Drug compound technology attribute evaluation method based on multi-modal deep learning

Abstract

The invention discloses a multi-modal deep learning-based drug compound technical attribute evaluation method, which comprises the following steps of: acquiring multi-source technical data of drug compounds by using a web crawler, acquiring key influence factors of drug technical attributes, and dividing the key influence factors into a low-correlation influence factor set and a high-correlation influence factor set, distributing different multi-modal fusion strategies for the low-correlation influence factor set and the high-correlation influence factor set; based on a long short-term memory network and a graph neural network, a multi-modal technology attribute evaluation model is constructed according to a hybrid multi-modal architecture and a decomposition-integration architecture, and characteristic parameters of key influence factors of a to-be-evaluated drug compound are obtained; and obtaining a compound technology attribute evaluation result for guiding drug research and development by using the multi-modal technology attribute evaluation model. According to the method, chemical informatics data are integrated through multi-modal deep learning, and efficient decision support is provided for drug research and development and synthetic route optimization scenes through accurate evaluation of compound technical attributes.

Description

Technical Field

[0001] The present invention relates to the field of chemical information processing technology, and more specifically, to a method for evaluating the technical properties of pharmaceutical compounds based on multimodal deep learning. Background Art

[0002] In the field of drug development and chemical synthesis, accurate evaluation of the technical properties of compounds (such as synthetic feasibility, drug-likeness, patent coverage, etc.) is crucial to improving R&D efficiency, reducing costs, and optimizing resource allocation. However, traditional evaluation methods mainly rely on empirical rules (or single data modality, and have the following limitations: (1) Traditional methods are usually based only on SMILES sequences or molecular descriptors, and fail to effectively integrate multi-source heterogeneous data (such as molecular structure images, synthesis process parameters, patent texts, etc.), resulting in insufficient information utilization. (2) Relying on artificial rules or simple statistical models (such as QSAR), it is difficult to capture the complex nonlinear relationship between molecular structure and properties, especially for rare skeleton compounds, the prediction deviation is large. (3) Models trained on a single dataset are prone to overfitting and are unstable when applied to new structural compounds or cross-scenario applications (such as evaluation of different synthetic routes). (4) Relying heavily on expert experience, the evaluation process is time-consuming and difficult to scale, and cannot meet the needs of high-throughput virtual screening.

[0003] Existing technologies use graph neural networks to process SMILES sequences or convolutional neural networks to analyze molecular images, but these methods fail to address the problem of multimodal data fusion. Different modal features are directly concatenated and fed into the prediction model, but the correlations and weight differences between the modalities are ignored, leading to information redundancy or conflict. Therefore, there is an urgent need for a multimodal deep learning method that can deeply integrate multi-source data such as chemical structure, synthesis parameters, and patent information, output an interpretable technical attribute index, and directly support R&D decision-making. Summary of the Invention

[0004] In order to solve the above technical problems, the present invention proposes a method for evaluating the technical properties of pharmaceutical compounds based on multimodal deep learning. By integrating chemical informatics data through multimodal deep learning and accurately evaluating the technical properties of compounds, it provides efficient decision support for drug development and synthesis route optimization scenarios.

[0005] The present invention provides a method for evaluating the technical properties of pharmaceutical compounds based on multimodal deep learning, comprising the following steps:

[0006] Using a web crawler to obtain multi-source technical data of pharmaceutical compounds, and performing data preprocessing on the obtained multi-source technical data of pharmaceutical compounds;

[0007] The influencing factors of compound technical properties are obtained by feature selection through pre-processed multi-source technical data, and the key influencing factors are screened using the improved particle swarm algorithm;

[0008] Obtain the correlation of key influencing factors and classify the data types into low-correlation influencing factor sets and high-correlation influencing factor sets, and assign different multimodal fusion strategies to the low-correlation influencing factor sets and the high-correlation influencing factor sets;

[0009] Based on the long short-term memory network and graph neural network, a multimodal technical attribute evaluation model is constructed according to the model-level fusion architecture and the decision-level fusion architecture. The characteristic parameters of key influencing factors are obtained according to the SMILES formula and common structural formula images of the pharmaceutical compounds to be evaluated. The multimodal technical attribute evaluation model is used to obtain the compound technical attribute evaluation results for guiding drug research and development.

[0010] In this solution, a web crawler is used to obtain multi-source technical data of pharmaceutical compounds, and the obtained multi-source technical data of pharmaceutical compounds is preprocessed as follows:

[0011] Using web crawler technology to capture multi-source technical data of different types of pharmaceutical compounds, including SMILES sequences, molecular structure images, and synthesis process parameters;

[0012] The captured multi-source technical data of different types of pharmaceutical compounds are cleaned, the SMILES sequences are standardized, the bond angles and resolution of the molecular structure images are normalized, and the reaction conditions in the synthesis process parameters are converted into one-hot encoding.

