Molecular property prediction method and apparatus using adaptive multi-modal gated hybrid expert network

Abstract

This invention relates to a method and device for predicting molecular properties using an adaptive multimodal gated hybrid expert network. First, compound data with molecular structure and property annotations is acquired to construct multimodal molecular features. Then, high-level semantic information from fingerprints, descriptors, and molecular graph features is extracted using sub-expert models. A gated network is introduced to dynamically learn and normalize the weights of each modality, achieving adaptive weighted fusion of multimodal features. Finally, molecular properties are predicted based on the fused features. Compared with existing technologies, this invention overcomes the limitations of traditional single-modality or fixed-weight fusion, dynamically adjusting modal contributions according to molecular structure characteristics, significantly improving the accuracy and robustness of molecular property prediction.

Description

Technical Field

[0001] This invention relates to the field of molecular property prediction technology, and in particular to a method and device for molecular property prediction using an adaptive multimodal gated hybrid expert network. Background Technology

[0002] With the continuous development of artificial intelligence and molecular modeling techniques in drug development, molecular property prediction has become a crucial technical step in improving drug development efficiency, optimizing candidate molecule screening processes, and reducing clinical development risks. The various intrinsic properties of drug molecules directly affect their structure-function relationship, in vivo behavior, and potential safety, thus significantly impacting drug efficacy, reliability, and the success rate of clinical development.

[0003] In drug development, the various intrinsic properties of drug molecules have a decisive influence on their in vivo behavior and efficacy. Different molecular properties work together to affect drug absorption, distribution, efficacy, and potential safety risks, making them key factors influencing the success or failure of drug candidates. If a candidate molecule performs poorly in certain key properties, it may lead to insufficient efficacy, increased adverse reactions, or subsequent development failure, thereby increasing research and development costs and risks. Therefore, accurate and efficient prediction of the key molecular properties of candidate molecules in the early stages of drug screening is crucial for guiding molecular design and optimization, improving screening efficiency, and reducing the failure rate of research and development. It has become one of the core technical aspects of the new drug development process.

[0004] Artificial intelligence (AI) technology, with its powerful pattern recognition and data-driven modeling capabilities, can efficiently process high-dimensional, nonlinear, and complex biological data. With the development of automated experimental techniques such as high-throughput screening, a large amount of data on the biological activity, toxicity, target effects, and other molecular properties of candidate molecules has been accumulated, providing an important data foundation for AI-based molecular property prediction models. By deeply exploring the potential patterns between molecular structural features and molecular properties, AI technology can achieve high-precision prediction of molecular properties, providing technical support for improving the efficiency and success rate of new drug development.

[0005] Currently, methods for predicting molecular properties mainly rely on various molecular characterization information such as molecular fingerprints, molecular graphs, molecular sequences, and molecular descriptors, and deep learning models that integrate multimodal data are gradually developing. Although multimodal feature fusion improves prediction performance to some extent, existing methods typically represent each modality feature independently and then directly concatenate them before inputting them into a shallow neural network or predictor for prediction. This direct concatenation approach fails to fully consider the differences among different compounds in terms of structural complexity, physicochemical properties, and various molecular characteristics, as well as the significant individual differences in the dependence on each modality's information. Therefore, existing models lack the ability to adaptively select and adjust features for individual molecular samples. The simple feature concatenation method implicitly assumes that the contribution of each modality feature in the prediction task is fixed and the same, lacking a dynamic modeling mechanism for modal importance. It is difficult to adaptively adjust the weights of different modal information according to the specific characteristics of the molecular sample, thus limiting the model's prediction accuracy and generalization ability.

[0006] A search revealed Chinese invention patent application publication number CN121075484A, which discloses a molecular property prediction method based on multimodal gating and contrastive learning. First, molecular fingerprint and molecular graph features are generated based on SMILES strings. A bimodal encoder is pre-trained using self-supervised contrastive learning to obtain a general feature representation. Then, a cross-modal attention mechanism is introduced to fuse the two modal features. Finally, molecular property prediction is completed based on the fused features. This scheme focuses on contrastive learning enhancement of bimodal features and fixed-structure attention fusion, but it does not introduce a molecular descriptor modality or implement dynamic weight allocation for individual molecular samples. This existing patent application suffers from the problem of using a fixed-structure cross-modal attention fusion, which cannot dynamically allocate modal weights according to individual molecular characteristics, thus limiting prediction accuracy and robustness.

