A multi-task drug screening method and system integrating multimodal characterization

Abstract

This invention discloses a multi-task drug screening method and system integrating multimodal representations, belonging to the field of multi-task learning technology. The method includes collecting datasets on drug-target affinity, blood-brain barrier permeability, and neurotoxicity; employing multimodal feature representations (such as SMILES sequences, molecular graph structures, and images); and extracting drug molecular features using deep learning methods. Simultaneously, protein FASTA sequences are digitized and protein features are extracted using a Gated-CNN module. A multi-task learning mechanism is employed to jointly optimize drug and protein features, improving prediction performance. To resolve gradient conflicts in multi-task training, this invention also introduces the NeuGradBalancer optimization method, significantly improving training stability and convergence speed. This method can improve the accuracy and efficiency of drug screening, and is particularly suitable for the screening and optimization of drugs for the central nervous system.

Description

Technical Field

[0001] This invention relates to the field of multi-task learning technology, and in particular to a multi-task drug screening method and system that integrates multimodal representations. Background Technology

[0002] Existing computational drug screening methods mainly focus on predicting drug-target interaction affinity, blood-brain barrier permeability, or toxicity. However, these tasks are mostly isolated, single-prediction processes, performed independently, lacking an effective joint modeling mechanism. Traditional methods often rely on artificially designed molecular descriptors, sequence codes, or simple topological representations. These superficial features cannot fully capture the complex three-dimensional structural relationships between molecules and targets, as well as the mechanisms of action of functional substructures, thus limiting the expressive power and predictive accuracy of the models.

[0003] While existing deep learning methods have improved prediction accuracy to some extent, they generally suffer from the following limitations: First, most methods only utilize data from a single modality, such as using only chemical sequences or only graph structure information, making it difficult to integrate complementary information between different modalities. Second, the learning processes of each task are isolated and lack interaction, making it impossible to share potentially meaningful representations, which easily leads to accumulated errors and limits the overall performance of the model. Third, traditional models lack effective interpretability, making it difficult to locate key drug-target binding regions and their contribution to the prediction results. Fourth, for drug screening in the central nervous system, ontology methods often ignore the joint requirements that drugs must simultaneously meet the three dimensions of target binding, blood-brain barrier penetration, and nervous system safety, making it difficult for single-task models to balance safety and effectiveness. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0006] In a first aspect, the present invention provides a multi-task drug screening method that integrates multimodal characterization, comprising collecting drug-target affinity datasets, drug blood-brain barrier permeability datasets, and neurotoxicity datasets.

[0007] Preprocess the data in each dataset, removing invalid data and augmenting it;

[0008] The drug's SMILES sequences were converted into integer sequences, graph structures, and image features, and deep learning methods were used to extract multi-dimensional features of the molecular structure.

[0009] By converting the FASTA sequence of the target protein into a digital sequence and using the Gated-CNN module to extract protein features, long-range dependency information in the sequence is captured.

[0010] A multi-task learning mechanism is adopted, which learns and optimizes multiple preset prediction tasks by sharing features; the multi-dimensional features of the drug and the features of the protein are used to make predictions for each prediction task.

[0011] Among them, the NeuGradBalancer gradient optimization method is used to handle gradient conflicts in multi-task learning, thereby improving the stability and convergence speed of training.

[0012] As a preferred embodiment of the multi-task drug screening method integrating multimodal characterization described in this invention, wherein: the drug-target affinity dataset includes binding affinity data between small molecules and target proteins; and the neurotoxicity dataset includes neurocytotoxicity, neurobehavioral toxicity, and neurodevelopmental toxicity data.

[0013] The data preprocessing includes: removing invalid data, and removing non-covalent compounds, inorganic compounds, mixtures, molecules containing only salts, and compounds with a molecular weight greater than 1000 Da;

[0014] Enumeration enhancements are applied to the SMILES sequences by generating multiple equivalent SMILES representations to expand the dataset and improve the model's generalization ability.

[0015] As a preferred embodiment of the multi-task drug screening method integrating multimodal representations described in this invention, the multi-dimensional features include converting SMILES sequences into fixed-length integer sequences and inputting them into a three-layer bidirectional GRU network to extract potential sequence patterns in the drug sequences.

