final fusion is represented as: A multimodal molecular attribute prediction method based on hierarchical divergence perception uncertainty fusion according to claim 7, characterized in that, The molecular attribute prediction step described in S7 includes: inputting the final fused molecular representation into a linear prediction layer to obtain the molecular attribute prediction output. When the task is a classification task, the prediction layer outputs the category logits corresponding to each task, and obtains the category probability through the Softmax function. The model is trained using the cross-entropy loss function, and the prediction performance is evaluated using ROC-AUC. When the task is a regression task, the prediction layer outputs the continuous attribute prediction values, and the model is trained using the mean squared error loss function. The prediction performance is evaluated using RMSE.

Images