Molecular property prediction method and system based on multi-task pre-training and multi-modal fusion

Abstract

The invention relates to a molecular property prediction method and system based on multi-task pre-training and multi-modal fusion. The method comprises the following steps: collecting a data set and performing molecular conversion; performing multi-task pre-training, which comprises the following steps: generating a heterogeneous enhanced view; constructing a pseudo label; performing comparative learning according to the structure enhanced view and the heterogeneous enhanced view, and constructing a maximum similar task; capturing semantic differences among molecules according to the pseudo labels; carrying out multi-modal fusion, namely introducing functional group structure information, and extracting molecular sequence characteristics based on Transform and Mamba2 to obtain fused molecular multi-modal representation; and analyzing and predicting the classification or regression task. A heterogeneous enhanced view is established in a multi-task pre-training stage, multi-task self-supervision is performed in combination with multi-granularity features of molecular fingerprints, and functional group structure information is introduced in a multi-modal fusion stage, so that deep cross-modal interaction is realized, downstream prediction performance is improved, and candidate drug screening and molecular property evaluation processes are accelerated.

Description

Technical Field

[0001] This invention relates to the field of molecular property prediction technology, and in particular to a molecular property prediction method and system based on multi-task pre-training and multimodal fusion. Background Technology

[0002] Molecular property prediction is a crucial step in the modern drug development process, spanning multiple stages including lead compound screening, drug toxicity assessment, and preliminary determination of therapeutic activity. While wet laboratory assays offer high accuracy, their operation is typically costly and time-consuming, failing to meet the timeliness requirements of large-scale drug screening. Faced with the increasing complexity of drug design, relying on traditional experimental methods is no longer sufficient to support the efficiency demands of new drug development. Therefore, developing efficient and reliable computational methods for predicting molecular properties has become an important means of accelerating drug discovery. Existing molecular property prediction methods mainly include traditional machine learning methods based on molecular fingerprints, deep learning methods based on graph neural networks, modeling methods based on SMILES sequences, and multimodal fusion methods. Traditional methods typically rely on expert-constructed molecular fingerprints (such as ECFP and MACCS) as input, combined with models such as support vector machines and random forests for property prediction, offering good interpretability but limited expressive power. With the development of deep learning, graph neural networks (such as GCN, GAT, and MPNN) are widely used to model the topological structure of molecules, automatically learning the representation of atoms and their neighborhoods through message passing mechanisms between nodes, significantly improving prediction accuracy. Meanwhile, models based on SMILES sequences represent molecules as strings and utilize RNN, CNN, or Transformer structures to extract semantic features at the sequence level, offering greater modeling flexibility. In recent years, multimodal methods have further enhanced molecular characterization capabilities by fusing graph structures, sequences, fingerprints, and other representations.

[0003] However, current mainstream self-supervised learning methods mostly rely on constructing augmented views to learn robust representations, but these augmentation strategies often disrupt the original semantic environment of molecules, causing the learned representations to deviate from the true chemical space. Simultaneously, existing methods typically ignore prior chemical knowledge in molecular structures, such as important information like interatomic interactions and functional groups, limiting the chemical rationality and interpretability of the models. Contrastive learning, as a commonly used self-supervised framework, relies excessively on the selection and quantity of negative samples, making it susceptible to the influence of sampling strategies and affecting model performance and stability. Furthermore, single-modal representations (such as using only molecular diagrams or SMILES sequences) cannot comprehensively capture the multi-level structural and semantic information of molecules, limiting the expressive power of the models.

[0004] Therefore, traditional molecular property prediction methods often suffer from low prediction efficiency and lengthy evaluation processes because the original semantic environment of molecules is destroyed, prior chemical knowledge in molecular structure is ignored, and the selection and quantity of negative samples are over-reliant. Summary of the Invention

[0005] In order to solve the above-mentioned technical problems, a molecular property prediction method and system based on multi-task pre-training and multi-modal fusion is provided, which can realize deep cross-modal interaction and improve downstream prediction performance, and accelerate the process of candidate drug screening and molecular property evaluation.

[0006] A molecular property prediction method based on multi-task pre-training and multimodal fusion, the method comprising:

[0007] Collect a dataset containing molecular structure and property labels, and convert the molecules in the dataset into SMILES molecular sequences, molecular maps, and molecular fingerprints;

[0008] Multi-task pre-training of the model includes: constructing a structure-enhanced view based on the molecular graph; constructing an atomic-level molecular graph and a structural fragment graph based on the SMILES molecular sequence, and generating a heterogeneous enhanced view connecting atomic nodes and fragment nodes; and constructing pseudo-labels based on the molecular fingerprint.