[0013] In this solution, feature selection is performed on pre-processed multi-source technical data to obtain the influencing factors of the technical properties of the compound, specifically:

[0014] Obtain pre-processed multi-source technical data for different types of pharmaceutical compounds, use the Pearson correlation coefficient to extract influencing factors related to the technical attributes of pharmaceutical compounds, and construct an initial set of influencing factors;

[0015] Based on the initial influencing factor set, four filtering feature selection methods, namely FCBF feature selection, ReliefF feature selection, Fisher score feature selection, and mRMR feature selection, are used as base selectors to perform basic feature selection to obtain influencing factor selection results;

[0016] The influencing factors output by each base selector are selected to construct a feature vector, and the feature vectors corresponding to each base selector are concatenated to generate a meta-feature matrix, which is used to train the meta-selector constructed based on the random forest algorithm;

[0017] The difference between the importance of the factors selected by the meta-selector and the final importance of the factors is minimized in the training process of the meta-selector, cross-validation is used to prevent overfitting, and the meta-selector hyperparameters are adjusted to optimize performance;

[0018] The trained meta-selector is used to predict the importance score of each factor in the initial factor set, the factors are sorted according to the importance score, and the top pre-set number of factors are extracted to form the initial factor set.

[0019] In this scheme, the improved particle swarm optimization algorithm is used to select the key factors, specifically:

[0020] The selected pre-set number of factors are feature-encoded, the feature-encoded factors are interval-divided, the initial factor combinations are randomly generated in each interval, the particle population is initialized according to the randomly generated initial factor combinations, and each particle corresponds to an initial factor combination;

[0021] The average importance score corresponding to each factor combination is obtained to construct an importance evaluation index, the dimension of the factor combination is obtained to construct a dimension evaluation index, the target function is constructed according to the importance evaluation index and the dimension evaluation index, the target function is associated with the fitness function, the fitness of each particle is calculated, and the particle individual optimal and global optimal are updated;

[0022] In the iterative search process of the particle swarm, particles with a fitness greater than a pre-set threshold are selected for copying, the search direction is determined, the inertia weight and learning factor dynamically change with the number of iterations to optimize the search process, and the speed and position of the particles are updated;

[0023] The encoding of the initial factor combination corresponding to the particle is changed in each search to expand new factor combinations, and it is judged whether the target function meets the termination criterion. When it is met, the best factor combination is obtained according to the optimal solution, and the factors in the best factor combination are used as the key factors.

[0024] In this scheme, the correlation degree of the key factors is obtained to classify the data types, and the low-correlation factor set and the high-correlation factor set are divided, and different multi-modal fusion strategies are assigned to the low-correlation factor set and the high-correlation factor set, specifically:

[0025] The Pearson correlation coefficients between different key factors are obtained, the correlation degree of the correlated factors is represented according to the Pearson correlation coefficients, a pre-set Pearson correlation coefficient threshold is set, the key factors lower than the Pearson correlation coefficient threshold are used as low-correlation factors, and the key factors not lower than the Pearson correlation coefficient threshold are used as high-correlation factors;

[0026] A multimodal fusion strategy is used to match the low-correlation influencing factors to the model level, and features are extracted from the digital parameter vectors corresponding to the low-correlation influencing factors, and then the feature representations of the different influencing factors are integrated;

[0027] A multimodal fusion strategy for matching the highly correlated influencing factors at the decision level is used, and the digital parameter vectors corresponding to each highly correlated influencing factor are used to independently train the model to generate prediction results, which are then integrated at the decision level.

[0028] In this solution, a multimodal technology attribute evaluation model is constructed based on the long short-term memory network and the graph neural network according to the model-level fusion architecture and the decision-level fusion architecture. Specifically:

[0029] A dual-channel multimodal technology attribute evaluation model is constructed based on the model-level fusion architecture and the decision-level fusion architecture. The feature extraction network of the model-level fusion channel and the independent model of the decision-level fusion channel are constructed by combining the long short-term memory network with the graph neural network.

[0030] Training data is generated by digitizing parameter vectors of low-correlation influencing factors and high-correlation influencing factors. The temporal dependencies between training data are extracted using parallel long-short-term memory networks with different time windows, and multi-scale temporal features are integrated through skip connections.

[0031] An influencing factor relationship graph is constructed based on the correlation between key influencing factors to characterize the synergistic effect between different technical attributes, the multi-scale time characteristics and the digital parameter vector are used to initialize nodes, and an adjacency matrix is ​​constructed based on the initialized nodes to represent the influencing factor relationship graph;

[0032] Importing the adjacency matrix into a graph neural network, performing aggregate updates on the nodes through a message passing mechanism, obtaining updated node representations, performing global pooling on the updated node representations of the entire graph, obtaining spatial features, and fusing the multi-scale temporal features with the spatial features to generate spatiotemporal features;

[0033] In the model-level fusion channel, the feature extraction networks corresponding to different influencing factors interact in the abstract feature space to build a joint inference network. In the joint inference network, the prediction technology attribute index corresponding to the spatiotemporal features of the training data is obtained through a fully connected layer.

[0034] In the decision-level fusion channel, the prediction technology attribute index corresponding to the spatiotemporal characteristics is obtained according to the independent models corresponding to different influencing factors. The prediction technology attribute index of each independent model is integrated using the decision integrator to obtain the final prediction technology attribute index.

[0035] The prediction technology attribute indexes output by the model-level fusion channel and the decision-level fusion channel are averaged, the verification data is used for result verification, and when the prediction performance meets the preset standard, the trained multi-modal technology attribute evaluation model is output.