[0016] Drug molecules are represented as graph structures, and information is transferred and aggregated through a three-layer GCNConv graph convolutional network to extract high-order structural features from the molecules.

[0017] The drug's SMILES sequence is converted into a two-dimensional molecular structure image, which is then input into a ResNet18 network to extract image features, thereby enhancing the model's ability to perceive the overall molecular structure.

[0018] As a preferred embodiment of the multi-task drug screening method integrating multimodal characterization described in this invention, the protein feature extraction step includes:

[0019] The FASTA sequence of the target protein was converted into a digital sequence and normalized to a length of 1000.

[0020] The encoded protein sequence is embedded into a feature matrix and input into the Gated-CNN module to extract protein features, thereby capturing long-range dependency information;

[0021] The Gated-CNN module is a deep learning module that combines convolutional neural networks and gating mechanisms to extract and enhance key features in the input data.

[0022] As a preferred embodiment of the multi-task drug screening method integrating multimodal representations described in this invention, the multi-task learning mechanism jointly trains multiple tasks by sharing a feature space, the tasks including drug-target affinity prediction, blood-brain barrier permeability prediction, and neurotoxicity prediction.

[0023] As a preferred embodiment of the multi-task drug screening method integrating multimodal characterization described in this invention, the NeuGradBalancer gradient optimization method includes weighting the gradients of multiple tasks and aligning them on each shared parameter dimension.

[0024] Only gradient components aligned with the dominant gradient direction are retained for parameter updates, thereby reducing gradient conflicts between tasks.

[0025] As a preferred embodiment of the multi-task drug screening method integrating multimodal characterization described in this invention, the output of the prediction task includes predicted values ​​of protein-small molecule affinity, blood-brain barrier permeability, and neurotoxicity. The results are presented in the form of a heatmap to show the contribution of different molecular features to the prediction results, thereby enhancing the interpretability of the model.

[0026] Secondly, the present invention provides a multi-task drug screening system integrating multimodal characterization, including a data acquisition unit for collecting drug-target affinity datasets, drug blood-brain barrier permeability datasets, and neurotoxicity datasets; preprocessing the data in each dataset to remove invalid data and enhance it;

[0027] The analysis unit converts the drug's SMILES sequences into integer sequences, graph structures, and image features, and uses deep learning methods to extract multi-dimensional features of the molecular structure; it converts the FASTA sequence of the target protein into a digital sequence and uses the Gated-CNN module to extract protein features, while capturing long-range dependency information in the sequence.

[0028] The prediction unit employs a multi-task learning mechanism, learning and optimizing multiple preset prediction tasks through shared features; it uses the multi-dimensional features of the drug and the features of the protein to make predictions for each prediction task.

[0029] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein the computer program, when executed by the processor, implements any step of the multi-task drug screening method integrating multimodal characterization as described in the first aspect of the present invention.

[0030] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the multi-task drug screening method integrating multimodal characterization as described in the first aspect of the present invention.

[0031] The beneficial effects of this invention are as follows: By constructing a multimodal molecular characterization and multi-task joint learning framework, it achieves synergistic prediction of multiple key pharmacological properties such as drug-target affinity, blood-brain barrier permeability, and neurotoxicity. Compared with existing single-modal or single-task methods, it can fully explore the deep correlation information between molecular structure, protein sequence, and their interactions, significantly improving prediction accuracy and model stability. At the same time, by introducing a gradient dynamic balancing mechanism, it effectively alleviates the gradient conflict problem in the multi-task training process, improving the model convergence speed and generalization ability. In addition, this invention also has good interpretability and scalability, which can provide reliable technical support for the efficient screening and optimization of drugs for the central nervous system, reduce experimental costs, and improve R&D efficiency. Attached Figure Description

[0032] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 A flowchart for a multi-task drug screening method that integrates multimodal characterization.

[0034] Figure 2 An enumeration-enhanced graph for SMILES, a multi-task drug screening method that integrates multimodal characterization.

[0035] Figure 3 A representation of drug molecule characteristics for integrating multimodal characterization into a multi-task drug screening method.

[0036] Figure 4 Protein feature representation diagram for integrating multimodal characterization into a multi-task drug screening method.

[0037] Figure 5 A shared module diagram for integrating multi-modal characterization of multi-task drug screening methods.

[0038] Figure 6 Task-specific module diagram for integrating multi-modal characterization into a multi-task drug screening approach.