[0007] How to achieve accuracy and robustness in molecular property prediction through multimodal adaptive adjustment has become a technical problem that needs to be solved. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the existing technology by providing an adaptive multimodal gated hybrid expert network for molecular property prediction, which provides important technical support for drug molecule design and optimization, pharmacokinetic property evaluation and preclinical research.

[0009] The objective of this invention can be achieved through the following technical solutions: According to one aspect of the present invention, a method for predicting molecular properties using an adaptive multimodal gated hybrid expert network is provided, the method comprising the following steps: Acquire compound data including annotations of molecular structure and molecular properties; The molecular structure is characterized to generate multimodal molecular features including molecular fingerprint features, molecular descriptor features, and molecular graph structure features; An adaptive multimodal gated hybrid expert network is constructed for adaptive fusion and joint representation of multimodal molecular features. The adaptive multimodal gated hybrid expert network includes a sub-expert model that extracts high-level semantic features for each modality of molecular features, and a gated network that adaptively fuses the high-level semantic features of each modality to output a fused feature representation. The fused feature representation is input into the prediction layer, and the molecular property prediction results are output.

[0010] As a preferred technical solution, in the adaptive multimodal gated hybrid expert network, corresponding molecular fingerprint sub-expert models, molecular descriptor sub-expert models, and molecular graph sub-expert models are constructed for molecular fingerprint features, molecular descriptor features, and molecular graph structure features, respectively, and high-level semantic features of each modality are obtained.

[0011] As a preferred technical solution, both the molecular fingerprint sub-expert model and the molecular descriptor sub-expert model are deep neural networks, which perform nonlinear mapping and high-level feature extraction on molecular fingerprint features and molecular descriptor features, respectively.

[0012] As a preferred technical solution, the molecular graph sub-expert model is a graph neural network, which takes atomic feature matrix, edge feature tensor and adjacency matrix as input, and obtains high-level semantic features of molecular graph modalities through graph encoding and global pooling.

[0013] As a preferred technical solution, the concatenated vector of molecular fingerprint features, molecular descriptor features and molecular graph encoding features is input into the gating network. The weights of each modality are learned and normalized through the gating network. The high-level semantic features of each modality are adaptively weighted and summed to obtain the fused feature representation.

[0014] As a preferred technical solution, in the feature construction, the string representing the molecular structure is parsed into a molecular object, a molecular fingerprint feature is generated based on the molecular object, and a molecular descriptor feature is calculated.

[0015] As a preferred technical solution, in the feature construction, molecular graph structure data is constructed based on the molecular structure to obtain molecular graph structure features.

[0016] As a preferred technical solution, the molecular fingerprint features are high-dimensional sparse or dense bit vectors, and the molecular descriptor features are continuous numerical vectors with physicochemical significance.

[0017] As a preferred technical solution, the molecular graph features are graph structure data with atoms as nodes and bonds as edges.

[0018] According to another aspect of the present invention, an electronic device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the program to implement the method described thereon.

[0019] Compared with the prior art, the present invention has the following beneficial effects: 1) This invention introduces three complementary modal features—molecular fingerprint, molecular descriptor, and molecular graph structure—simultaneously, and designs sub-expert models to extract high-level semantic features for each modal feature. It also introduces a gating mechanism to achieve adaptive weighted fusion, breaking through the limitations of traditional single-modality or fixed-weight fusion methods. This invention achieves effective collaboration and dynamic fusion of multimodal features, significantly improving the accuracy, robustness, and generalization ability of molecular property prediction.

[0020] 2) This invention achieves deep decoupling and targeted high-level semantic extraction of molecular fingerprints, molecular descriptors, and molecular graph structural features by constructing dedicated sub-expert models for different modal features. This avoids feature interference caused by multimodal mixed modeling, fully explores the unique advantages of each modal feature, and provides a high-quality single-modal feature foundation for subsequent adaptive fusion.

[0021] 3) This invention employs a deep neural network as a sub-expert model for molecular fingerprints and molecular descriptors, which can perform nonlinear mapping and feature extraction on high-dimensional discrete molecular fingerprints and physically meaningful molecular descriptors, effectively capturing the implicit patterns related to molecular properties in the two types of features, and improving the representation ability and discriminative power of single-modal features; it also employs a graph neural network as a molecular graph sub-expert model, which directly takes atomic feature matrices, edge feature tensors and adjacency matrices as inputs, and can naturally adapt to the topological structure information of molecules. Through graph encoding and global pooling, it automatically learns the deep interaction features of molecular structures, overcoming the shortcomings of traditional fingerprints and descriptors in representing the three-dimensional structural information of molecules, and greatly improving the prediction ability of structure-sensitive molecular properties.