[0036] In the scheme, according to the characteristic parameters of the key influencing factors obtained from the SMILES formula and the general structural formula image of the drug compound to be evaluated, the multi-modal technology attribute evaluation model is used to obtain the compound technology attribute evaluation result for guiding drug research and development, specifically:

[0037] The SMILES formula and the general structural formula of the drug compound to be evaluated are obtained, data crawling is performed according to the SMILES formula and the general structural formula of the drug compound to be evaluated, the crawled data is divided using the key influencing factors, and the characteristic parameters of the key influencing factors corresponding to the drug compound to be evaluated are generated;

[0038] The characteristic parameters of the key influencing factors corresponding to the drug compound to be evaluated are imported into the multi-modal technology attribute evaluation model, the prediction technology attribute indexes output by the model-level fusion channel and the decision-level fusion channel are obtained, and the prediction technology attribute indexes of the drug compound to be evaluated are obtained by averaging operation.

[0039] In the scheme, the multi-modal technology attribute evaluation model is evaluated by cross-validation, specifically:

[0040] The multi-modal technology attribute evaluation model is evaluated by cross-validation, the prediction technology attribute indexes output by the multi-modal technology attribute evaluation model are K-fold cross-validated, the overall accuracy of the model and the standard deviation of the K times validation results are calculated;

[0041] The index calculation result is compared with the preset index benchmark, and the multi-modal technology attribute evaluation model parameters are adjusted according to the cross-validation result.

[0042] Compared with the prior art, the beneficial effects of the present application are:

[0043] The present application integrates chemical informatics data (SMILES sequence, molecular structure image, synthesis process parameter, etc.) through multi-modal deep learning, constructs a compound technology attribute evaluation model, and provides efficient and accurate decision support in drug research and development, synthesis route optimization, patent evaluation and other scenes. Important factors are selected from implicit key factors, breaking through the limitations of traditional subjective experience or single data source; a crawler system and data cleaning are designed to obtain more efficient data; different multi-modal architectures are used for different data types, combining model-level and decision-level multi-modal fusion strategies, combining LSTM and GNN deep learning algorithms, introducing an evaluation model for more accurate compound technology attribute evaluation, outputting interpretable technology attribute evaluation indexes, and enhancing the decision support capability of drug research and development. BRIEF DESCRIPTION OF DRAWINGS

[0044] In order to more clearly illustrate the technical solutions in the embodiments or examples of the present application, the drawings needed to be used in the embodiments or examples will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0045] Figure 1 A flowchart of a method for evaluating technical attributes of drug compounds based on multi-modal deep learning is shown.

[0046] Figure 2 A flowchart of screening key influencing factors using an improved particle swarm algorithm is shown.

[0047] Figure 3 A flowchart of constructing a multi-modal technical attribute evaluation model is shown.

[0048] Figure 4 A block diagram of a system for evaluating technical attributes of drug compounds based on multi-modal deep learning is shown. DETAILED DESCRIPTION

[0049] In order to more clearly illustrate the technical solutions in the embodiments or examples of the present application, the drawings needed to be used in the embodiments or examples will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0050] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application, however, the present application can also be implemented in other ways different from those described herein, therefore, the scope of protection of the present application is not limited by the specific embodiments disclosed below.

[0051] Figure 1 A flowchart of a method for evaluating technical attributes of drug compounds based on multi-modal deep learning is shown.

[0052] As Figure 1 shown, the first embodiment of the present application provides a method for evaluating technical attributes of drug compounds based on multi-modal deep learning, comprising:

[0053] S102, using a web crawler to obtain multi-source technical data of drug compounds, and performing data preprocessing on the obtained multi-source technical data of drug compounds;

[0054] S104, selecting features through the preprocessed multi-source technical data to obtain influencing factors of compound technical attributes, and screening key influencing factors using an improved particle swarm algorithm;

[0055] S106, the correlation degree of the key influencing factors is obtained to perform data type classification, and is divided into a low-correlation influencing factor set and a high-correlation influencing factor set, and different multi-modal fusion strategies are assigned to the low-correlation influencing factor set and the high-correlation influencing factor set;

[0056] S108, a multi-modal technology attribute evaluation model is constructed based on a long short-term memory network and a graph neural network according to a model-level fusion architecture and a decision-level fusion architecture, feature parameters of the key influencing factors are obtained according to a SMILES formula and a general structural formula image of a to-be-evaluated drug compound, and a compound technology attribute evaluation result used for guiding drug research and development is obtained by using the multi-modal technology attribute evaluation model.

[0057] It should be noted that different kinds of multi-source technical data of drug compounds are captured based on a network crawler technology, and the multi-source technical data includes SMILES sequences, molecular structure images, and synthesis process parameters, etc.; the captured multi-source technical data of different kinds of drug compounds is cleaned, the SMILES sequences are standardized by using an RDKit toolkit to eliminate stereochemical differences; the molecular structure images are subjected to bond angle normalization and resolution standardization; and the reaction conditions in the synthesis process parameters are subjected to One-hot encoding conversion.