[0039] Figure 7A diagram of the NeuGradBalancer gradient optimization method for integrating multimodal characterization into a multi-task drug screening approach. Detailed Implementation

[0040] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0041] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0042] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0043] Reference Figures 1-7 This is one embodiment of the present invention, which provides a multi-task drug screening method integrating multimodal characterization, comprising the following steps:

[0044] S1: Collect drug-target affinity datasets, drug blood-brain barrier permeability datasets, and neurotoxicity datasets.

[0045] Specifically, for neurological diseases, drug-target affinity (protein small molecule affinity) datasets were collected, including KIBA, Davis, and BindingDB datasets, while obtaining the SMILES strings of the corresponding drugs and the FASTA sequences of the target proteins.

[0046] Because drugs for neurological diseases must simultaneously meet the requirements of effective target action, ability to cross the blood-brain barrier, and good neurological safety, we collected a blood-brain barrier (BBB) ​​permeability dataset and three types of neurotoxicity datasets, including neurocytotoxicity (NC), neurobehavioral toxicity (NA), and neurodevelopmental toxicity (NT).

[0047] S2: Preprocess the data in each dataset, remove invalid data, and perform augmentation.

[0048] The raw data was filtered to remove non-covalent compounds, inorganic compounds, mixtures, molecules containing only salts, and compounds with a molecular weight greater than 1000 Da. Compounds with inconsistent BBB permeability labeling information were also removed.

[0049] SMILES enumeration augmentation expands the training samples by generating multiple equivalent SMILES expressions for the same molecule without altering the fundamental molecular structure. This reduces sequence representation bias, improves model generalization ability, and stabilizes multi-task shared feature learning. SMILES-based enumeration augmentation was applied to the neurotoxicity dataset, resulting in expanded dataset sizes of NC: 9762 entries, NA: 9886 entries, and NT: 10262 entries.

[0050] For ease of understanding, the SMILES enumeration is enhanced as follows: Figure 2 As shown.

[0051] Protein-small molecule affinity is used to characterize the strength of binding between small molecules and target proteins, and is a key indicator for evaluating whether candidate drugs can effectively act on specific targets and produce therapeutic effects. The affinity values ​​in the protein-small molecule affinity dataset are processed to... value conversion This approach aims to standardize numerical scales and improve the stability of model training.

[0052] S3: The drug's SMILES sequence is converted into integer sequences, graph structures, and image features, and deep learning methods are used to extract multi-dimensional features of the molecular structure.

[0053] The SMILES sequences are converted into fixed-length integer sequences and fed into a three-layer bidirectional GRU network to extract latent sequence patterns from the drug sequence. The drug molecule is represented as a graph structure, and a three-layer GCNConv graph convolutional network is used to transfer and aggregate information from the molecular graph, thereby extracting high-order structural features. The drug's SMILES sequences are converted into two-dimensional molecular structure images and fed into a ResNet18 network to extract image features, enhancing the model's ability to perceive the overall molecular structure.

[0054] Furthermore, (1) Sequence feature representation: The drug SMILES is converted into a fixed-length integer sequence. If the length is less than 138, zero padding is performed, and the part exceeding the length is truncated. The sequence is then input into a three-layer bidirectional GRU network to model the atomic arrangement order and local chemical context information hidden in the SMILES sequence. 512-dimensional context-aware sequence features are extracted to characterize the sequence patterns in the molecule that are potentially related to the interaction with the target.

[0055] (2) Graph Structure Representation: Molecules are represented as graph structures, where atoms are mapped to graph nodes and chemical bonds to graph edges. The node feature dimension is 89, used to characterize the type, electronic properties, and chemical environment of atoms, while the edge feature dimension is 5, used to describe the bond types and connections between atoms. A three-layer GCNConv graph convolutional network is used to transfer and aggregate information from the molecular graph, enabling each atomic node to integrate its local chemical neighborhood information layer by layer, thereby learning the high-order structural representations related to the distribution of functional groups, electronic properties, and spatial connections in the molecule. This graph structure feature is optimized under the supervision of downstream drug-target affinity, blood-brain barrier permeability, and neurotoxicity tasks, allowing the model to focus on key atoms and local structures that contribute to target binding and biological activity. Subsequently, a graph-level molecular representation is obtained through node pooling operations, used to characterize the structural features related to targeting at the overall molecular level, rather than simply geometric or stability descriptions.