[0009] Based on the comparative learning between the structure-enhanced view and the heterogeneous-enhanced view, a maximum similarity task is constructed between the structure-enhanced view and the heterogeneous-enhanced view; and semantic differences between molecules are captured based on the pseudo-labels; thus completing multi-task pre-training.

[0010] Multimodal fusion includes: introducing functional group structure information extracted based on prior rules, integrating the structure information obtained in multi-task pre-training with the functional group structure information, and extracting molecular sequence features based on Transformer and Mamba2 to obtain the fused molecular multimodal representation;

[0011] Based on the fused molecular multimodal representation, analytical prediction results are obtained for classification or regression tasks.

[0012] In one embodiment, an atomic-level molecular map and a structural fragment map are constructed based on the SMILES molecular sequence, generating a heterogeneous enhanced view connecting atomic nodes and fragment nodes, including:

[0013] Based on the SMILES molecular sequence, the BRICS algorithm is used to fragment the molecule and construct a structural fragment map in units of fragments;

[0014] The RDKit tool was used to extract atomic and bond features from the molecular diagram to construct an atomic-level molecular diagram.

[0015] By combining the atomic-level molecular diagram with the structural fragment diagram, the connection edges between fragments and the heterogeneous edges connecting atoms and fragments are determined, generating a heterogeneous enhanced view.

[0016] In one embodiment, constructing a pseudo-tag based on the molecular fingerprint includes:

[0017] The molecular fingerprint is clustered using a multi-granularity clustering algorithm to generate pseudo-labels;

[0018] A classification task based on the pseudo-labels is constructed to capture the semantic differences between different molecules.

[0019] In one embodiment, the task of constructing the maximum similarity between the structurally enhanced view and the heterogeneous enhanced view through comparative learning includes:

[0020] Based on the structural enhancement view and the heterogeneous enhancement view, positive and negative sample pairs are constructed;

[0021] Based on the positive and negative sample pairs, maximize the similarity between the structure-enhanced view and the heterogeneous-enhanced view in the representation space;

[0022] The Euclidean distance is used to measure the representational similarity between the structurally enhanced view and the heterogeneous enhanced view, and a maximum similarity task is constructed.

[0023] In one embodiment, the method further includes:

[0024] Based on the pseudo-labels, determine the label matching task and obtain the cross-entropy loss and clustering labels for multi-label classification;

[0025] Based on the maximum similarity task, cross-entropy loss, and clustering labels, the minimum joint loss in the multi-task pre-training stage is calculated to complete the multi-task pre-training.

[0026] In one embodiment, the method further includes:

[0027] Functional group structure information extracted based on prior rules is introduced, and an identifier is added before the functional group structure information and the structure information obtained in multi-task pre-training to form target information;

[0028] The target information is modeled in context using a self-attention mechanism to obtain a global feature representation.

[0029] In one embodiment, the structural information obtained in multi-task pre-training is integrated with the functional group structural information, and molecular sequence features are extracted based on Transformer and Mamba2 to obtain a fused molecular multimodal representation, including:

[0030] The SMILES molecular sequence is used as input, and the final sequence features are extracted using Transformer and Mamba2.

[0031] The integrated structural information and functional group structural information are used as functional group features, and the functional group features are combined and fused with the final sequence features to obtain a fused molecular multimodal representation.

[0032] In one embodiment, binary cross-entropy loss is used as the loss function in a classification task that performs analysis and prediction.

[0033] In regression tasks that perform analysis and prediction, mean squared error is used as the loss function.

[0034] A molecular property prediction system based on multi-task pre-training and multimodal fusion, the system comprising:

[0035] The data collection module is used to collect datasets containing molecular structure and property labels, and convert the molecules in the dataset into SMILES molecular sequences, molecular maps, and molecular fingerprints;

[0036] A multi-task pre-training module is used for multi-task pre-training of the model, including: constructing a structure-enhanced view based on the molecular graph; constructing an atomic-level molecular graph and a structural fragment graph based on the SMILES molecular sequence, and generating a heterogeneous enhanced view connecting atomic nodes and fragment nodes; and constructing pseudo-labels based on the molecular fingerprint.

[0037] The self-supervised task design module is used to construct the maximum similarity task between the structure-enhanced view and the heterogeneous-enhanced view by comparing and learning the structure-enhanced view and the heterogeneous-enhanced view; and to capture the semantic differences between molecules based on the pseudo-labels; and to complete multi-task pre-training.

[0038] A multimodal fusion module is used to perform multimodal fusion, including: introducing functional group structure information extracted based on prior rules, integrating the structure information obtained in multi-task pre-training with the functional group structure information, and extracting molecular sequence features based on Transformer and Mamba2 to obtain a fused molecular multimodal representation;

[0039] The analysis and prediction module is used to obtain analysis and prediction results for classification or regression tasks based on the fused molecular multimodal representation.