[0058] The application of stacking ensemble idea in feature selection is to combine the results of multiple base feature selection methods, and then use a meta-learner to integrate these results to obtain a better feature subset than a single method. The stacking ensemble feature selection process is used based on FCBF, ReliefF, Fisher Score and mRMR four filtering methods to extract the influencing factors of the technical properties of compounds. The pre-processed multi-source technical data of different types of drug compounds are obtained, the Pearson correlation coefficient is used to extract the influencing factors related to the technical properties of the drug compounds, an initial influencing factor set is constructed, and the preferred influencing factors in the initial influencing factor set include the number of hydrogen bond donors, the number of aromatic rings, etc. The influencing factors of the physical and chemical properties include LogP, TPSA, etc. The influencing factors of the synthesis process include step economy, maximum reaction temperature, etc. The influencing factors of the spatial characteristics include functional group spatial density, chiral center distribution, etc. Based on the initial influencing factor set, FCBF feature selection, ReliefF feature selection, Fisher score feature selection and mRMR feature selection are used as base selectors to perform base feature selection and obtain the influencing factor selection results. The FCBF feature selection calculates the symmetric uncertainty of each feature and the target variable, retains the features with symmetric uncertainty higher than a predetermined threshold, and removes redundant features from the retained features. The ReliefF feature selection finds the nearest neighbor (hit) in the same class and the nearest neighbor (miss) in the different class for each sample, updates the feature weight according to the distinguishing ability of the feature between different class samples, and arranges the features in descending order of feature weight after iteration for all samples. The Fisher score feature selection calculates the ratio of the between-class variance to the within-class variance of each feature, sorts the features according to the score from high to low, and selects the top k features with the highest score. The mRMR feature selection selects the feature with the maximum mutual information with the target variable, then iteratively selects the feature with the minimum redundancy with the selected features and the maximum correlation with the target, and uses an incremental search method to gradually construct the feature subset.

[0059] The feature vector of the selected result of each base selector output is constructed, for FCBF and mRMR, the selected features are set to 1, and the unselected features are set to 0, and for ReliefF and Fisher Score, the feature vector is constructed using the normalized feature weight or score, and the feature vectors corresponding to each base selector are spliced to generate a meta-feature matrix, and the meta-feature matrix is used to train a meta-selector constructed based on a random forest algorithm; the difference between the importance of the selected impact factors and the final importance of the impact factors is minimized in the training process of the meta-selector, cross-validation is used to prevent overfitting, and the performance of the meta-selector is optimized by adjusting the hyperparameters; the trained meta-selector is used to predict the importance score of each impact factor in the initial impact factor set, the importance score is sorted, and a preset number of impact factors are extracted according to the sorting result, and reverse feature elimination is performed to further optimize the impact factor subset, and the preferred key impact factors include molecular weight, hydrogen bond donor number, synthetic step complexity, molecular topological polarity, and number of chiral centers.

[0060] Figure 2 A flowchart of screening key impact factors using an improved particle swarm algorithm is shown.

[0061] According to an embodiment of the present application, the key impact factors are screened using an improved particle swarm algorithm, specifically:

[0062] S202, feature encoding is performed on the selected preset number of impact factors, the impact factors after feature encoding are interval divided, an initial impact factor combination is randomly generated in each interval, a particle population is initialized according to the randomly generated initial impact factor combination, and each particle corresponds to an initial impact factor combination;

[0063] S204, an importance evaluation index is constructed by obtaining the average importance score corresponding to each impact factor combination, a dimension evaluation index is constructed by obtaining the dimension of the impact factor combination, the target function is constructed according to the importance evaluation index and the dimension evaluation index, the target function is associated with the fitness function, the fitness of each particle is calculated, and the particle individual optimum and global optimum are updated;

[0064] S206, in the iterative search process of the particle swarm, the particles with a fitness greater than a preset threshold are selected for copying, the search direction is determined, the inertia weight and learning factor dynamically changing with the iteration number are used to optimize the search process, and the speed and position of the particles are updated;

[0065] S208, the coding of the initial impact factor combination corresponding to each particle is changed in each search, new impact factor combinations are expanded, it is judged whether the target function meets the termination criterion, when it meets, the best impact factor combination is obtained according to the optimal solution, and the impact factors in the best impact factor combination are taken as the key impact factors.

[0066] It should be noted that the traditional particle swarm algorithm mostly adopts fixed inertia weight and fixed learning factor, which is easy to fall into local optimal solution and cannot find global optimal solution. By adjusting the inertia weight and improving the learning factor, the search efficiency of the particle swarm algorithm is improved. The inertia weight dynamically changes with the iteration number, and the appropriate inertia weight shortens the iteration time and speeds up the optimization speed. The inertia weight w is expressed as where w max , w min are the maximum and minimum values of the inertia weight, t is the current iteration number, and T max is the maximum iteration number. The learning factor in the particle swarm algorithm determines the speed of the particle updating in the direction of the individual optimal and global optimal. A dynamic learning factor that nonlinearly changes with the iteration number is used to improve the efficiency of global optimization, where c1 and c2 are the individual learning factor and social learning factor, and q1, q2, and s are constants. According to the continuous adjustment of the improved inertia weight and learning factor to the position and speed of the particle itself, the iteration number is reached or the global optimal solution is found.

