A multi-source information fusion drug response prediction method and system
By employing a multi-source information fusion-based drug response prediction method, and utilizing graph coding networks and pre-trained language models, the unified modeling problem for single-drug and multi-drug predictions is solved, thereby improving the accuracy of drug response prediction and the adaptability of the model.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ANHUI UNIVERSITY OF TRADITIONAL CHINESE MEDICINE
- Filing Date
- 2026-04-29
- Publication Date
- 2026-07-24
AI Technical Summary
In existing technologies, there is a lack of a unified modeling framework for single-drug sensitivity prediction and multi-drug synergistic prediction. Multimodal information is not fully utilized, making it difficult to characterize the deep relationship between drugs and cells. Furthermore, the predictive performance of the models is unstable in small-sample real-world scenarios.
By acquiring drug structure information, drug text description information, and cell background feature information, we use graph coding networks and pre-trained language models for encoding, combine attention mechanisms for information fusion, and construct a unified response prediction model to achieve unified processing of single-drug and multi-drug tasks.
It improves the ability to characterize the relationship between drugs and cells, enhances the model's generalization and adaptability in small-sample real-world scenarios, and improves the accuracy and consistency of predictions.
Smart Images
Figure CN122455404A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of bioinformatics and computational pharmacology, specifically to a method and system for predicting drug response through multi-source information fusion. Background Technology
[0002] Drug response prediction is a key technology in precision medicine, computational pharmacology, and computer-aided drug discovery. Its purpose is to predict the inhibitory intensity of a drug in target cells or the synergistic effect of drug combinations based on drug characteristics, cellular molecular states, and their interactions. In existing technologies, researchers typically utilize the chemical structure representation of the drug, the gene or transcriptome expression characteristics of the cell line, and the corresponding half-maximal inhibitory concentration (IC50) or synergistic scoring labels to construct machine learning or deep learning models to support candidate drug screening and combination therapy design. In the area of single-drug response prediction, existing methods often employ one or more representations from molecular fingerprints, SMILES sequences, molecular graphs, or drug description texts, combined with cell line expression vectors to build regression models. For example, existing technologies include schemes that use graph neural networks to encode drug molecular graphs and predict drug sensitivity, as well as schemes that combine drug graph structure with cell line omics information for multimodal modeling.
[0003] However, the existing technologies still have the following shortcomings: Existing methods typically establish independent models for single-drug sensitivity prediction and multi-drug synergistic prediction, lacking a unified modeling framework, making it difficult to transfer the drug-cell interaction patterns learned in single-drug scenarios to multi-drug scenarios; the utilization of multimodal information is still insufficient, with most methods only using single-modal information or fusing multimodal inputs through simple splicing, making it difficult to characterize the deep correlation between drug structural information, drug semantic information, and cellular background information; the granularity of cell state characterization is relatively coarse, usually based on gene level or overall expression vector modeling, making it difficult to reflect the differential drug responses under different cellular backgrounds; and training and evaluation are mostly based on publicly available datasets, while in real-world application scenarios, the data scale is small and there are distribution differences, resulting in unstable predictive performance of the model in real-world scenarios. Summary of the Invention
[0004] The purpose of this invention is to provide a drug response prediction method and system that integrates multi-source information to solve the above-mentioned technical problems. This method and system can unify the prediction of single-drug sensitivity and the prediction of multi-drug synergistic prediction, and improve the generalization and adaptability of the model in small sample real-world scenarios while integrating multi-source information.
[0005] The objective of this invention can be achieved through the following technical solutions: A drug response prediction method based on multi-source information fusion, comprising: S1. Obtain sample data, which includes single-drug samples and multi-drug samples. Both single-drug samples and multi-drug samples shall include at least: drug structure information, drug text description information, and cell background feature information. S2. Encode drug structural information and drug textual description information, and output drug structural feature representation and drug semantic feature representation; construct cell representation based on cell background feature information; S3. Based on the drug structural feature representation, guide the fusion of the drug semantic feature representation and output the drug fusion representation; S4. Based on the task type, combine the drug fusion representation with the cell representation to construct the model input features, where the task type is used to indicate single-drug response prediction tasks and multi-drug synergistic prediction tasks. S5. Input the model input features into a unified response prediction model and output the corresponding drug response prediction results. When the task type is a single-drug response prediction task, output the single-drug response value; when the task type is a multi-drug synergistic prediction task, output the multi-drug synergistic response value.
[0006] Furthermore, the drug structure information includes simplified molecular linear input specifications, the drug text description information includes textual descriptions of the drug, and the cell background feature information includes transcriptomic expression characteristics of the cell line and the corresponding half-maximal inhibitory concentration.