[0022] 4) This invention achieves adaptive weighted fusion based on sample characteristics by concatenating multimodal features into a gating network, learning and normalizing the dynamic weights of each modality, thus ensuring the stability and interpretability of the weight allocation, and dynamically adjusting the contribution of each modality feature according to the structural characteristics of different molecules. This realizes the complementary advantages of multimodal features and significantly improves the expressive power of the fused features and the overall performance of the prediction model. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of different characterization modes of the molecular structure in this invention; Figure 2 This is a flowchart illustrating the molecular property prediction method of the present invention; Figure 3 This is a schematic diagram of the adaptive multimodal gated hybrid expert network in this invention; Figure 4 This is a comparison chart of the equilibrium accuracy metrics of the adaptive multimodal gated hybrid expert network and the comparative model in a classification-based molecular property prediction task. Figure 5 This is a comparison chart of the Matthews correlation coefficient between the adaptive multimodal gated hybrid expert network and the comparative model of the present invention in the molecular property prediction task of classification. Figure 6 This is a comparison chart of the F1 scores of the adaptive multimodal gated hybrid expert network and the comparative model in the molecular property prediction task of classification. Detailed Implementation

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

[0025] This embodiment relates to a molecular property prediction method based on an adaptive multimodal gated hybrid expert network. The method is based on a hybrid expert model framework with a gating mechanism. This framework designs multiple sub-expert models, each of which performs representation learning for molecular features of different modalities. The outputs of each sub-expert model are dynamically weighted and fused through a gating network, thereby achieving adaptive integration of multimodal features and improving the accuracy and robustness of molecular property prediction.

[0026] This method is based on multimodal feature fusion through the synergistic interaction of sub-expert models and gating networks. It aims to fully exploit the complementary information among multimodal features, enhancing the model's accuracy and robustness in molecular property prediction tasks. Each sub-expert model corresponds to a specific modality, used to extract feature representations related to that modality, thereby enhancing the specificity and effectiveness of multimodal feature representations. A gating network is employed as the fusion mechanism, dynamically adjusting the output weights of each sub-expert model based on the feature information of the input molecular samples. This achieves adaptive fusion of multimodal features, thereby improving the model's representation ability and generalization performance for different molecular samples.

[0027] The steps of this method are as follows: S1. Obtain compound molecular data from a public database. The compound molecular data includes the molecular structure represented in SMILES format and the corresponding molecular property annotations. For example... Figure 1The acquired compound molecular data is organized and processed according to the requirements of the molecular property prediction task and the preset data screening, cleaning and labeling rules. The samples are classified or numerically labeled according to the corresponding molecular property labels, and the dataset is divided to form a high-quality dataset for subsequent training, validation and testing of molecular property prediction models.

[0028] S2 involves feature construction of molecular structures represented in SMILES format within the dataset, using chemical tools to generate molecular fingerprint features, molecular descriptor features, and molecular graph structure features, such as... Figure 1 As shown. Specifically: as Figure 2 As shown, molecular structures are converted into molecular fingerprint features and molecular descriptor features, and corresponding molecular graph structure features are constructed using the DeepChem tool for subsequent model training and analysis. The molecular descriptor features are preprocessed and aligned, including missing value handling, outlier detection and handling, and feature standardization, to ensure the scale consistency and comparability of different samples in the feature space.

[0029] Molecules in the dataset are typically represented as SMILES. The RDKit tool library is used to parse the SMILES strings into molecular objects (Mol), and molecular fingerprints are generated and molecular descriptors are calculated based on these objects. The DeepChem tool is used to construct molecular graph structure data, providing a rich information foundation for subsequent multimodal feature fusion and model training.