[0067] An importance evaluation index is constructed by obtaining the average importance score corresponding to each factor combination. The importance evaluation index is obtained based on the Gini impurity of the random forest feature importance. The dimension evaluation index is constructed by obtaining the dimension of the factor combination. The more factors are selected, the larger the dimension is, and the larger the calculation amount is. The objective function is constructed by maximizing the importance evaluation index and minimizing the dimension evaluation index.

[0068] It should be noted that the Pearson correlation coefficient between different key influencing factors is obtained, the correlation degree of the related influencing factors is represented according to the Pearson correlation coefficient, a Pearson correlation coefficient threshold is preset, the key influencing factors lower than the Pearson correlation coefficient threshold are regarded as low-correlation influencing factors, and the key influencing factors not lower than the Pearson correlation coefficient threshold are regarded as high-correlation influencing factors; a model-level multi-modal fusion strategy is matched for the low-correlation influencing factors, the digital parameter vectors corresponding to each low-correlation influencing factor are respectively extracted for features, different modal feature representations are integrated inside the model, and cross-modal feature combinations at the neuron level are allowed in feature interaction, a joint model is constructed for end-to-end prediction; a decision-level multi-modal fusion strategy is matched for the high-correlation influencing factors, models are independently trained using the digital parameter vectors corresponding to each high-correlation influencing factor to generate prediction results, and each model gradient flow is isolated from each other and integrated at the decision level.

[0069] Figure 3 A flowchart for constructing a multi-modal technology attribute evaluation model is shown.

[0070] According to an embodiment of the present application, a multi-modal technical attribute evaluation model is constructed based on a long short-term memory network and a graph neural network according to a model-level fusion architecture and a decision-level fusion architecture, specifically:

[0071] In S302, a double-channel multi-modal technical attribute evaluation model is constructed according to the model-level fusion architecture and the decision-level fusion architecture, and a feature extraction network of the model-level fusion channel and an independent model of the decision-level fusion channel are respectively constructed by combining the long short-term memory network with the graph neural network.

[0072] In S304, training data is generated by digital parameter vectors of low-correlation influence factors and high-correlation influence factors, time dependence between the training data is extracted by using parallel long short-term memory networks of different time windows, and multi-scale time features are integrated by a skip connection.

[0073] In S306, an influence factor relationship graph is constructed according to the correlation degree between key influence factors, representing the synergistic effect between different technical attributes, and the multi-scale time features and the digital parameter vectors are used for node initialization, and the influence factor relationship graph is represented based on the initialized nodes and an adjacency matrix.

[0074] In S308, the adjacency matrix is imported into the graph neural network, the nodes are aggregated and updated by a message passing mechanism, the updated node representations are obtained, the global pooling is performed on the updated node representations of the whole graph, the spatial features are obtained, and the multi-scale time features and the spatial features are fused to generate spatio-temporal features.

[0075] In S310, the feature extraction networks corresponding to different influence factors in the model-level fusion channel are interacted in an abstract feature space to construct a joint reasoning network, and in the joint reasoning network, the prediction technical attribute index corresponding to the spatio-temporal features of the training data is obtained by a fully connected layer.

[0076] In S312, in the decision-level fusion channel, the prediction technical attribute index corresponding to the spatio-temporal features is obtained according to the independent model corresponding to each influence factor, the prediction technical attribute index of each independent model is integrated by using a decision integrator, and the final prediction technical attribute index is obtained.

[0077] In S314, the prediction technical attribute indexes output by the model-level fusion channel and the decision-level fusion channel are averaged, the result is verified by using verification data, and when the prediction performance meets the preset standard, the trained multi-modal technical attribute evaluation model is output.

[0078] It should be noted that in the LSTM network, the parameters such as the synthesis process parameters are captured by using a bidirectional LSTM structure to capture the context dependence, 2-3 hidden layers are set to enhance the feature expression ability, an attention mechanism is added to dynamically weight the important time steps, and a Dropout layer (0.2-0.5 proportion) is used to prevent overfitting. Through parallel time windows of different lengths, multi-scale feature extraction is performed, the dynamic importance weight of each factor is learned for feature enhancement, the hidden state of the last time step is taken as the feature representation, and the multi-granularity time features are obtained by splicing the LSTM output of each layer. In the graph neural network, a graph attention network (GAT) or a graph convolution network (GCN) is used to set a 2-3 layer message passing network, convert the edge structure of the influence factor relationship graph into a high-dimensional representation, evaluate the synergistic effect between different technical attributes, and fuse the multi-scale time features output by the LSTM and the digital parameter vector for node initialization. After 2-3 rounds of graph convolution using the message passing mechanism, the final node representation is obtained, the global pooling of the node features of the whole graph is performed, and the spatial features with the same time feature dimension are output. The time features and the spatial features are fused to generate spatio-temporal features. Preferably, the time features and the spatial features are fused by using a sigmoid function to generate a dynamic fusion gate, and the time features and the spatial features are fused by using a gating mechanism. In the model-level fusion channel, the feature extraction networks corresponding to different influence factors interact in the abstract feature space, a joint reasoning network is constructed, and a softmax layer is used to output a predicted technical attribute index; in the decision-level fusion channel, a softmax layer is used to obtain a predicted technical attribute index corresponding to the spatio-temporal features in the prediction head of the independent model corresponding to each influence factor, and a decision integrator is used to integrate the predicted technical attribute indexes of each independent model to obtain a final predicted technical attribute index. The multi-modal technical attribute evaluation model combines the multi-modal fusion strategies at the model level and the decision level, combines the LSTM and GNN deep learning algorithms, accurately predicts the technical attribute index, and provides efficient and accurate decision support in the fields of drug research and development, synthesis route optimization, and patent evaluation. The technical attribute index includes a molecular synthesizability score, a biological activity prediction confidence, and a patent coverage score, etc. The molecular synthesizability score quantifies the difficulty of synthesizing a compound in a laboratory (0-100 points, the higher the score, the easier to synthesize), and the corresponding training data comes from synthesis process data and experimental verification labels corresponding to synthesis yield and synthesis difficulty expert scores. The biological activity prediction confidence predicts the reliability of the binding activity of a compound and a target (0-1, >0.8 for high confidence), and the corresponding training data comes from molecular docking scores calculated from biological experimental data, RMSD fluctuation values of molecular dynamics simulation, and uncertainty labels corresponding to experimental standard deviation.Patent coverage score assesses the extent to which a molecular structure is protected by existing patents (0-100 score, the higher the score, the greater the patent risk), the corresponding training data comes from the USPTO / EPO patent database Markush structure of claims, molecular structure matching data and the corresponding legal labels of patent infringement litigation results.