[0040] In one embodiment, the multi-task pre-training module is further configured to fragment the molecule based on the SMILES molecular sequence using the BRICS algorithm to construct a structural fragment map in units of fragments; extract the atomic and bond features in the molecular map using the RDKit tool to construct an atomic-level molecular map; and combine the atomic-level molecular map with the structural fragment map to determine the connection edges between fragments and the heterogeneous edges connecting atoms and fragments to generate a heterogeneous enhanced view.

[0041] The aforementioned molecular property prediction method and system based on multi-task pre-training and multimodal fusion constructs a heterogeneous enhanced view during the multi-task pre-training stage and performs multi-task self-supervision by combining multi-granular features of molecular fingerprints to effectively train the graph encoder to capture molecular structural information with higher semantic fidelity. In the multimodal fusion stage, functional group structural information is introduced, and molecular sequence features are extracted by combining Transformer and Mamba2 to complete effective fusion, thereby achieving deep cross-modal interaction and improving downstream prediction performance, thus accelerating the process of candidate drug screening and molecular property evaluation. Attached Figure Description

[0042] Figure 1 This is an application environment diagram of a molecular property prediction method based on multi-task pre-training and multi-modal fusion in one embodiment;

[0043] Figure 2 This is a flowchart illustrating a molecular property prediction method based on multi-task pre-training and multi-modal fusion in one embodiment.

[0044] Figure 3 This is a schematic diagram of the framework structure for multi-task pre-training and multi-modal fusion in one embodiment;

[0045] Figure 4 This is a block diagram of a molecular property prediction system based on multi-task pre-training and multi-modal fusion in one embodiment;

[0046] Figure 5 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0048] The molecular property prediction method based on multi-task pre-training and multi-modal fusion provided in this application embodiment can be applied to, for example... Figure 1 The application environment shown. For example... Figure 1As shown, the application environment includes computer equipment 110. Computer device 110 can collect datasets containing molecular structure and property labels, and convert the molecules in the dataset into SMILES molecular sequences, molecular graphs, and molecular fingerprints. Computer device 110 can perform multi-task pre-training of the model, including: constructing structure-enhanced views based on molecular graphs; constructing atomic-level molecular graphs and structural fragment graphs based on SMILES molecular sequences, generating heterogeneous enhanced views connecting atomic nodes and fragment nodes; constructing pseudo-labels based on molecular fingerprints; computer device 110 can perform comparative learning between structure-enhanced views and heterogeneous enhanced views to construct a maximum similarity task between the structure-enhanced views and heterogeneous enhanced views; and capture semantic differences between molecules based on pseudo-labels; completing multi-task pre-training; computer device 110 can perform multimodal fusion, including: introducing functional group structure information extracted based on prior rules, integrating the structure information obtained in multi-task pre-training with the functional group structure information, and extracting molecular sequence features based on Transformer and Mamba2 to obtain a fused molecular multimodal representation; computer device 110 can obtain analysis and prediction results based on the fused molecular multimodal representation for classification or regression tasks. Among them, computer equipment 110 may include, but is not limited to, various personal computers, laptops, smartphones, robots, tablets and other devices.

[0049] In one embodiment, such as Figure 2 As shown, a molecular property prediction method based on multi-task pre-training and multi-modal fusion is provided, including the following steps:

[0050] Step 202: Collect a dataset containing molecular structure and property labels, and convert the molecules in the dataset into SMILES molecular sequences, molecular maps, and molecular fingerprints.

[0051] Computer equipment can first perform data preparation work, which includes: collecting public datasets containing molecular structure and property tags (such as bace, esol, tox21, etc.), and converting molecules into SMILES sequences, molecular maps, and molecular fingerprints (MACSS).

[0052] Step 204 involves multi-task pre-training of the model, including: constructing a structure-enhanced view based on the molecular graph; constructing an atomic-level molecular graph and structural fragment graph based on the SMILES molecular sequence, generating a heterogeneous enhanced view connecting atomic nodes and fragment nodes; and constructing pseudo-labels based on molecular fingerprints.

[0053] During the pre-training phase, the computer device can construct structurally enhanced views centered on molecular graphs and mine potential semantic features through multi-task learning strategies to achieve structure awareness and semantic alignment. The computer device can construct paired sample pairs based on fragment enhancement and design two types of self-supervised tasks: one is to maximize the representational similarity between enhanced views to improve the model's robustness to structural changes; the other is to construct pseudo-labels based on molecular fingerprints to guide the model in label matching tasks, thereby improving the discriminative ability of the graph encoder under unsupervised conditions.