[0030] S3, constructing an adaptive multimodal gated hybrid expert network to achieve adaptive fusion and joint characterization of multimodal molecular features. For example... Figure 3 Adaptive multimodal gated hybrid expert networks include multimodal hybrid expert models and gated networks, including: S31, Construct a multimodal hybrid expert model: For the different modal features obtained in S2, construct corresponding sub-expert models. Each sub-expert model specializes in processing one type of modal feature and is used to extract high-level feature representations from its respective modal features. Each sub-expert model is used to extract and learn features from the corresponding modal features, mine the deep semantic representation of its respective modal information, and enhance the expressive power and modal specificity of the corresponding modal features. Three types of modal features serve as inputs to the sub-expert models. Molecular fingerprint features are high-dimensional sparse or dense bit vectors; molecular descriptor features are continuous numerical vectors with physicochemical significance; and molecular graph features are graph structure data with atoms as nodes and bonds as edges. For molecular fingerprint and molecular descriptor features, deep neural network-based sub-expert models are constructed to perform nonlinear mapping and feature extraction on the corresponding features, outputting the embedded representation of the fingerprint modality and the deep representation vector of the descriptor, respectively. For molecular graph features, a graph neural network-based molecular graph sub-expert model is constructed. This model takes atomic feature matrices, edge feature tensors, and adjacency matrices as inputs. The graph neural network captures the topological structure and inter-atomic relationships of the molecular graph, encodes the molecular graph to obtain a graph-level feature representation, and further maps this graph-level feature representation using a deep neural network to obtain a high-level feature representation of the molecular graph modality. Global pooling is then used to obtain the graph-level molecular representation, and finally, a deep neural network is used to extract the high-level feature representation.

[0031] For molecular samples Its multimodal features are represented as By inputting the molecular fingerprint sub-expert model, molecular descriptor sub-expert model, and molecular graph sub-expert model respectively, the corresponding multimodal high-level semantic features are obtained: in, , , These respectively represent the molecular fingerprint features Molecular descriptor features and molecular diagram features The designed sub-expert model is used to extract high-level semantic features of each modality.

[0032] S32 introduces a gating mechanism to achieve adaptive fusion of multimodal high-level semantic features. A gating network is constructed to achieve adaptive fusion of multimodal features, taking molecular fingerprint features, molecular descriptor features, and molecular graph structure feature representations encoded by a graph neural network as input. To fully utilize the complementarity between different modal features and improve the model's comprehensive representation ability of molecular features, the gating network dynamically adjusts the weights of different modal features based on the feature information of the input molecular samples, and performs adaptive weighted fusion of feature representations to obtain the fused multimodal feature representation. The input to the gated network is a concatenation of molecular fingerprint, molecular descriptor, and molecular graph encoded features. The feature vectors are concatenated into a joint representation. , in, This is the concatenated joint feature vector. It is a molecular fingerprint feature. Features of molecular descriptors; Molecular diagram features Global feature representation of the molecular graph obtained after graph neural network encoding and global pooling; Indicates feature splicing, The modal weights for molecular fingerprint features, molecular descriptor features, and molecular graph features are respectively learned by the gating network.

[0033] By mapping through a multi-layer neural network, a weight vector corresponding to each sub-expert is output. The obtained weights are then used to perform a weighted summation of the features of each modality to obtain the final fused feature representation. : .

[0034] The gating network determines the importance of each modal feature and outputs the corresponding weight coefficients. These coefficients are then normalized to ensure the sum of all modal weights equals 1, guaranteeing the stability and interpretability of the fusion process. The weighted sum of each modal feature is then calculated based on its weight coefficients to obtain the fused multimodal comprehensive feature representation.

[0035] S4, fusing feature representations Input is fed into the prediction module, which outputs the predicted molecular properties of candidate molecules; The prediction module performs nonlinear mapping and discriminant analysis on the fused feature representation, and generates corresponding molecular property prediction results based on the multimodal feature representation to characterize the potential features of candidate molecules in terms of biological activity, drug-related properties or chemical behavior.

[0036] During the training of the integrated model (adaptive multimodal gated hybrid expert network and prediction model), labeled training data is used to jointly optimize the parameters of the prediction module, each sub-expert model, and the gated network. To improve the model's generalization ability and prevent overfitting, regularization techniques can be employed during training, and the training process can be regulated by monitoring the performance of the validation set. An appropriate loss function is selected based on the target molecule properties and the prediction task type to supervise the model's training. The training set is used for model fitting, the validation set is used for model parameter tuning and model selection, and finally, the model performance is evaluated on the test set.