[0079] Simplified molecular-input line-entry system (SMILES) is based on the principle of molecular graph theory, which constructs a complete molecular structure description through the linear arrangement of atomic symbols, chemical bond markers and ring closure instructions. Obtain the SMILES formula and ordinary structural formula of the drug compound to be evaluated, perform data crawling according to the SMILES formula and ordinary structural formula of the drug compound to be evaluated, divide the crawled data using key influencing factors, and generate feature parameters of the key influencing factors corresponding to the drug compound to be evaluated. Convert the SMILES formula into an atomic coordinate matrix, which is equivalent to converting the molecular structure from ordinary electronic information into a three-dimensional model. This conversion not only identifies the chemical properties of the molecule, but also integrates the synthesis feasibility data and pharmacokinetic parameters (such as how long the drug will be metabolized in the human body) for comprehensive analysis. The feature parameters of the key influencing factors corresponding to the drug compound to be evaluated are imported into the multi-modal technical property evaluation model, the predicted technical property index output by the model-level fusion channel and the decision-level fusion channel is obtained, and the predicted technical property index of the drug compound to be evaluated is obtained by averaging operation. Preferably, the technical property index is used to screen candidate molecules with high synthesis feasibility and high biological activity, and a priority synthesis decision for drug research and development is obtained.

[0080] It should be noted that the cross-validation is used to evaluate the prediction ability of the multi-modal technical property evaluation model for the technical property index (such as synthesis feasibility score, patent coverage, drug-like property, etc.) of the compound, so as to ensure that the model has strong generalization, low bias and high stability. The data set is randomly and uniformly divided into K subsets, ensuring consistent data distribution in each fold. The multi-modal model is trained using K-1 fold data, the multi-modal output is integrated through the feature fusion layer, and the technical property index is predicted. The overall accuracy is divided into classification tasks and regression tasks, wherein the classification task calculates the average accuracy of K validations, and the regression task calculates the average R 2 score or root mean square error; the stability evaluation calculates the standard deviation (SD) of K validation results, which measures the volatility of the model. According to the field experience, the minimum acceptable standard is set, if the cross-validation result is better than the benchmark, the model can be directly deployed, if it does not meet the standard, the optimization process needs to be triggered. In the model optimization, the model adjustment strategies include adding or deleting key influencing factors, performing PCA dimensionality reduction on high-variance features; adjusting the number of graph convolution layers of GNN and the number of hidden units of LSTM; performing SMILES enumeration or image rotation enhancement on scarce data (such as rare molecular skeletons) and the like.

[0081] Figure 4A block diagram of a drug compound technical attribute evaluation system based on multi-modal deep learning is shown.

[0082] The second embodiment of the present application provides a drug compound technical attribute evaluation system 4 based on multi-modal deep learning, which comprises a web crawler module 401, an influence factor screening module 402, a multi-modal technical attribute evaluation module 403, and a precision evaluation module 404.

[0083] The web crawler module uses a web crawler to obtain multi-source technical data of drug compounds, and performs data preprocessing on the obtained multi-source technical data of drug compounds.

[0084] The influence factor screening module obtains the influence factors of the compound technical attributes through feature selection based on the preprocessed multi-source technical data, screens the key influence factors using an improved particle swarm algorithm, and divides them into a low-correlation influence factor set and a high-correlation influence factor set.

[0085] The multi-modal technical attribute evaluation module assigns different multi-modal fusion strategies to the low-correlation influence factor set and the high-correlation influence factor set, constructs a multi-modal technical attribute evaluation model based on a long short-term memory network and a graph neural network according to a model-level fusion architecture and a decision-level fusion architecture, obtains feature parameters of the key influence factors according to the SMILES formula and the general structural formula image of the drug compound to be evaluated, and uses the multi-modal technical attribute evaluation model to obtain the compound technical attribute evaluation results for guiding drug research and development.