[0054] In one embodiment, a molecular property prediction method based on multi-task pre-training and multimodal fusion may further include a process of generating a heterogeneous enhanced view. The specific process includes: based on the SMILES molecular sequence, using the BRICS algorithm to fragment the molecule and construct a structural fragment map in units of fragments; using the RDKit tool to extract atomic and bond features from the molecular map and construct an atomic-level molecular map; combining the atomic-level molecular map with the structural fragment map to determine the connection edges between fragments and the heterogeneous edges connecting atoms and fragments, thereby generating a heterogeneous enhanced view.

[0055] Computer devices can utilize cheminformatics tools (such as RDKit) to extract atomic and bond features from molecular graphs. Simultaneously, the BRICS algorithm is used to fragment the molecule, constructing a fragment-based view, which is then combined with the atomic-level molecular graph to generate a heterogeneous enhanced view containing different semantic levels. Specifically, based on the input SMILES molecular sequence, the computer device can use RDKit to construct both an atomic-level molecular graph and a structural fragment graph obtained from BRICS rule segmentation, and further generate a heterogeneous graph connecting atomic nodes and fragment nodes, thereby achieving the construction of a semantically preserved heterogeneous enhanced view.

[0056] In this embodiment, the computer device uses the BRICS algorithm to fragment decompose the original molecular graph, generating a heterogeneous enhancement graph containing atomic nodes and fragment nodes. The structural enhancement view can be defined as G = (V... atom E atom The heterogeneous enhanced view is G. het =(V atom ∪V frag E atom ∪E frag ∪E inter ), where V frag E represents the structural fragment nodes obtained based on BRICS partitioning. frag E represents the connecting edges between segments. inter This connects atoms to their respective fragments via heterogeneous edges. The enhanced view preserves the structural semantic integrity, avoiding the problem of traditional random deletion and replacement strategies violating chemical priors.

[0057] Step 206: Construct the maximum similarity task between the structure-enhanced view and the heterogeneous-enhanced view by comparing and learning the structure-enhanced view and the heterogeneous-enhanced view; and capture the semantic differences between molecules based on pseudo-labels; complete multi-task pre-training.

[0058] In this embodiment, a dual-task joint self-supervised learning mechanism is designed in the pre-training stage: First, by comparing the structural augmented view and the heterogeneous augmented view, the model is guided to capture the semantic consistency of the graph structure and obtain a robust shared representation; Second, pseudo-labels are generated based on molecular fingerprint clustering to construct an auxiliary classification task, which enhances the discrimination ability between molecules and alleviates the problem of excessive dependence on negative samples in the comparative learning.

[0059] In one embodiment, a molecular property prediction method based on multi-task pre-training and multimodal fusion may further include a contrastive learning process, specifically including: constructing positive and negative sample pairs based on structural augmented views and heterogeneous augmented views; maximizing the similarity between structural augmented views and heterogeneous augmented views in the representation space according to the positive and negative sample pairs; and constructing a maximum similarity task by using Euclidean distance to measure the representational similarity between structural augmented views and heterogeneous augmented views.

[0060] Computer devices can enable encoders to learn segments by maximizing the similarity between structurally augmented views and heterogeneously augmented views. Specifically, Euclidean distance can be used to measure the representational similarity between the original and augmented images, and maximizing similarity is achieved by minimizing the distance. The calculation formula can be expressed as: Where f(*) represents the original graph encoder CMPNN, f ' (*) indicates the heterogeneous enhanced graph encoder PharmHGT; G n This represents an enhanced structural view. This represents a heterogeneous augmented view; N represents the number of samples in the pre-trained dataset.

[0061] In one embodiment, a molecular property prediction method based on multi-task pre-training and multimodal fusion may further include a process of constructing pseudo-labels. The specific process includes: clustering molecular fingerprints using a multi-granularity clustering algorithm to generate pseudo-labels; constructing a classification task based on pseudo-labels; and capturing semantic differences between different molecules according to the classification task.

[0062] Computer devices can be used to construct classification tasks based on clustered fingerprint labels, enabling models to capture semantic differences between different molecules. Specifically, multi-granularity clustering algorithms can be used to cluster molecular MACCS fingerprints to generate pseudo-labels, training the encoder to enhance its semantic consistency. The calculation formula for multi-granularity clustering algorithms can be expressed as: Where l1, l2, and l3 are the cross-entropy losses for multi-label classification. and These are K-means clustering labels based on molecular MACSS fingerprints using K values ​​of 100, 1000, and 10000; W represents the classification parameters, W... 100 W 1000 W 10000 These correspond to layers of 100, 1000, and 10000 neurons, respectively.