[0007] Furthermore, in step S2, the drug structure information of each sample data is represented as a three-dimensional molecular graph, and the drug structure features are extracted using a graph coding network; the three-dimensional molecular graph uses atoms as nodes, chemical bonds as edges, and contains the spatial coordinate information of atoms.
[0008] Furthermore, in step S2, the drug text description information is segmented and encoded using a pre-trained language model to output the drug semantic feature representation; in the multi-drug task, multiple drugs are processed using a parameter-shared structural encoder and a text encoder respectively.
[0009] Furthermore, in step S3, based on the attention mechanism, the drug structural feature representation is used as conditional information to perform weighted processing on the drug semantic feature representation, and the drug fusion representation is output.
[0010] Furthermore, in step S2, the cell background feature information is vectorized to output a cell representation.
[0011] Furthermore, in step S4, when the task type is a single-drug response prediction task, the single drug fusion representation is combined with the cell representation and the task type identifier; when the task type is a multi-drug synergistic prediction task, the multiple drug fusion representations are combined with the cell representation and the task type identifier.
[0012] Furthermore, in step S5, the response prediction model is a regression model based on a neural network.
[0013] Furthermore, after obtaining the response prediction model, the parameters of the response prediction model are adjusted based on the target data, and the drug response is predicted based on the adjusted response prediction model.
[0014] A drug response prediction system based on multi-source information fusion, the system comprising: The data acquisition module is used to acquire sample data, which includes single-drug samples and multi-drug samples. The sample data includes at least: drug structure information, drug text description information, and cell background feature information. The encoding module is used to encode drug structural information and drug textual description information respectively to obtain drug structural feature representation and drug semantic feature representation, and to construct cell representation based on cell background feature information; The fusion module is used to guide the fusion of drug semantic feature representation based on drug structural feature representation to obtain drug fusion representation; The input building module is used to combine a single drug fusion representation or a two-drug fusion representation with a cell representation to build model input features based on the task type, where the task type is used to indicate a single-drug response prediction task or a multi-drug synergistic prediction task. The response prediction module is used to input the model input features into a unified response prediction model and output the corresponding drug response prediction results. Specifically, when the task type is a single-drug response prediction task, the single-drug response value is output; when the task type is a multi-drug synergistic prediction task, the multi-drug synergistic response value is output.
[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention unifies the modeling of single-drug response prediction tasks and multi-drug synergistic prediction tasks in the same response prediction model, so that the drug-cell interaction relationship obtained based on drug fusion representation and cell representation learning can be shared between different tasks, thereby avoiding the structural fragmentation problem caused by separate modeling. 2. This invention utilizes drug structural information, drug textual description information, and cell background feature information simultaneously, and guides the fusion of drug semantic feature representation based on drug structural feature representation to obtain drug fusion representation. This enables different modal information to be associated and modeled in the same representation space, thereby improving the ability to represent the relationship between drugs and cells. 3. This invention constructs a unified response prediction model and processes the model input features, enabling the response prediction model to handle different numbers of drug fusion representations and different types of drug response prediction tasks under the same input form, thereby improving the consistency of the model structure. 4. After the response prediction model is trained, the parameters of the response prediction model are adjusted so that the response prediction model can adapt to the feature distribution of the target data, thereby completing drug response prediction when the sample size is small. Attached Figure Description
[0016] Figure 1 This is a flowchart illustrating a drug response prediction method based on multi-source information fusion according to the present invention. Detailed Implementation
[0017] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.
[0018] like Figure 1 The method shown is a drug response prediction method based on multi-source information fusion, the method comprising: S1. Obtain sample data, including single-drug and multi-drug samples. Both single-drug and multi-drug samples must include at least: drug structure information, drug text description information, and cell background feature information. Single-drug samples include the SMILES (Simplified Molecular Input Line Entry System) representation of the drug, the drug text description, the transcriptome expression characteristics of the cell line, and the corresponding IC50 (half maximal inhibitory concentration) label. Multi-drug samples include the SMILES representation of the first drug, the SMILES representation of the second drug, the text descriptions of each drug, the transcriptome expression characteristics of the cell line, and the corresponding co-scoring label. The transcriptome expression characteristics are preferably transcript-level expression profiles to provide a more fine-grained characterization of the cell background.
[0019] S2. Encode drug structural information and drug textual description information to output drug structural feature representation and drug semantic feature representation; construct cell representation based on cell background feature information; by transforming raw data into feature representation, different types of data can be processed in a unified space.
[0020] S3. Based on the drug structure feature representation, guide the fusion of drug semantic feature representation and output the drug fusion representation; guide the drug semantic feature representation through the drug structure feature representation to establish the correlation between different modal information.
