CN-122025208-A - Mechanism perception-based synergistic anticancer drug combination prediction method and system

CN122025208ACN 122025208 ACN122025208 ACN 122025208ACN-122025208-A

Abstract

The invention provides a mechanism perception based collaborative anticancer drug combination prediction method and system, which belong to the technical field of biological medical treatment, and comprise the steps of extracting node-level feature characterization of a first drug and a second drug through a drug encoder, extracting molecular characterization of a cell line through a cell line encoder, generating attention tensors through a tri-linear cross-modal collaborative strategy based on a tri-linear attention network to explicitly capture collaborative dependency of drug combinations in a specific cell environment, updating and enhancing the features through context fusion, and finally splicing the enhanced features and outputting collaborative scores through a multi-layer perceptron. The invention can directly model the joint interaction relation of the first medicine, the second medicine and the cell line, thereby accurately predicting the synergistic effect of the anticancer medicine combination.

Inventors

LIU JUNTAO
Shang Ziru
Xin Gaojia
ZHU YANHAO

Assignees

山东大学

Dates

Publication Date: 20260512
Application Date: 20260408

Claims (10)

1. The mechanism perception-based synergistic anticancer drug combination prediction method is characterized by comprising the following steps of: Obtaining biological information of a drug combination to be predicted and a targeted cell line, wherein the drug combination comprises a first drug and a second drug; Converting the SMILES character strings of the first medicine and the second medicine into a molecular graph representation, and respectively extracting the characteristics of the molecular structures of the first medicine and the second medicine through a medicine encoder to obtain a first medicine node level characteristic representation and a second medicine node level characteristic representation; Performing feature extraction on the multiple sets of chemical data of the targeted cell line through a cell line encoder to obtain molecular characterization of the cell line; Generating attention tensors related to the first medicine, the second medicine and the target cell line through a tri-linear attention network, and updating and enhancing three characteristic representations through a tri-linear context fusion module by utilizing the attention tensors; Flattening and splicing the enhanced first drug characteristic, the enhanced second drug characteristic and the enhanced molecular characteristic of the cell line into vectors, and outputting predicted synergy scores through the multi-layer perceptron.
2. The mechanism-aware synergistic anticancer drug combination prediction method as claimed in claim 1, wherein the feature extraction of the molecular structure of the target drug by the drug encoder comprises: The method comprises the steps of mapping atomic initial characteristics to an embedding space through a linear layer, carrying out characteristic coding on a molecular graph by adopting a graph convolution network to obtain molecular graph characteristic representation, introducing a transducer encoder into the graph convolution network, taking node embedding as input, modeling interaction among nodes through a self-attention mechanism, not introducing position coding to keep displacement invariance of node ordering, carrying out stabilizing treatment on attention output through residual connection and layer normalization, and carrying out independent nonlinear mapping and dimension transformation on characteristic representation of each position through a feedforward neural network to obtain node level characteristic representation.
3. The method of claim 1, wherein the feature extraction of the plurality of sets of data of the targeted cell line by the cell line encoder comprises: The method comprises the steps of constructing a plurality of groups of chemical feature matrixes by integrating gene expression, gene mutation and copy number variation information based on a preset gene set, processing the plurality of groups of chemical feature matrixes by adopting a one-dimensional convolutional neural network, mapping the linear layer to a unified embedding dimension to obtain feature representation, introducing a pre-layer normalized transducer encoder to model global relations among genes represented by the features, modeling global dependency relations among different features in a sequence by a multi-head self-attention module, and carrying out nonlinear feature transformation by a feedforward neural network to generate molecular representation of a cell line.
4. The method for mechanism-aware based collaborative anticancer drug combination prediction according to claim 1, wherein the generation of the attention tensor comprises: The method comprises the steps of aligning a first drug node level feature representation, a second drug node level feature representation and a molecular representation of a cell line to the same embedded dimension through a full connection layer, constructing a tri-linear attention network among the three features, constructing a tri-linear attention mechanism through Tucker decomposition, decomposing a high-dimensional tensor into a factor matrix and a low-dimensional core tensor, and carrying out normalization processing on the non-normalized interaction tensor through a Softmax function to obtain the attention tensor.
5. The method of mechanism-aware collaborative anticancer drug combination prediction of claim 4, wherein said generating of attention tensors further comprises applying symmetry constraints on input drug characteristics to reflect interactions between the first drug and the second drug.
6. The method for predicting a synergistic anticancer drug combination based on mechanism perception as claimed in claim 1, wherein the updating and enhancing of three feature representations by the attention tensor through the tri-linear context fusion module comprises: The method comprises the steps of carrying out linear projection on a first medicine node level characteristic representation, a second medicine node level characteristic representation and a molecular representation of a cell line, constructing pairwise interactive representations through a broadcasting mechanism to respectively obtain a combined representation of the cell line and a first medicine, a combined representation of the cell line and a second medicine and a combined representation of the first medicine and the second medicine, introducing the attention tensor into a subspace, weighting and polymerizing context information from all the characteristics to obtain a cell environment context matrix, a first medicine context matrix and a second medicine context matrix obtained by polymerization of other two modes, and fusing the context information with the original characteristic representation by adopting residual connection and linear transformation to obtain the enhanced first medicine characteristics, second medicine characteristics and molecular characteristics of the cell line.
7. The method for predicting a synergistic anticancer drug combination based on mechanism perception as claimed in claim 1, wherein each interlayer of the composition structure of the multi-layered perceptron adopts a ReLU activation function, and Dropout layers are introduced at the same time to alleviate the overfitting.
8. A mechanism-aware-based synergistic anticancer drug combination prediction system, comprising: A data acquisition module configured to acquire biological information of a drug combination to be predicted and a targeted cell line, the drug combination comprising a first drug and a second drug; the medicine structure feature construction module is configured to convert SMILES character strings of the first medicine and the second medicine into molecular graph representations, and respectively extract the features of the molecular structures of the first medicine and the second medicine through a medicine encoder to obtain a first medicine node level feature representation and a second medicine node level feature representation; a cell line molecular feature capture module configured to extract, by a cell line encoder, a plurality of sets of chemical data of the targeted cell line, resulting in a molecular characterization of the cell line; the tri-linear cross-modal collaborative processing module is configured to perform characterization processing based on a tri-linear cross-modal collaborative strategy, specifically, generate attention tensors related to the first drug, the second drug and the target cell line through a tri-linear attention network, and update and enhance three characteristic representations through a tri-linear context fusion module by utilizing the attention tensors; And the medicine combination prediction module is configured to flatten and splice the enhanced first medicine characteristic, the enhanced second medicine characteristic and the molecular characteristic of the cell line into vectors, and output predicted synergy scores through the multi-layer perceptron.
9. A computer readable storage medium having stored thereon a program, which when executed by a processor, implements the steps of a mechanism-aware based collaborative anticancer drug combination prediction method according to any of claims 1-7.
10. An electronic device comprising a memory, a processor and a program stored on the memory and executable on the processor, wherein the processor implements the steps of a mechanism-aware collaborative anticancer drug combination prediction method according to any one of claims 1-7 when executing the program.