[0063] In one embodiment, a molecular property prediction method based on multi-task pre-training and multimodal fusion may further include a process of calculating a loss function. The specific process includes: determining the label matching task based on pseudo-labels and obtaining the cross-entropy loss and clustering labels for multi-label classification; calculating the minimum joint loss in the multi-task pre-training stage based on the maximum similarity task, cross-entropy loss, and clustering labels to complete the multi-task pre-training.

[0064] Ultimately, the goal of the pre-training phase is to minimize the joint loss: L pre =argmin(L contrast +L cluster ).

[0065] By combining pseudo-supervised matching tasks based on molecular fingerprint clustering labels and tasks that enhance the similarity maximization between views, the molecular representation capabilities of graph encoders are improved from two dimensions: label semantics and structure preservation. This effectively alleviates problems such as strong dependence on negative samples and semantic destruction in contrastive learning.

[0066] Step 208 involves multimodal fusion, including: introducing functional group structure information extracted based on prior rules, integrating the structure information obtained in multi-task pre-training with the functional group structure information, and extracting molecular sequence features based on Transformer and Mamba2 to obtain the fused molecular multimodal representation.

[0067] The computer device can inject the pre-trained graph features into the initial representation of the functional group to guide it to learn structure-aware features; then, the global representation of the functional group is extracted using the Transformer; for the SMILES sequence, the multi-scale semantic features are extracted using both the Transformer and the Mamba2 structure; finally, the functional group features are concatenated with the sequence features to form a fused molecular multimodal representation.

[0068] In one embodiment, a molecular property prediction method based on multi-task pre-training and multimodal fusion may further include an information integration process, specifically including: taking the SMILES molecular sequence as input, extracting the final sequence features through Transformer and Mamba2; taking the integrated structural information and functional group structural information as functional group features, and performing feature concatenation and fusion of the functional group features and the final sequence features to obtain a fused molecular multimodal representation.

[0069] To address the crucial role of functional groups in determining molecular properties, this embodiment introduces functional group structure information extracted based on prior rules during the fine-tuning stage. Let the set of functional groups contained in the molecule be F = {fg1, fg2, ...}, and the corresponding embedding representation is... Since functional groups are also structural information, and pre-training also yields structural information, they are integrated with the pre-trained structural information in the early stages of functional group learning. This integration is done on a unit basis, with each functional group containing atoms, to form an initial representation: M max (V i = maxpooling(v f(G) ,v∈V i ); where V i V represents the set of atoms contained in the i-th functional group; W represents the trainable weight matrix, v f(G) It is the atomic representation after being encoded by a pre-trained encoder.

[0070] In one embodiment, a molecular property prediction method based on multi-task pre-training and multi-modal fusion may further include a process using a self-attention mechanism. The specific process includes: introducing functional group structure information extracted based on prior rules, and adding a notation before the functional group structure information and the structure information obtained in multi-task pre-training to form target information; and performing contextual modeling on the target information through a self-attention mechanism to obtain a global feature representation.

[0071] Add a special symbol before all functional groups constitute We use a self-attention mechanism to model its context and obtain a global representation: Thus, functional group vectors representing the graph structure are extracted.

[0072] In this embodiment, when the computer device performs sequence feature extraction, the SMILES sequence embedding is specifically represented as S. em The final sequence features, H, are obtained by sequentially inputting a Transformer with local encoding capabilities and a Mamba2 algorithm that captures long-range dependencies. S=Mamba2(Transformer(S em )).

[0073] In this embodiment, when the computer device performs feature fusion and prediction, functional group features are integrated with graph features in the early stage of feature extraction, and then integrated with sequence features in the later stage of feature extraction to obtain the final molecular features for downstream task prediction.

[0074] Multimodal fusion during model fine-tuning, incorporating functional group and sequence information, can enhance the model's interpretability and predictive ability. Specifically, during fine-tuning, the computing device can incorporate rule-based functional group structural information and use the graph features obtained in the pre-training stage as its initial embedding. Multi-layer self-attention modules are used to perform contextual modeling of functional groups, capturing their global semantic features to improve the model's structural interpretability. Simultaneously, the SMILES sequences are processed using Transformer and Mamba2 to extract local and long-range dependency features, forming semantically rich sequence representations. Finally, the fused functional group and sequence representations are input into the downstream prediction network to achieve molecular property classification or regression tasks.

[0075] By injecting the pre-trained molecular graph features into the initial representation of functional groups, and using Transformer to extract global dependency features between functional groups, the semantic expressive power is enhanced. At the same time, by combining the SMILES sequence input, multi-scale semantic features are extracted by Transformer and Mamba2 respectively, and fused with functional group features, the complementary advantages of multimodal information can be fully utilized.

[0076] Step 210: Based on the fused molecular multimodal representation, obtain the analysis and prediction results for the classification or regression task.

