Anticancer drug response prediction method and system based on feature fusion and gating mechanism

Abstract

The invention discloses an anti-cancer drug response prediction method and system based on feature fusion and a gating mechanism, and the method comprises the steps: extracting the initial feature representation of a drug from an extended connectivity fingerprint, an interpretable substructure partition fingerprint and a molecular fingerprint based on drug molecules; the initial feature representation of the cancer cell line is extracted through multi-omics data gene expression, mutation, DNA methylation and pathway enrichment score; based on the initial features of the drug, constructing an embedded representation of the drug; based on the initial features of the cancer cell line, a cancer cell line embedding expression is obtained by adopting a dual-scale and gating fusion mechanism; and finally, performing interactive mapping and fusion on the drug and cell characteristics through a double-projection module, obtaining fusion characterization through global average pooling, and inputting the fusion characterization into a full-connection neural network to predict the half inhibition concentration value of the drug on cancer cells. Experimental results show that the accuracy of anticancer drug response prediction is effectively improved.

Description

A Method and System for Predicting Anticancer Drug Response Based on Feature Fusion and Gating Mechanism Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to a method and system for predicting anticancer drug response based on feature fusion and gating mechanisms. Background Technology

[0002] The goal of precision cancer medicine is to find the most effective treatment for each patient, and predicting cancer drug response is a core challenge in achieving this goal. Because each patient's tumor has unique genomic characteristics, the efficacy of the same drug varies greatly among different patients. Therefore, accurately predicting the sensitivity of drugs to specific cancer cell lines using computational methods is of great significance for new drug development, drug retargeting, and guiding personalized clinical treatment.

[0003] Research in the last three years has focused on effectively integrating multi-omics data and capturing the complex nonlinear interactions between drugs and cancer cell lines. For example, the BANDRP model (Wang et al., 2024) proposed a bilinear attention network to integrate the molecular fingerprint of drugs and multi-omics data of cell lines, and automatically learn the interaction information between drugs and cell lines. Furthermore, the PASO model (Wu et al., 2025) integrates multi-omics data and the SMILES representation of drugs, utilizing multi-scale convolutional networks and attention mechanisms to predict the sensitivity of cell lines to drugs. The MultiDRP model (Zhang et al., 2024) combines multi-omics features with drug characterization, effectively improving predictive performance through deep representation learning. The DRN-CDR model (Li et al., 2023) designed a feature interaction module based on graph convolution and attention mechanisms, achieving efficient fusion of drug and cell characterization. These studies demonstrate the powerful ability of deep learning to capture complex biological relationships.

[0004] Despite significant progress, these methods still have some shortcomings. On the one hand, existing methods often fail to fully exploit the potential information in multi-source heterogeneous data, or neglect the complex nonlinear interaction mechanisms between drugs and cell lines during the feature fusion stage. On the other hand, some models have limited ability to capture and represent key omics features during cross-modal information integration, easily leading to feature redundancy and information loss, resulting in insufficient multi-omics data fusion and thus weakening the model's generalization performance and prediction accuracy. This invention aims to address these problems through a novel feature extraction and fusion mechanism. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for predicting anticancer drug response based on feature fusion and gating mechanisms, so as to solve the technical problems of insufficient utilization of drug-cancer cell line interaction information, insufficient fusion of multi-omics data, and low prediction accuracy in the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: Firstly, the present invention provides a method for predicting anticancer drug response based on feature fusion and gating mechanisms, comprising: extracting an initial feature representation of the drug based on extended connectivity fingerprints, interpretable substructure partitioning fingerprints, and molecular fingerprints of drug molecules; extracting an initial feature representation of cancer cell lines based on multi-omics data of cancer cell lines, wherein the multi-omics data of cancer cell lines includes: gene expression, mutation, DNA methylation, and pathway enrichment scores; constructing a drug embedding representation based on the initial feature representation of the drug using a combination of a multilayer perceptron and a gated recurrent unit; obtaining a cancer cell line embedding representation based on the initial feature representation of the cancer cell lines using a dual-scale gating fusion mechanism; performing interactive mapping and fusion of the drug embedding representation and the cancer cell line embedding representation through a dual-projection module, and obtaining a fused representation through global average pooling; inputting the fused representation into a fully connected neural network to predict the half-maximal inhibitory concentration (IC50) of the drug on cancer cells. 50 value.