[0086] The precision evaluation module uses cross-validation to evaluate the multi-modal technical attribute evaluation model, and optimizes and adjusts the multi-modal technical attribute evaluation model according to the model evaluation results.

[0087] The third embodiment of the present application provides a computer-readable storage medium comprising a drug compound technical attribute evaluation method program based on multi-modal deep learning, which, when executed by a processor, implements the steps of the drug compound technical attribute evaluation method based on multi-modal deep learning.

[0088] Those skilled in the art can understand that all or part of the steps of the above-mentioned method embodiments can be completed by program instruction-related hardware, and the aforementioned program can be stored in a computer-readable storage medium, which, when executed, executes steps including the above-mentioned method embodiments; and the aforementioned storage medium includes mobile storage devices, read-only memories (ROMs), random access memories (RAMs), magnetic discs or optical discs, and various media that can store program codes.

[0089] Alternatively, the above-mentioned integrated module of the present application, if realized in the form of a software function module and sold or used as an independent product, can also be stored in a computer-readable storage medium. Based on such an understanding, the technical solutions of the embodiments of the present application can be embodied in the form of a software product, which is stored in a storage medium and includes a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the methods described in the various embodiments of the present application. The aforementioned storage medium includes various media that can store program codes, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

[0090] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A method for evaluating the technical attributes of pharmaceutical compounds based on multimodal deep learning, characterized in that: The following steps are involved: Using a web crawler to obtain multi-source technical data of pharmaceutical compounds, and performing data preprocessing on the obtained multi-source technical data of pharmaceutical compounds; The influencing factors of compound technical properties are obtained by feature selection through pre-processed multi-source technical data, and the key influencing factors are screened using the improved particle swarm algorithm; Obtain the correlation of key influencing factors and classify the data types into low-correlation influencing factor sets and high-correlation influencing factor sets, and assign different multimodal fusion strategies to the low-correlation influencing factor sets and the high-correlation influencing factor sets; Based on the long short-term memory network and graph neural network, a multimodal technical attribute evaluation model is constructed according to the model-level fusion architecture and the decision-level fusion architecture. The characteristic parameters of key influencing factors are obtained according to the SMILES formula and common structural formula images of the pharmaceutical compounds to be evaluated. The multimodal technical attribute evaluation model is used to obtain the compound technical attribute evaluation results for guiding drug research and development.

2. The method for evaluating technical properties of pharmaceutical compounds based on multimodal deep learning according to claim 1, characterized in that: Use a web crawler to obtain multi-source technical data of pharmaceutical compounds, and perform data preprocessing on the obtained multi-source technical data of pharmaceutical compounds, specifically: Using web crawler technology to capture multi-source technical data of different types of pharmaceutical compounds, including SMILES sequences, molecular structure images, and synthesis process parameters; The captured multi-source technical data of different types of pharmaceutical compounds are cleaned, the SMILES sequences are standardized, the bond angles and resolution of the molecular structure images are normalized, and the reaction conditions in the synthesis process parameters are converted into one-hot encoding.

3. The method for evaluating the technical properties of pharmaceutical compounds based on multimodal deep learning according to claim 1, characterized in that: The influencing factors of compound technical properties are obtained by feature selection through pre-processed multi-source technical data, specifically: Obtain pre-processed multi-source technical data for different types of pharmaceutical compounds, use the Pearson correlation coefficient to extract influencing factors related to the technical attributes of pharmaceutical compounds, and construct an initial set of influencing factors; Based on the initial influencing factor set, four filtering feature selection methods, namely FCBF feature selection, ReliefF feature selection, Fisher score feature selection, and mRMR feature selection, are used as base selectors to perform basic feature selection to obtain influencing factor selection results; The influencing factors output by each base selector are selected to construct a feature vector, and the feature vectors corresponding to each base selector are concatenated to generate a meta-feature matrix, which is used to train the meta-selector constructed based on the random forest algorithm; During the meta-selector training process, minimize the difference between the importance of the factors selected by the meta-selector and the final importance of the factors, use cross-validation to prevent overfitting, and adjust the meta-selector hyperparameters to optimize performance; The trained meta-selector is used to predict the importance score of each influencing factor in the initial influencing factor set, and the factors are sorted according to the importance score. A preset number of influencing factors are extracted according to the sorting results to construct the initial influencing factor set.

4. The method for evaluating technical properties of pharmaceutical compounds based on multimodal deep learning according to claim 1, characterized in that: The improved particle swarm algorithm is used to screen the key influencing factors, specifically: Perform feature coding on a preset number of selected influencing factors, divide the feature-coded influencing factors into intervals, randomly generate initial influencing factor combinations within the intervals, and initialize the particle population based on the randomly generated initial influencing factor combinations, with each particle corresponding to one initial influencing factor combination; Obtain the average importance score corresponding to each influencing factor combination to construct an importance evaluation index, obtain the dimension of the influencing factor combination to construct a dimension evaluation index, construct an objective function based on the importance evaluation index and the dimension evaluation index, associate the objective function with the fitness function, calculate the fitness of each particle, and update the individual optimal particle and the global optimal particle. During the iterative search process of the particle swarm, particles with fitness greater than the preset threshold are selected for replication, the search direction is determined, and the inertia weight and learning factor that change dynamically with the number of iterations are used to optimize the search process and update the particle speed and position; In each search, the code of the particle corresponding to the initial influencing factor combination is changed, a new influencing factor combination is expanded, and it is judged whether the objective function meets the termination criterion. When it meets the criterion, the best influencing factor combination is obtained according to the optimal solution, and the influencing factors in the best influencing factor combination are used as key influencing factors.