Description

Mechanism perception-based synergistic anticancer drug combination prediction method and system Technical Field The invention belongs to the technical field of biomedical science, and particularly relates to a coordinated anticancer drug combination prediction method and system based on mechanism perception. Background The statements in this section merely provide background information related to the present disclosure and may not necessarily constitute prior art. In the course of cancer treatment, it is often difficult for a single drug to effectively block multiple pathogenic pathways involved in a tumor, and drug resistance and adverse reactions are easily generated. Combination therapy, which can reduce the risk of drug resistance while enhancing therapeutic effects by acting on multiple molecular targets and key signaling pathways simultaneously, has become one of the important strategies for cancer treatment. Under the background, accurately predicting the synergistic drug combination is not only beneficial to screening potential high-efficiency combinations, but also can obviously improve the efficiency and safety of the design of the combined treatment scheme, and provides scientific basis for individualized accurate treatment. Currently, a large number of computational approaches have emerged for collaborative drug prediction. Existing methods for predicting such synergistic drugs typically model the effects of individual drugs in a particular cell line separately and then post-fuse drug pair representations. The modeling mode essentially ignores the joint interaction relationship among the medicine A, the medicine B and the cell line, and is difficult to describe a synergistic mechanism generated when the two medicines jointly act on a specific cell environment, so that the characterization capability of the model on the synergistic effect is insufficient, and the prediction performance is limited. Disclosure of Invention In order to overcome the defects in the prior art, the invention provides a mechanism perception-based collaborative anticancer drug combination prediction method and system, which can directly model the joint interaction relationship of a first drug, a second drug and a cell line, so as to accurately predict the collaborative effect of anticancer drug combination. To achieve the above object, one or more embodiments of the present invention provide the following technical solutions: The first aspect of the invention provides a mechanism-perception-based synergistic anticancer drug combination prediction method. A mechanism perception-based synergistic anticancer drug combination prediction method comprises the following steps: Obtaining biological information of a drug combination to be predicted and a targeted cell line, wherein the drug combination comprises a first drug and a second drug; Converting the SMILES character strings of the first medicine and the second medicine into a molecular graph representation, and respectively extracting the characteristics of the molecular structures of the first medicine and the second medicine through a medicine encoder to obtain a first medicine node level characteristic representation and a second medicine node level characteristic representation; Performing feature extraction on the multiple sets of chemical data of the targeted cell line through a cell line encoder to obtain molecular characterization of the cell line; Generating attention tensors related to the first medicine, the second medicine and the target cell line through a tri-linear attention network, and updating and enhancing three characteristic representations through a tri-linear context fusion module by utilizing the attention tensors; Flattening and splicing the enhanced first drug characteristic, the enhanced second drug characteristic and the enhanced molecular characteristic of the cell line into vectors, and outputting predicted synergy scores through the multi-layer perceptron. Further, feature extraction of the molecular structure of the target drug by the drug encoder includes: The method comprises the steps of mapping atomic initial characteristics to an embedding space through a linear layer, carrying out characteristic coding on a molecular graph by adopting a graph convolution network to obtain molecular graph characteristic representation, introducing a transducer encoder into the graph convolution network, taking node embedding as input, modeling interaction among nodes through a self-attention mechanism, not introducing position coding to keep displacement invariance of node ordering, carrying out stabilizing treatment on attention output through residual connection and layer normalization, and carrying out independent nonlinear mapping and dimension transformation on characteristic representation of each position through a feedforward neural network to obtain node level characteristic representation. Further, the feature extraction of the multi-set of