[0021] S4. Based on the task type, combine the drug fusion representation with the cell representation to construct model input features. The task type is used to indicate single-drug response prediction tasks and multi-drug synergistic prediction tasks. By uniformly constructing model input features, the same model can handle different drug response prediction tasks.
[0022] S5. Input the model's input features into a unified response prediction model and output the corresponding drug response prediction results. Specifically, when the task type is a single-drug response prediction task, output the single-drug response value; when the task type is a multi-drug synergistic prediction task, output the multi-drug synergistic response value. Establish a mapping relationship between input features and prediction results through the response prediction model.
[0023] Specifically, in step S2, the drug structure information of each sample data is converted into a three-dimensional molecular graph, and a graph coding network is used to extract drug structure features. The three-dimensional molecular graph uses atoms as nodes and chemical bonds as edges, and includes the spatial coordinate information of atoms. In this embodiment, the input data is first preprocessed. For each drug, its drug structure information is read, and a molecular graph is constructed based on the drug structure information. The molecular graph includes node features, edge connections, and spatial location information. Among them, node features include atom categories, charge states, and bond connections, and edge features include bond types and adjacency relationships. Drug structure feature representations are extracted based on the molecular graph. For the drug text description information of each drug, a pre-trained language model is used for word segmentation and encoding to obtain drug semantic feature representations. To ensure consistency between the single-drug response prediction task and the multi-drug collaborative prediction task in the feature space, in the multi-drug collaborative prediction task, two drugs are processed using parameter-shared structure coding and text coding methods respectively to obtain corresponding drug structure feature representations and drug semantic feature representations. Cell background feature information is vectorized to output cell representations. Vectorization enables cell background feature information to participate in model calculations.
[0024] Specifically, in step S3, based on an attention mechanism, the drug structural feature representation is used as conditional information to weight the drug semantic feature representation, outputting a drug fusion representation. In this embodiment, an attention-based fusion method is adopted, using the drug structural feature representation as conditional information to weight the drug semantic feature representation, thereby obtaining the drug fusion representation. Compared to a simple concatenation method, this step can explicitly establish the correspondence between structural information and semantic information, improving the discriminative ability and transferability of drug representation. After obtaining the drug fusion representation, cell background feature information is read, and a cell representation is constructed based on the cell background feature information. The drug fusion representation and the cell representation are jointly used to construct the model input features.
[0025] In this embodiment, when the task type is a single-drug response prediction task, the fusion representation of a single drug, the cell representation, and the task type identifier are combined and input into a unified response prediction model; when the task type is a multi-drug synergistic prediction task, the fusion representation of two drugs, the cell representation, and the task type identifier are combined and input into the response prediction model. The task type identifier is used to distinguish different drug response prediction tasks.
[0026] Specifically, in step S5, the response prediction model is a neural network-based regression model. After obtaining the response prediction model, the parameters of the response prediction model are adjusted based on the target data, and drug response prediction is performed based on the adjusted response prediction model. The response prediction model preferably uses a Transformer-based regression network to learn the nonlinear mapping relationship between drug representation, cell background, and task type information. During training, the regression loss is calculated between the predicted response value output by the model and the true label, and the model parameters are updated based on the training set. During testing, only drug-related information, cell line-related information, and task type identification information are input, and the predicted response value of the test sample is output, without needing to input the true response label.
[0027] In one implementation, the model achieved high correlation and low prediction error in the drug response prediction task, for example, the Pearson correlation coefficient was 0.8963, the Spearman correlation coefficient was 0.8768, and the coefficient of determination R0 was [missing data]. 2 The Pearson correlation coefficient was 0.7969. When the target domain was not adapted, the basic model was less affected by the distribution shift. However, after a small sample adaptation, the Pearson correlation coefficient increased from -0.2176 to 0.4685, and the mean absolute error (MAE) decreased from 2.2089 to 0.6476. This shows that the present invention has good technical effects in both standard data scenarios and real low-resource scenarios.
[0028] A drug response prediction system based on multi-source information fusion, the system comprising: The data acquisition module is used to acquire sample data, which includes single-drug samples and multi-drug samples. The sample data includes at least: drug structure information, drug text description information, and cell background feature information. The encoding module is used to encode drug structural information and drug textual description information respectively to obtain drug structural feature representation and drug semantic feature representation, and to construct cell representation based on cell background feature information; The fusion module is used to guide the fusion of drug semantic feature representation based on drug structural feature representation to obtain drug fusion representation; The input building module is used to combine a single drug fusion representation or a two-drug fusion representation with a cell representation to build model input features based on the task type, where the task type is used to indicate a single-drug response prediction task or a multi-drug synergistic prediction task. The response prediction module is used to input the model input features into a unified response prediction model and output the corresponding drug response prediction results. Specifically, when the task type is a single-drug response prediction task, the single-drug response value is output; when the task type is a multi-drug synergistic prediction task, the multi-drug synergistic response value is output.