[0077] Computer devices can input the fused molecular multimodal representation into a fully connected prediction module, and output corresponding molecular prediction results for specific classification or regression tasks, thereby achieving accurate prediction of molecular properties such as toxicity, solubility, and biological activity.

[0078] In one embodiment, binary cross-entropy loss is used as the loss function in classification tasks that perform analysis and prediction; mean squared error is used as the loss function in regression tasks that perform analysis and prediction.

[0079] Specifically, for classification tasks, binary cross-entropy (BCE) loss is used as the loss function.

[0080] For regression tasks, mean squared error (MSE) is used as the loss function.

[0081] In terms of model evaluation, the area under the receiver operating characteristic curve (ROC-AUC) is used as the main performance indicator for classification tasks, while the root mean squared error (RMSE) is used to measure prediction accuracy for regression tasks.

[0082] In one embodiment, such as Figure 3 As shown, a molecular property prediction method based on multi-task pre-training and multimodal fusion can be provided, including a multi-task pre-training part A and a multimodal fusion, i.e., fine-tuning part B. In the multi-task pre-training part A, molecular transformation can be performed from a dataset without property labels to obtain transformed SMILES molecular sequences, molecular maps, and molecular fingerprints; based on the molecular map, an enhanced structure view can be constructed, i.e. Figure 3 The homomorphic graph convolution process in [the context of the graph]; based on the SMILES molecular sequence, a heterogeneous enhanced view can be constructed, i.e. Figure 3 The heterogeneous graph convolution process is performed in the middle; then, comparative learning can be carried out between the structure-enhanced view and the heterogeneous enhanced view to construct the maximum similarity task between the structure-enhanced view and the heterogeneous enhanced view; at the same time, pseudo-labels can be constructed based on molecular fingerprints to capture semantic differences between molecules and complete multi-task pre-training.

[0083] Next, in the fine-tuning part B, functional group structure information can be extracted from the dataset containing molecular structure and property labels based on prior rules as an embedding; molecular sequence features in SMILES molecular sequences are extracted using Transformer and Mamba2, and the structural information obtained in multi-task pre-training is integrated with the functional group structure information to obtain a fused molecular multimodal representation, which further completes the prediction of molecular properties.

[0084] In this application, by introducing a multi-task self-supervised pre-training task based on molecular fingerprint label matching and enhanced view similarity alignment, the model significantly improves the molecular graph encoder's ability to represent structural semantics. Multi-task self-supervised pre-training and multimodal fine-tuning are combined for molecular property prediction. During pre-training, heterogeneous enhanced views are constructed based on molecular fragments to preserve the original molecular semantics, enabling the graph encoder to capture fragment-level information. Simultaneously, fingerprint features are integrated into the multi-task learning objective, reducing dependence on negative sampling and enhancing the encoder's representational ability. In the fine-tuning stage, pre-trained graph features are injected into the initial representation of functional groups and combined with a Transformer to extract their global dependency features. Simultaneously, the Transformer and Mamba2 structure are used to extract multi-scale semantic information of the SMILES sequence, achieving effective fusion of graph structure, functional groups, and sequence features. Pre-trained graph features are injected into the initial representation of hanging functional groups to guide feature extraction, and then fused with SMILES features to achieve deep cross-modal interaction and improve downstream prediction performance, enabling efficient and accurate prediction of molecular properties. Furthermore, through structure-guided pre-training strategies and multimodal fusion design, this approach enhances the model's sensitivity to key functional groups and molecular structural features while maintaining prediction accuracy. It also possesses stronger generalization ability and chemical interpretability, which helps to accelerate the screening of candidate drugs and the evaluation of molecular properties.

[0085] It should be understood that although the steps in the above flowcharts are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the above flowcharts may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0086] In one embodiment, such as Figure 4 As shown, a molecular property prediction system based on multi-task pre-training and multimodal fusion is provided, including: a data collection module 410, a multi-task pre-training module 420, a self-supervised task design module 430, a multimodal fusion module 440, and an analysis and prediction module 450, wherein:

[0087] The data collection module 410 is used to collect a dataset containing molecular structure and property labels, and to convert the molecules in the dataset into SMILES molecular sequences, molecular maps, and molecular fingerprints.

[0088] The multi-task pre-training module 420 is used for multi-task pre-training of the model, including: constructing structure-enhanced views based on molecular graphs; constructing atomic-level molecular graphs and structural fragment graphs based on SMILES molecular sequences, generating heterogeneous enhanced views connecting atomic nodes and fragment nodes; and constructing pseudo-labels based on molecular fingerprints.

[0089] The self-supervised task design module 430 is used to construct the maximum similarity task between the structure-enhanced view and the heterogeneous-enhanced view by comparing and learning from the structure-enhanced view and the heterogeneous-enhanced view; and to capture the semantic differences between molecules based on pseudo-labels; and to complete multi-task pre-training.