[0007] Optionally, constructing the embedded representation of the drug includes: mapping the extended connectivity fingerprint, interpretable substructure partitioning fingerprint, and molecular fingerprint from the initial features of the drug into 128-dimensional embedding vectors using a multilayer perceptron (MLP), and concatenating them to form a drug pseudo-sequence tensor. The concatenation formula satisfies the following relationship: In the formula, , , These represent the 128-dimensional embedding vectors of the extended connectivity fingerprint, the interpretable substructure partitioning fingerprint, and the molecular fingerprint after MLP mapping, respectively. Concat represents the vector concatenation operation, and the result forms the drug pseudo-sequence tensor. ; the drug pseudo-sequence tensor Input a single-layer unidirectional gated recurrent unit (GRU), and recursively calculate the hidden state step by step, satisfying the following relationship: In the formula, express The Middle Input features at each time step, This is the hidden state from the previous time step. The current hidden state is given; GRU is used to capture the temporal dependencies among the three fingerprint features of the drug; by concatenating or pooling the hidden states at all time steps, the overall embedding representation of the drug is obtained, which satisfies the following relationship: In the formula, For drug embedding representation, The connection operation is represented by h1, h2, ..., hT, where h1, h2, ..., hT all represent the hidden states obtained by GRU at each time step.

[0008] Optionally, the method of obtaining the cancer cell line embedding representation using a dual-scale gating fusion mechanism specifically includes: for each omics modality Input features We construct two-level MLP branches in parallel, one with 128 dimensions and the other with 256 dimensions. The two branches are as follows: ; In the formula, For activation function, , It is a 128-dimensional learnable weight matrix. , For a 128-dimensional bias term, , It is a 256-dimensional learnable weight matrix. , The bias term is 256 dimensions; modal features are obtained through residual fusion, where the modal features satisfy the following relationship: In the formula, The residual mapping matrix is... The modal features are obtained through residual fusion; the features of each omics modality are adaptively fused through a gating fusion mechanism, and the gating coefficients are calculated, which satisfy the following relationship: ; In the formula, For the Sigmoid function, , For the gate layer weights and biases, This indicates element-wise multiplication. It is the gating factor. The features are fused after adjustment by gating coefficients; the weighted features of all omics modalities are summed, and the summation satisfies the following relationship: In the formula, u is a weighted feature representation for each omics modality; the summation results of different omics modalities are concatenated to form the cancer cell line embedding representation, which satisfies the following relationship: In the formula, For cancer cell embedding representation, , , and Concatg( represents the weighted sum of gene expression, gene mutation, DNA methylation, and biological pathway characteristics, respectively). ) represents the splicing operation.

[0009] Optionally, the step of interactively mapping and fusing the drug embedding representation and the cancer cell line embedding representation through the dual projection module includes: mapping and fusing the drug embedding representation... Embedded representation of cancer cells Projected onto the same output dimension In the space, its projection satisfies the following relationship: ; In the formula, , , , The projection results of drug features and cell features are shown respectively; mean pooling is performed on the cell feature sequences, and the pooling satisfies the following relationship: In the formula, MeanPool represents average pooling. Represent the cell feature sequence after average pooling; calculate the interaction mapping matrix between the drug and cell features, satisfying the following relationship: In the formula, The interaction is element-wise multiplication, where F represents the result of the interaction between the drug and cellular features. Global average pooling is applied to the interaction mapping matrix, and the pooling satisfies the following relationship: ; In the formula, GAP represents Global Average Pooling, and Dropout is used to prevent overfitting; the final fusion representation is obtained. .

[0010] Optionally, the fully connected neural network includes: a first hidden layer, a second hidden layer, and an output layer; the first hidden layer has a dimension of 512, the second hidden layer has a dimension of 128, the activation function is ReLU, and it is combined with batch normalization (BatchNorm); the output layer is a single-neuron regression layer used to output IC. 50 value.

[0011] Optionally, the mean squared error (MSE) can be used as the loss function during training, and the loss function satisfies the following relationship: In the formula, For the sample size, For the true value, These are the predicted values; the parameters are updated using the Adam optimizer.

[0012] Secondly, this invention also provides an anticancer drug response prediction system based on feature fusion and gating mechanisms, comprising: a feature acquisition module, used to extract initial feature representations of the drug based on extended connectivity fingerprints, interpretable substructure partition fingerprints, and molecular fingerprints of drug molecules, and to extract initial feature representations of cancer cell lines based on multi-omics data of cancer cell lines, including gene expression, mutations, DNA methylation, and pathway enrichment scores; a drug embedding representation module, used to construct an embedding representation of the drug based on the initial features of the drug using a structure combining a multilayer perceptron (MLP) and a gated recurrent unit (GRU); a cancer cell line multi-scale characterization module, used to obtain a cancer cell line embedding representation based on the initial features of the cancer cell line using a dual-scale plus gating fusion mechanism; and a dual-projection fusion prediction module, used to perform interactive mapping and fusion of the drug embedding representation and the cancer cell line embedding representation through the dual-projection module, and obtain a fused representation through global average pooling, and input the fused representation into a fully connected neural network to predict the half-maximal inhibitory concentration (IC50) of the drug on cancer cells. 50 value.