[0037] The performance of the constructed adaptive multimodal gated hybrid expert network is evaluated. Using pre-defined evaluation metrics, the molecular property prediction results of the adaptive multimodal gated hybrid expert network in the molecular property prediction task are quantitatively evaluated to analyze its prediction performance.

[0038] In one embodiment, when the molecular property prediction task is a binary classification task, Balanced Accuracy (BA), F1 score, and Matthews Correlation Coefficient (MCC) are used as performance evaluation metrics, specifically defined as follows: , , , Where TP is the number of true positive cases, TN is the number of true negative cases, FP is the number of false positive cases, and FN is the number of false negative cases.

[0039] Based on the above evaluation indicators, using different cytochrome P450 (CYP) enzyme inhibitor prediction tasks as examples, the prediction results of the model proposed in this invention and existing models under the same data conditions are compared and analyzed.

[0040] Figure 4 This is a comparison chart showing the balanced accuracy metrics of the adaptive multimodal gated hybrid expert network (denoted as AGMoE-CYP in this embodiment) and a contrasting model in classification and prediction tasks, provided in this embodiment of the invention. Figure 5 This presentation shows the comparison results of each model on the MCC metric. Figure 6 The results show the comparison of the F1 score of each model.

[0041] The following is an explanation of the comparative model: GIN (Graph Isomorphic Network): A graph neural network model based solely on molecular graph structural features. It makes predictions by learning the representation of molecular topology, without introducing other modal features such as molecular fingerprints or molecular descriptors. It relies solely on single structural information and has limited ability to characterize the physicochemical properties of molecules.

[0042] GraphSAGE (Graph Sampling and Generative Embedding Model): Also a graph neural network model based solely on molecular graphs, it learns the embedding representation of molecular graphs by sampling neighbor nodes, but its structural feature extraction method is relatively simple and does not incorporate multimodal information for enhancement.

[0043] iCYP-MFE (Multi-Feature Integration Model): This multi-feature integration model uses molecular fingerprints and SMILES sequence representation. It only integrates features from two modalities, does not introduce molecular descriptor modalities, and does not achieve adaptive weight allocation. It belongs to an early multi-feature splicing and fusion method.

[0044] MuMCyp_Net (Multimodal CYP Prediction Network): It adopts a multimodal fusion model with fixed weights and uses molecular graphs, molecular fingerprints and molecular descriptor features. However, it uses a fixed weight or simple splicing fusion method and cannot dynamically adjust the modal contribution for different molecular samples.

[0045] DEEPCYPs (Deep CYP Prediction Model): A molecular property prediction model based on deep learning. It takes molecular fingerprint and molecular graph features as inputs. In order to further introduce molecular descriptor modes, it does not design a sample-level adaptive weight allocation mechanism for multimodal features.

[0046] GTransCYPs (Graph Transformer-based Prediction Model): This model is mainly based on molecular graph structure for modeling. It enhances the graph representation learning ability through Transformer convolutional layers and global attention pooling. It adopts fixed architecture attention fusion and single molecular graph modality modeling, without further integrating multi-source modal information such as molecular fingerprints and molecular descriptors, and does not achieve adaptive modality weight allocation for individual molecules.

[0047] It can be seen that the adaptive multimodal gated hybrid expert network proposed in this invention outperforms the comparative models in all evaluation metrics, indicating that the multimodal information fusion and adaptive modal weight adjustment mechanism can effectively improve the model discrimination ability and the stability of prediction results in molecular property prediction tasks.

[0048] Compared with existing technologies, this invention introduces a multimodal hybrid expert model with a gating mechanism to achieve adaptive fusion of multimodal molecular features. This enables the dynamic mining of complementary information between different features to address the heterogeneity of molecular structure and physicochemical properties, significantly improving the robustness of drug molecule property prediction tasks. It provides an intelligent solution for complex drug feature modeling and has strong practical value.

[0049] This embodiment also relates to a molecular property prediction method using an adaptive multimodal gated hybrid expert network. Taking aspirin as the target molecule and its water solubility (logS) regression prediction task as an example, it combines... Figure 1 The multimodal feature construction process shown in the figure fully illustrates the execution flow of the method of the present invention.

[0050] Step 1: Acquire compound molecular data and construct a dataset Obtain molecular data for aspirin from public databases (such as the TDC database), including: The molecular structure information is a linear string represented in SMILES format: CC(=O)OC1=CC=CC=C1C(=O)O; The molecular property information is labeled with the corresponding experimentally measured logS value of water solubility, which is labeled as -1.72.