[0090] The multimodal fusion module 440 is used to perform multimodal fusion, including: introducing functional group structure information extracted based on prior rules, integrating the structure information obtained in multi-task pre-training with the functional group structure information, and extracting molecular sequence features based on Transformer and Mamba2 to obtain the fused molecular multimodal representation;

[0091] The analysis and prediction module 450 is used for fusion-based molecular multimodal representation to obtain analysis and prediction results for classification or regression tasks.

[0092] In one embodiment, the multi-task pre-training module 420 is further used to fragment the molecule based on the SMILES molecular sequence using the BRICS algorithm to construct a structural fragment map in units of fragments; to extract the atomic and bond features in the molecular map using the RDKit tool to construct an atomic-level molecular map; and to combine the atomic-level molecular map with the structural fragment map to determine the connection edges between fragments and the heterogeneous edges connecting atoms and fragments to generate a heterogeneous enhanced view.

[0093] In one embodiment, the multi-task pre-training module 420 is also used to cluster molecular fingerprints using a multi-granularity clustering algorithm to generate pseudo-labels; construct a classification task based on pseudo-labels, and capture the semantic differences between different molecules according to the classification task.

[0094] In one embodiment, the self-supervised task design module 430 is further configured to construct positive and negative sample pairs based on the structural augmented view and the heterogeneous augmented view; maximize the similarity between the structural augmented view and the heterogeneous augmented view in the representation space according to the positive and negative sample pairs; and use Euclidean distance to measure the representation similarity between the structural augmented view and the heterogeneous augmented view to construct the maximum similarity task.

[0095] In one embodiment, the self-supervised task design module 430 is further configured to determine the label matching task based on the pseudo-labels, and obtain the cross-entropy loss and clustering labels for multi-label classification; and calculate the minimum joint loss for the multi-task pre-training stage based on the maximum similarity task, cross-entropy loss, and clustering labels, thereby completing the multi-task pre-training.

[0096] In one embodiment, the multimodal fusion module 440 is further used to introduce functional group structure information extracted based on prior rules, and add an identifier before the functional group structure information and the structure information obtained in multi-task pre-training to form target information; and perform context modeling on the target information through a self-attention mechanism to obtain a global feature representation.

[0097] In one embodiment, the multimodal fusion module 440 is further configured to take the SMILES molecular sequence as input, extract the final sequence features through Transformer and Mamba2, take the integrated structural information and functional group structural information as functional group features, and perform feature splicing and fusion with the functional group features and the final sequence features to obtain the fused molecular multimodal representation.

[0098] In one embodiment, binary cross-entropy loss is used as the loss function in classification tasks that perform analysis and prediction; mean squared error is used as the loss function in regression tasks that perform analysis and prediction.

[0099] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 5 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When executed by the processor, the computer program implements a molecular property prediction method based on multi-task pre-training and multimodal fusion. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0100] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0101] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement steps of a molecular property prediction method based on multi-task pre-training and multimodal fusion.

[0102] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of a molecular property prediction method based on multi-task pre-training and multimodal fusion.

[0103] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0104] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0105] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for predicting molecular properties based on multi-task pre-training and multimodal fusion, characterized in that, The method includes: Collect a dataset containing molecular structure and property labels, and convert the molecules in the dataset into SMILES molecular sequences, molecular maps, and molecular fingerprints; Multi-task pre-training of the model includes: constructing a structure-enhanced view based on the molecular graph; constructing an atomic-level molecular graph and a structural fragment graph based on the SMILES molecular sequence, and generating a heterogeneous enhanced view connecting atomic nodes and fragment nodes; and constructing pseudo-labels based on the molecular fingerprint. Based on the comparative learning between the structure-enhanced view and the heterogeneous-enhanced view, a maximum similarity task is constructed between the structure-enhanced view and the heterogeneous-enhanced view; and semantic differences between molecules are captured based on the pseudo-labels; thus completing multi-task pre-training. Multimodal fusion includes: introducing functional group structure information extracted based on prior rules, integrating the structure information obtained in multi-task pre-training with the functional group structure information, and extracting molecular sequence features based on Transformer and Mamba2 to obtain the fused molecular multimodal representation; Based on the fused molecular multimodal representation, analytical prediction results are obtained for classification or regression tasks.

2. The molecular property prediction method based on multi-task pre-training and multi-modal fusion according to claim 1, characterized in that, Based on the SMILES molecular sequence, an atomic-level molecular map and structural fragment map are constructed, generating a heterogeneous enhanced view connecting atomic nodes and fragment nodes, including: Based on the SMILES molecular sequence, the BRICS algorithm is used to fragment the molecule and construct a structural fragment map in units of fragments; The RDKit tool was used to extract atomic and bond features from the molecular diagram to construct an atomic-level molecular diagram. By combining the atomic-level molecular diagram with the structural fragment diagram, the connection edges between fragments and the heterogeneous edges connecting atoms and fragments are determined, generating a heterogeneous enhanced view.