[0056] (3) Image feature representation: SMILES is converted into two-dimensional molecular structure images and unified into 112×112 RGB image format. The images are then input into the ResNet18 network to extract 512-dimensional image features to supplement information on the molecular topology and spatial configuration in two dimensions, thereby enhancing the model's ability to perceive the overall molecular structure pattern.

[0057] (4) Feature modeling and compression: Two Transformer encoders are introduced for sequence features and graph structure features to model the dependencies between long-range atoms or structural fragments in the molecule. The importance of different structural components is weighted by attention pooling mechanism to finally obtain a fixed-length and interpretable molecular feature representation, providing a unified input for subsequent multi-task prediction.

[0058] Specifically, drug molecular characterization includes, for example: Figure 3 As shown.

[0059] S4: By converting the FASTA sequence of the target protein into a digital sequence and using the Gated-CNN module to extract protein features, long-range dependency information in the sequence is captured.

[0060] The target protein FASTA sequence was mapped to a numerical sequence using ASCII encoding, and the length was standardized to 1000. Parts with insufficient length were padded with zeros, and parts with excessive length were truncated.

[0061] The encoded sequence is embedded into a 128×1000 feature matrix and fed into a Gated-CNN module to extract protein features while capturing long-range dependencies in the sequence. The Gated-CNN module is a deep learning module combining convolutional neural networks (CNNs) and gating mechanisms to extract and enhance key features from the input data. In this module, convolutional operations are used to extract local features, while the gating mechanism automatically selects which features should be preserved and which should be suppressed by controlling the activation level of the information flow, thereby improving the model's feature learning ability. By introducing a gating mechanism, this module can effectively capture long-range dependencies in the input data. The Gated-CNN module includes: convolutional layers for extracting local features from the input protein sequence data to learn local patterns in the sequence; gating layers for controlling the activation level of information flow using activation functions (e.g., sigmoid or tanh functions) to determine which features should be enhanced and which should be suppressed, thereby strengthening the focus on important features; pooling layers for pooling the features after convolution to reduce feature dimensionality and retain the most important feature information; and fully connected layers for mapping the extracted features to a higher-level representation for subsequent prediction tasks.

[0062] For ease of understanding, protein characteristics are represented as follows: Figure 4 As shown

[0063] S5: Employ a multi-task learning mechanism, learn and optimize multiple preset prediction tasks by sharing features; and use the multi-dimensional features of the drug and the features of the protein to make predictions for each prediction task.

[0064] Multimodal Sharing Module: Early Fusion and Construction of a Shared Latent Space for Multimodal Features: After high-level semantic abstraction of drug sequence features, graph structure features, and image features through deep networks, a unified feature modeling and compression mechanism is introduced to ensure consistency in dimensionality, semantic level, and numerical distribution across different modalities. Specifically, before fusion, sequence and graph structure features are modeled with long-range dependencies using a Transformer encoder and mapped to fixed-length vectors via attention pooling; image features are represented by high-level feature mapping using ResNet18 to match their semantic level. Through this process, different modal features are no longer original heterogeneous inputs but comparable high-order molecular semantic representations. Based on this, a weighted summation-based early fusion method is used to map the three types of modal features to a unified shared latent space, thus avoiding scale inconsistencies, feature redundancy, and noise accumulation problems caused by direct splicing.

[0065] The shared module's multi-task universal representation learning mechanism: The constructed shared latent space is not a simple superposition of multimodal features, but rather dynamically optimized during multi-task joint training through joint supervision from tasks such as protein-small molecule affinity, BBB permeability, and neurotoxicity. This allows the shared representation to simultaneously retain key information discriminative for target binding, pharmacokinetic properties, and safety assessment. This shared module effectively alleviates feature conflicts between different tasks, avoids single-task-dominated feature learning, and enables the model to extract biologically meaningful universal molecular representations in multi-task scenarios, serving as common input features for subsequent prediction tasks.

[0066] For ease of understanding, the structure of the multimodal shared module is as follows: Figure 5 As shown.

[0067] Task-specific module: Protein small molecule affinity task module: Input drug features and protein features into the mutual attention mechanism module for feature interaction, and output drug-target affinity prediction results through a three-layer fully connected network and gated attention mechanism.