[0051] The molecular data was included in the water solubility prediction dataset and divided into training, validation, and test sets in an 8:1:1 ratio along with other compound data. Aspirin molecules were used as test set samples to verify the model's predictive ability.

[0052] Step 2: Constructing multimodal molecular features based on SMILES Feature construction was performed on the SMILES string of aspirin to generate three types of modal features: Molecular fingerprint feature construction: The SMILES string is parsed into a molecular object (Mol) using the RDKit tool library. Based on this molecular object, the Morgan molecular fingerprint (radius 2, length 1024 bits) is generated, thus obtaining the molecular fingerprint feature. This fingerprint maps the atomic, bond, and functional group information of a molecule into a binary vector, where the position of "1" corresponds to the presence of specific substructures (such as ester groups, carboxyl groups, and benzene rings) in the aspirin molecule.

[0053] Molecular descriptor feature construction and preprocessing: Based on the aspirin molecular object, RDKit was used to calculate the physicochemical molecular descriptors, including molecular weight (MW=184), topological polar surface area (PSA=63.3), logP, and number of hydrogen bond donors. All descriptor features were preprocessed, including missing value handling, outlier handling, and Z-score normalization, to obtain aligned molecular descriptor features. .

[0054] Molecular graph feature construction: The DeepChem tool was used to parse the SMILES of aspirin into molecular graph structure data. Atomic features include atomic number, electronegativity, and hybridization type, while edge features include bond order and bond length. The data was then converted into a format that can be processed by a graph neural network using a graph encoding module to obtain the molecular graph features. This provides a data foundation for the input of subsequent sub-expert models.

[0055] Step 3: Construct an adaptive multimodal gated hybrid expert network and perform feature fusion. Multimodal sub-expert model construction and high-level feature extraction: Sub-expert models are constructed for the three types of modal features, including: Molecular fingerprint sub-expert model: Input is 1024-dimensional molecular fingerprint features After being mapped through a multi-layer fully connected network, high-level semantic features X are output. head1 ; Molecular descriptor sub-expert model: Input is standardized molecular descriptor features After being mapped through a multi-layer fully connected network, high-level semantic features X are output. head2 ; Molecular graph sub-expert model: Input is the molecular graph features of aspirin After graph convolutional neural network encoding and global pooling, the output molecular graph encoded features are obtained. and the corresponding high-level semantic features X head3 .

[0056] Gated networks and adaptive weighted fusion: concatenating molecular fingerprint features, molecular descriptor features, and molecular graph features into a joint feature vector. The joint feature vector serves as the input to the gating network, which learns and outputs the weight vectors for each modality. ω ={ } and after normalization, it satisfies + + =1. Adaptive weighted summation of the high-level semantic features of each modality is performed to obtain the fused feature representation. .

[0057] Step 4: Model Training and Molecular Property Prediction Representing fusion features Input prediction layer Output molecular property prediction results This embodiment is a regression task, using the mean squared error (MSE) loss function as a supervision signal to perform end-to-end joint training on the entire network (sub-expert model, gating network, and prediction layer). The training set is used for model fitting, and the validation set is used for model parameter tuning and early stopping to prevent overfitting.

[0058] The fusion characteristics of aspirin Input the trained prediction layer and output its predicted water solubility logS value of -1.74, with an absolute error of only 0.02 compared to the experimental true value of -1.72.

[0059] In this embodiment, the prediction result of aspirin water solubility based on the method of the present invention is significantly better than the prediction effect of traditional single-modality or fixed-weight fusion methods: The model using only molecular fingerprint features predicted a value of -1.22 with an absolute error of 0.50. The model using only molecular descriptor features predicted a value of -1.88 with an absolute error of 0.16. The model using only molecular graph features predicted a value of -1.99 with an absolute error of 0.25. The model using fixed-weight splicing and fusion has a prediction value of -1.12 and an absolute error of 0.60.

[0060] The above comparison shows that the adaptive multimodal gating fusion method of the present invention can dynamically allocate modal weights according to the structural characteristics of aspirin molecules, realize the complementary advantages of molecular fingerprints, descriptors and graph features, effectively reduce the prediction bias caused by single-modal or fixed-weight fusion, and significantly improve the accuracy and robustness of molecular property prediction.