3. The molecular property prediction method based on multi-task pre-training and multi-modal fusion according to claim 1, characterized in that, Constructing pseudo-tags based on the molecular fingerprint includes: The molecular fingerprint is clustered using a multi-granularity clustering algorithm to generate pseudo-labels; A classification task based on the pseudo-labels is constructed to capture the semantic differences between different molecules.

4. The molecular property prediction method based on multi-task pre-training and multi-modal fusion according to claim 1, characterized in that, Based on comparative learning between the structurally enhanced view and the heterogeneous enhanced view, the task of constructing the maximum similarity between the structurally enhanced view and the heterogeneous enhanced view includes: Based on the structural enhancement view and the heterogeneous enhancement view, positive and negative sample pairs are constructed; Based on the positive and negative sample pairs, maximize the similarity between the structure-enhanced view and the heterogeneous-enhanced view in the representation space; The Euclidean distance is used to measure the representational similarity between the structurally enhanced view and the heterogeneous enhanced view, and a maximum similarity task is constructed.

5. The molecular property prediction method based on multi-task pre-training and multi-modal fusion according to claim 1, characterized in that, The method further includes: Based on the pseudo-labels, determine the label matching task and obtain the cross-entropy loss and clustering labels for multi-label classification; Based on the maximum similarity task, cross-entropy loss, and clustering labels, the minimum joint loss in the multi-task pre-training stage is calculated to complete the multi-task pre-training.

6. The molecular property prediction method based on multi-task pre-training and multi-modal fusion according to claim 1, characterized in that, The method further includes: Functional group structure information extracted based on prior rules is introduced, and an identifier is added before the functional group structure information and the structure information obtained in multi-task pre-training to form target information; The target information is modeled in context using a self-attention mechanism to obtain a global feature representation.

7. The molecular property prediction method based on multi-task pre-training and multi-modal fusion according to claim 6, characterized in that, The structural information obtained from multi-task pre-training is integrated with the functional group structural information, and molecular sequence features are extracted based on Transformer and Mamba2 to obtain a fused molecular multimodal representation, including: The SMILES molecular sequence is used as input, and the final sequence features are extracted using Transformer and Mamba2. The integrated structural information and functional group structural information are used as functional group features, and the functional group features are combined and fused with the final sequence features to obtain a fused molecular multimodal representation.

8. The molecular property prediction method based on multi-task pre-training and multi-modal fusion according to claim 1, characterized in that, In classification tasks that perform analysis and prediction, binary cross-entropy loss is used as the loss function; In regression tasks that perform analysis and prediction, mean squared error is used as the loss function.

9. A molecular property prediction system based on multi-task pre-training and multimodal fusion, characterized in that, The system includes: The data collection module is used to collect datasets containing molecular structure and property labels, and convert the molecules in the dataset into SMILES molecular sequences, molecular maps, and molecular fingerprints; A multi-task pre-training module is used for multi-task pre-training of the model, including: constructing a structure-enhanced view based on the molecular graph; constructing an atomic-level molecular graph and a structural fragment graph based on the SMILES molecular sequence, and generating a heterogeneous enhanced view connecting atomic nodes and fragment nodes; and constructing pseudo-labels based on the molecular fingerprint. The self-supervised task design module is used to construct the maximum similarity task between the structure-enhanced view and the heterogeneous-enhanced view by comparing and learning the structure-enhanced view and the heterogeneous-enhanced view; and to capture the semantic differences between molecules based on the pseudo-labels; and to complete multi-task pre-training. A multimodal fusion module is used to perform multimodal fusion, including: introducing functional group structure information extracted based on prior rules, integrating the structure information obtained in multi-task pre-training with the functional group structure information, and extracting molecular sequence features based on Transformer and Mamba2 to obtain a fused molecular multimodal representation; The analysis and prediction module is used to obtain analysis and prediction results for classification or regression tasks based on the fused molecular multimodal representation.

10. The molecular property prediction system based on multi-task pre-training and multi-modal fusion according to claim 9, characterized in that, The multi-task pre-training module is also used to fragment the molecule based on the SMILES molecular sequence using the BRICS algorithm to construct a structural fragment map in units of fragments; to extract the atomic and bond features in the molecular map using the RDKit tool to construct an atomic-level molecular map; and to combine the atomic-level molecular map with the structural fragment map to determine the connection edges between fragments and the heterogeneous edges connecting atoms and fragments to generate a heterogeneous enhanced view.

Images