[0068] BBB and Neurotoxicity Task Module: For BBB permeability and NC, NA, and NT neurotoxicity tasks, a two-layer fully connected classification network is used, and the predicted probability of the corresponding category is output through the Sigmoid function.

[0069] For ease of understanding, the structure of a task-specific module is as follows: Figure 6 As shown.

[0070] Feature fusion strategies: Early fusion strategy: Weighted summation of feature vectors from different modalities according to set weights to form a unified input representation. Late fusion strategy: Weighted average of prediction results from each modality or sub-model to output the final prediction result.

[0071] Among them, the NeuGradBalancer gradient optimization method is used to handle gradient conflicts in multi-task learning, thereby improving the stability and convergence speed of training.

[0072] In multi-task learning, the gradient directions and magnitudes of different tasks are inconsistent, which can easily lead to unstable updates of shared parameters and even performance degradation. Traditional gradient weighting or simple summation is difficult to reconcile task conflicts. The NeuGradBalancer method of this invention effectively avoids gradient interference by aligning and filtering gradients of multiple tasks along the parameter dimension, achieving task co-optimization and significantly improving training stability, convergence speed, and overall performance.

[0073] First, the gradients of the shared parameters from the five tasks are weighted and aligned on each shared parameter dimension.

[0074] Subsequently, the dominant direction of the multi-task gradient is determined across various parameter dimensions.

[0075] Only the gradient component consistent with the dominant direction is retained for parameter updates.

[0076] The above mechanism effectively alleviates gradient conflicts between different tasks, thereby improving the stability and optimization efficiency of the model training process.

[0077] For ease of understanding, the NeuGradBalancer gradient optimization method is as follows: Figure 7 As shown.

[0078] Furthermore, regarding the loss function and model training: the mean squared error (MSE) was used as the loss function for the small molecule protein affinity task, while the binary cross-entropy (BCE) loss function was used for the BBB permeability and neurotoxicity tasks. The total model loss was the weighted sum of the losses from the five tasks. The Adam optimizer was used for parameter updates, with a learning rate of 0.0002, a batch size of 128, and 100 training epochs. The ResNet18 module was initialized using ImageNet pre-trained parameters, while the remaining network modules were trained using random initialization.

[0079] The NeuGradBalancer module calculates the gradient information for each task and calculates the purity of gradient signs across different tasks for each parameter dimension. It selects gradient components based on a random thresholding strategy to construct an aggregated gradient. This aggregated gradient is then used to update the parameters of the shared module, improving the stability and convergence of the multi-task joint optimization process.

[0080] The system supports both single-molecule and batch molecule prediction modes, outputting predicted small molecule protein affinity, BBB permeability probability, and neurotoxicity risk assessment results. The contribution of different molecular features to the prediction results is displayed through visual heatmaps, enhancing the model's interpretability. Users can export the prediction results for subsequent analysis or experimental validation.

[0081] This embodiment also provides a multi-task drug screening system that integrates multimodal characterization, including: a data acquisition unit that collects drug-target affinity datasets, drug blood-brain barrier permeability datasets, and neurotoxicity datasets; and preprocesses the data in each dataset, removes invalid data, and enhances the data.

[0082] The analysis unit converts the drug's SMILES sequences into integer sequences, graph structures, and image features, and uses deep learning methods to extract multi-dimensional features of the molecular structure; it converts the FASTA sequence of the target protein into a digital sequence and uses the Gated-CNN module to extract protein features, while capturing long-range dependency information in the sequence.

[0083] The prediction unit employs a multi-task learning mechanism, learning and optimizing multiple preset prediction tasks through shared features; it uses the multi-dimensional features of the drug and the features of the protein to make predictions for each prediction task.

[0084] This embodiment also provides a computer device suitable for multi-task drug screening methods that integrate multimodal characterization, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the multi-task drug screening method that integrates multimodal characterization as proposed in the above embodiment.