[0061] The electronic device of this invention includes a central processing unit (CPU), which can perform various appropriate actions and processes according to computer program instructions stored in read-only memory (ROM) or loaded from a storage unit into random access memory (RAM). The RAM may also store various programs and data required for device operation. The CPU, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0062] Multiple components in the device are connected to the I / O interface, including: input units such as keyboards and mice; output units such as various types of displays and speakers; storage units such as disks and optical discs; and communication units such as network interface cards (NICs), modems, and wireless transceivers. The communication unit allows the device to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0063] The processing unit performs the various methods and processes described above. For example, in some embodiments, the methods may be implemented as computer software programs tangibly contained in a machine-readable medium, such as a storage unit. In some embodiments, part or all of the computer program may be loaded and / or installed on the device via ROM and / or a communication unit. When the computer program is loaded into RAM and executed by the CPU, one or more steps of the methods described above may be performed. Alternatively, in other embodiments, the CPU may be configured to execute the methods by any other suitable means (e.g., by means of firmware).

[0064] The functions described above in this document can be performed, at least in part, by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: Field Programmable Gate Arrays (FPGAs), Application-Specific Integrated Circuits (ASICs), Application Standard Products (ASSPs), System-on-Chip (SoCs), Complex Programmable Logic Devices (CPLDs), and so on.

[0065] The program code used to implement the methods of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0066] In the context of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory, optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0067] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for predicting molecular properties using an adaptive multimodal gated hybrid expert network, characterized in that, The method includes the following steps: Acquire compound data including annotations of molecular structure and molecular properties; The molecular structure is characterized to generate multimodal molecular features including molecular fingerprint features, molecular descriptor features, and molecular graph structure features; An adaptive multimodal gated hybrid expert network is constructed for adaptive fusion and joint representation of multimodal molecular features. The adaptive multimodal gated hybrid expert network includes a sub-expert model that extracts high-level semantic features for each modality of molecular features, and a gated network that adaptively fuses the high-level semantic features of each modality to output a fused feature representation. The fused feature representation is input into the prediction layer, and the molecular property prediction results are output.

2. The molecular property prediction method using an adaptive multimodal gated hybrid expert network according to claim 1, characterized in that, In the adaptive multimodal gated hybrid expert network, corresponding molecular fingerprint sub-expert models, molecular descriptor sub-expert models, and molecular graph sub-expert models are constructed for molecular fingerprint features, molecular descriptor features, and molecular graph structure features, respectively, and high-level semantic features of each modality are obtained.

3. The molecular property prediction method using an adaptive multimodal gated hybrid expert network according to claim 2, characterized in that, Both the molecular fingerprint sub-expert model and the molecular descriptor sub-expert model are deep neural networks, which perform nonlinear mapping and high-level feature extraction on molecular fingerprint features and molecular descriptor features, respectively.

4. The molecular property prediction method using an adaptive multimodal gated hybrid expert network according to claim 2, characterized in that, The molecular graph sub-expert model is a graph neural network that takes atomic feature matrix, edge feature tensor and adjacency matrix as input, and obtains high-level semantic features of molecular graph modalities through graph encoding and global pooling.

5. The molecular property prediction method using an adaptive multimodal gated hybrid expert network according to claim 1, characterized in that, The concatenated vector of molecular fingerprint features, molecular descriptor features, and molecular graph encoding features is input into the gating network. The gating network learns and normalizes the weights of each modality. The high-level semantic features of each modality are adaptively weighted and summed to obtain the fused feature representation.

6. The molecular property prediction method using an adaptive multimodal gated hybrid expert network according to claim 1, characterized in that, In the feature construction process, the string representing the molecular structure is parsed into a molecular object, a molecular fingerprint feature is generated based on the molecular object, and a molecular descriptor feature is calculated.

7. The molecular property prediction method using an adaptive multimodal gated hybrid expert network according to claim 1, characterized in that, In the feature construction, molecular graph structure data is constructed based on molecular structure to obtain molecular graph structure features.

8. The molecular property prediction method using an adaptive multimodal gated hybrid expert network according to claim 1, characterized in that, The molecular fingerprint features are high-dimensional sparse or dense bit vectors, and the molecular descriptor features are continuous numerical vectors with physicochemical significance.

9. The molecular property prediction method using an adaptive multimodal gated hybrid expert network according to claim 1, characterized in that, The molecular graph is characterized as graph structure data with atoms as nodes and bonds as edges.

10. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 9.

Images