5. The method for evaluating technical properties of pharmaceutical compounds based on multimodal deep learning according to claim 1, characterized in that: Obtain the correlation of key influencing factors and classify the data types into low-correlation influencing factor sets and high-correlation influencing factor sets. Different multimodal fusion strategies are assigned to the low-correlation influencing factor sets and the high-correlation influencing factor sets, specifically: Obtaining the Pearson correlation coefficients between different key influencing factors, characterizing the correlation of the associated influencing factors according to the Pearson correlation coefficients, presetting a Pearson correlation coefficient threshold, and defining key influencing factors below the Pearson correlation coefficient threshold as low-correlation influencing factors, and defining key influencing factors not lower than the Pearson correlation coefficient threshold as high-correlation influencing factors; A multimodal fusion strategy is used to match the low-correlation influencing factors to the model level, and features are extracted from the digital parameter vectors corresponding to the low-correlation influencing factors, and then the feature representations of the different influencing factors are integrated; A multimodal fusion strategy for matching the highly correlated influencing factors at the decision level is used, and the digital parameter vectors corresponding to each highly correlated influencing factor are used to independently train the model to generate prediction results, which are then integrated at the decision level.

6. The method for evaluating technical properties of pharmaceutical compounds based on multimodal deep learning according to claim 1, characterized in that: Based on the long short-term memory network and graph neural network, a multimodal technology attribute evaluation model is constructed according to the model-level fusion architecture and the decision-level fusion architecture. Specifically: A dual-channel multimodal technology attribute evaluation model is constructed based on the model-level fusion architecture and the decision-level fusion architecture. The feature extraction network of the model-level fusion channel and the independent model of the decision-level fusion channel are constructed by combining the long short-term memory network with the graph neural network. Training data is generated by digitizing parameter vectors of low-correlation influencing factors and high-correlation influencing factors. The temporal dependencies between training data are extracted using parallel long-short-term memory networks with different time windows, and multi-scale temporal features are integrated through skip connections. An influencing factor relationship graph is constructed based on the correlation between key influencing factors to characterize the synergistic effect between different technical attributes, the multi-scale time characteristics and the digital parameter vector are used to initialize nodes, and an adjacency matrix is ​​constructed based on the initialized nodes to represent the influencing factor relationship graph; Importing the adjacency matrix into a graph neural network, performing aggregate updates on the nodes through a message passing mechanism, obtaining updated node representations, performing global pooling on the updated node representations of the entire graph, obtaining spatial features, and fusing the multi-scale temporal features with the spatial features to generate spatiotemporal features; In the model-level fusion channel, the feature extraction networks corresponding to different influencing factors interact in the abstract feature space to build a joint inference network. In the joint inference network, the prediction technology attribute index corresponding to the spatiotemporal features of the training data is obtained through a fully connected layer. In the decision-level fusion channel, the prediction technology attribute index corresponding to the spatiotemporal characteristics is obtained according to the independent models corresponding to different influencing factors. The prediction technology attribute index of each independent model is integrated using the decision integrator to obtain the final prediction technology attribute index. The prediction technical attribute indexes output by the model-level fusion channel and the decision-level fusion channel are averaged, and the results are verified using validation data. When the prediction performance meets the preset standards, the trained multimodal technical attribute evaluation model is output.

7. The method for evaluating technical properties of pharmaceutical compounds based on multimodal deep learning according to claim 1, characterized in that: Based on the SMILES formula and common structural formula images of the pharmaceutical compound to be evaluated, the characteristic parameters of key influencing factors are obtained. A multimodal technical attribute evaluation model is used to obtain the compound technical attribute evaluation results used to guide drug development. Specifically: Obtain the SMILES formula and common structural formula diagram of the drug compound to be evaluated, crawl data based on the SMILES formula and common structural formula diagram of the drug compound to be evaluated, divide the crawled data using key influencing factors, and generate characteristic parameters corresponding to the key influencing factors of the drug compound to be evaluated; The characteristic parameters of the key influencing factors corresponding to the pharmaceutical compound to be evaluated are imported into the multimodal technical attribute evaluation model to obtain the predicted technical attribute index output by the model-level fusion channel and the decision-level fusion channel, and the predicted technical attribute index of the pharmaceutical compound to be evaluated is obtained by taking the average operation.

8. The method for evaluating technical properties of pharmaceutical compounds based on multimodal deep learning according to claim 1, characterized in that: Cross-validation is used to evaluate the multimodal technical attribute evaluation model, specifically: Cross-validation is used to evaluate the multimodal technical attribute evaluation model. The predicted technical attribute index output by the multimodal technical attribute evaluation model is subjected to K-fold cross-validation, and the overall accuracy of the model and the standard deviation of the K-fold validation results are calculated; Compare the indicator calculation results with the preset indicator benchmarks, and adjust the multimodal technology attribute evaluation model parameters based on the cross-validation results.

Images