[0085] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0086] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the multi-task drug screening method integrating multimodal characterization as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0087] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A multi-task drug screening method integrating multimodal characterization, characterized in that: include, Collect drug-target affinity datasets, drug blood-brain barrier permeability datasets, and neurotoxicity datasets; Preprocess the data in each dataset, removing invalid data and augmenting it; The drug's SMILES sequences were converted into integer sequences, graph structures, and image features, and deep learning methods were used to extract multi-dimensional features of the molecular structure. By converting the FASTA sequence of the target protein into a digital sequence and using the Gated-CNN module to extract protein features, long-range dependency information in the sequence is captured. A multi-task learning mechanism is adopted, which learns and optimizes multiple preset prediction tasks by sharing features; the multi-dimensional features of the drug and the features of the protein are used to make predictions for each prediction task. Among them, the NeuGradBalancer gradient optimization method is used to handle gradient conflicts in multi-task learning, thereby improving the stability and convergence speed of training.

2. The multi-task drug screening method integrating multimodal characterization as described in claim 1, characterized in that: The drug-target affinity dataset contains data on the binding affinity between small molecules and target proteins; the neurotoxicity dataset includes data on neurocytotoxicity, neurobehavioral toxicity, and neurodevelopmental toxicity. The data preprocessing includes: removing invalid data, and removing non-covalent compounds, inorganic compounds, mixtures, molecules containing only salts, and compounds with a molecular weight greater than 1000 Da; Enumeration enhancements are applied to the SMILES sequences by generating multiple equivalent SMILES representations to expand the dataset and improve the model's generalization ability.

3. The multi-task drug screening method integrating multimodal characterization as described in claim 2, characterized in that: The multidimensional features include converting the SMILES sequence into a fixed-length integer sequence and inputting it into a three-layer bidirectional GRU network to extract potential sequence patterns in the drug sequence. Drug molecules are represented as graph structures, and information is transferred and aggregated through a three-layer GCNConv graph convolutional network to extract high-order structural features from the molecules. The drug's SMILES sequence is converted into a two-dimensional molecular structure image, which is then input into a ResNet18 network to extract image features, thereby enhancing the model's ability to perceive the overall molecular structure.

4. The multi-task drug screening method integrating multimodal characterization as described in claim 3, characterized in that: The protein feature extraction steps include: The FASTA sequence of the target protein was converted into a digital sequence and normalized to a length of 1000. The encoded protein sequence is embedded into a feature matrix and input into the Gated-CNN module to extract protein features, thereby capturing long-range dependency information; The Gated-CNN module is a deep learning module that combines convolutional neural networks and gating mechanisms to extract and enhance key features in the input data.

5. The multi-task drug screening method integrating multimodal characterization as described in claim 4, characterized in that: The multi-task learning mechanism performs joint training on multiple tasks by sharing a feature space. These tasks include drug-target affinity prediction, blood-brain barrier permeability prediction, and neurotoxicity prediction.

6. The multi-task drug screening method integrating multimodal characterization as described in claim 5, characterized in that: The NeuGradBalancer gradient optimization method includes weighting the gradients of multiple tasks and aligning them on each shared parameter dimension. Only gradient components aligned with the dominant gradient direction are retained for parameter updates, thereby reducing gradient conflicts between tasks.

7. The multi-task drug screening method integrating multimodal characterization as described in claim 6, characterized in that: The output of the prediction task includes predicted values ​​for protein-small molecule affinity, blood-brain barrier permeability, and neurotoxicity. These results are presented in the form of heatmaps to show the contribution of different molecular features to the prediction results, thereby enhancing the interpretability of the model.

8. A multi-task drug screening system integrating multimodal characterization, based on the multi-task drug screening method integrating multimodal characterization as described in any one of claims 1 to 7, characterized in that: include, The data acquisition unit collects drug-target affinity datasets, drug blood-brain barrier permeability datasets, and neurotoxicity datasets. Preprocess the data in each dataset, removing invalid data and augmenting it; The analysis unit converts the drug's SMILES sequences into integer sequences, graph structures, and image features, and uses deep learning methods to extract multi-dimensional features of the molecular structure; it converts the FASTA sequence of the target protein into a digital sequence and uses the Gated-CNN module to extract protein features, while capturing long-range dependency information in the sequence. The prediction unit employs a multi-task learning mechanism, learning and optimizing multiple preset prediction tasks through shared features; it uses the multi-dimensional features of the drug and the features of the protein to make predictions for each prediction task.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the multi-task drug screening method integrating multimodal characterization as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the multi-task drug screening method integrating multimodal characterization as described in any one of claims 1 to 7.

Images