[0029] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.
[0030] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in sequences other than those illustrated or described herein.
[0031] The present invention has been further described above with reference to specific embodiments. However, it should be understood that the specific description herein should not be construed as limiting the nature and scope of the present invention. Various modifications made to the above embodiments by those skilled in the art after reading this specification are all within the scope of protection of the present invention.
Claims
1. A drug response prediction method based on multi-source information fusion, characterized in that, The method includes: S1. Obtain sample data, which includes single-drug samples and multi-drug samples, wherein both single-drug samples and multi-drug samples include at least: drug structure information, drug text description information, and cell background feature information; S2. Encode the drug structure information and drug text description information to output drug structure feature representation and drug semantic feature representation; construct cell representation based on the cell background feature information; S3. Based on the drug structural feature representation, guide the fusion of the drug semantic feature representation and output the drug fusion representation; S4. Based on the task type, combine the drug fusion representation with the cell representation to construct model input features, wherein the task type is used to indicate single-drug response prediction tasks and multi-drug synergistic prediction tasks. S5. Input the input features of the model into a unified response prediction model and output the corresponding drug response prediction results. When the task type is a single-drug response prediction task, output the single-drug response value; when the task type is a multi-drug synergistic prediction task, output the multi-drug synergistic response value.
2. The drug response prediction method based on multi-source information fusion according to claim 1, characterized in that, The drug structure information includes simplified molecular linear input specifications, the drug text description information includes textual descriptions of the drug, and the cell background feature information includes transcriptomic expression characteristics of the cell line and the corresponding half-maximal inhibitory concentration.
3. The drug response prediction method based on multi-source information fusion according to claim 1, characterized in that, In step S2, the drug structure information of each sample data is converted into a three-dimensional molecular graph, and the drug structure features are extracted using a graph coding network; The three-dimensional molecular graph uses atoms as nodes, chemical bonds as edges, and includes the spatial coordinate information of atoms.
4. The drug response prediction method based on multi-source information fusion according to claim 1, characterized in that, In step S2, the drug text description information is segmented and encoded using a pre-trained language model to output the semantic feature representation of the drug; in the multi-drug task, multiple drugs are processed using a parameter-shared structural encoder and a text encoder respectively.
5. The drug response prediction method based on multi-source information fusion according to claim 1, characterized in that, In step S3, based on the attention mechanism, the drug structural feature representation is used as conditional information to perform weighted processing on the drug semantic feature representation, and the drug fusion representation is output.
6. The drug response prediction method based on multi-source information fusion according to claim 1, characterized in that, In step S2, the cell background feature information is vectorized to output the cell representation.
7. The drug response prediction method based on multi-source information fusion according to claim 1, characterized in that, In step S4, when the task type is a single-drug response prediction task, the single drug fusion representation is combined with the cell representation and the task type identifier; when the task type is a multi-drug synergistic prediction task, multiple drug fusion representations are combined with the cell representation and the task type identifier.
8. The drug response prediction method based on multi-source information fusion according to claim 1, characterized in that, In step S5, the response prediction model is a regression model based on a neural network; the regression model is used to learn the nonlinear mapping relationship between drug fusion representation, cell background representation and task type identification information.
9. The drug response prediction method based on multi-source information fusion according to claim 1, characterized in that, After obtaining the response prediction model, the parameters of the response prediction model are adjusted based on the target data, and drug response prediction is performed based on the adjusted response prediction model.
10. A drug response prediction system based on multi-source information fusion, the system being used to run the method according to any one of claims 1-9, characterized in that, The system includes: The data acquisition module is used to acquire sample data, which includes single-drug samples and multi-drug samples. The sample data includes at least: drug structure information, drug text description information, and cell background feature information. The encoding module is used to encode the drug structure information and the drug text description information respectively to obtain drug structure feature representation and drug semantic feature representation, and to construct cell representation based on the cell background feature information; The fusion module is used to guide the fusion of the drug semantic feature representation based on the drug structural feature representation to obtain a drug fusion representation; An input building module is used to combine a single drug fusion representation or a two-drug fusion representation with the cell representation according to the task type to build model input features, wherein the task type is used to indicate a single-drug response prediction task or a multi-drug synergistic prediction task. The response prediction module is used to input the model input features into a unified response prediction model and output the corresponding drug response prediction results. Specifically, when the task type is a single-drug response prediction task, the single-drug response value is output; when the task type is a multi-drug synergistic prediction task, the multi-drug synergistic response value is output.