[0013] Optionally, the drug embedding representation module is used to: map the extended connectivity fingerprint, the interpretable substructure partitioning fingerprint, and the molecular fingerprint into 128-dimensional embedding vectors via a multilayer perceptron (MLP), and concatenate them to form a drug pseudo-sequence tensor. In the formula, Concat represents a vector concatenation operation; the drug pseudo-sequence tensor is... Input a single-layer unidirectional gated recurrent unit (GRU), recursively calculate the hidden state step by step, and obtain the overall embedding representation of the drug through splicing or pooling operations. , , and These represent the three fingerprint embedding vectors after MLP mapping.

[0014] Optionally, the cancer cell line multi-scale characterization module is used for: for each omics modality Input features We construct 128-dimensional and 256-dimensional two-layer MLP branches in parallel to compute feature representations; and obtain modal features through residual fusion. ;in, This is the residual mapping matrix, used to adjust the fusion ratio between 128-dimensional and 256-dimensional features; the gating coefficients and weighted features are calculated through a gating fusion mechanism. ; ;in, For the Sigmoid function, , For the gate layer weights and biases, This indicates element-wise multiplication. It is the gating factor. The features are fused after adjustment by gating coefficients; the weighted features of all omics modalities are summed and concatenated to form the cancer cell line embedding representation: In the formula, For cancer cell embedding representation, , , and Concatg( represents the weighted sum of gene expression, gene mutation, DNA methylation, and biological pathway characteristics, respectively). ) represents the splicing operation.

[0015] Optionally, the dual-projection fusion prediction module is used to: project the drug embedding representation and the cancer cell line embedding representation onto the same dimensional space; perform mean pooling on the cell feature sequence; calculate the interaction mapping matrix between the drug and cell features; perform global average pooling on the interaction mapping matrix, and optionally apply Dropout operation to obtain the fused representation; and input the fused representation into a fully connected neural network to predict IC. 50 value.

[0016] The present invention has the following beneficial effects: The anticancer drug response prediction method and system based on feature fusion and gating mechanisms of the present invention fully utilizes drug fingerprint and cellular multi-omics data, and effectively extracts deep features through multi-scale networks and gating recurrent mechanisms. Simultaneously, it employs dual-projection fusion to deeply learn the complex interactions between drugs and cell lines, effectively improving the accuracy of cancer drug response prediction. In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0017] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 is a flowchart illustrating the anticancer drug response prediction method based on feature fusion and gating mechanisms according to a preferred embodiment of the present invention; Figure 2 is a structural schematic diagram of the anticancer drug response prediction system based on feature fusion and gating mechanisms according to a preferred embodiment of the present invention; Figure 3 is a schematic diagram of the half-inhibitory concentration (WIC) values ​​of 169 drugs in an AML cell line according to a preferred embodiment of the present invention. Detailed Implementation

[0018] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings, but the present invention can be implemented in many different ways as defined and covered by the claims.

[0019] Figure 1 is a flowchart and structural diagram of the anticancer drug response prediction method based on feature fusion and gating mechanism as described in this embodiment.

[0020] Referring to Figure 1, the anticancer drug response prediction method based on feature fusion and gating mechanism according to an embodiment of the present invention includes the following steps: S1: Extracting the initial feature representation of the drug based on the extended connectivity fingerprint, interpretable substructure partition fingerprint, and molecular fingerprint of the drug molecule. The initial feature representation of the cancer cell line is extracted through its multi-omics data, including gene expression, mutation, DNA methylation, and pathway enrichment scores; S2: Based on the initial features of the drug, constructing the embedded representation of the drug using a structure combining a multilayer perceptron (MLP) and a gated recurrent unit (GRU); In practice, the drug fingerprint sequence is obtained by concatenating the three fingerprint embedding vectors transformed by the multilayer perceptron (MLP), denoted as:

[0021] in, , , These represent the 128-dimensional embedding vectors of the extended connectivity fingerprint, the interpretable substructure partitioning fingerprint, and the molecular fingerprint after MLP mapping, respectively. Concat represents the vector concatenation operation, and the result forms the drug pseudo-sequence tensor. .

[0022] In implementation, the gated recurrent unit is a single-layer unidirectional GRU, with the input dimension consistent with the drug fingerprint embedding dimension; for pseudo-sequences Each time step vector is used as the input to a gated recurrent unit (GRU), and the calculation is performed recursively step by step:

[0023] in, express The Middle Input features at each time step, This is the hidden state from the previous time step. This represents the current hidden state. GRU is used to capture the temporal dependencies among the three fingerprint features of a drug.

[0024] Finally, the overall embedded representation of the drug is obtained through splicing or pooling operations:

[0025] in, This represents drug embedding.

[0026] S3: Based on the initial features of the cancer cell line, a dual-scale gating fusion mechanism is used to obtain the embedded representation of the cancer cell line. In implementation, the multi-scale branching network constructs a 128-dimensional and 256-dimensional two-layer MLP in parallel for each omics feature, and outputs the result after residual connections. For any omics modality... Input features The two branches are as follows:

[0027] in For activation function, , For learnable weight matrix, , This is a bias term.

[0028]

[0029] The symbols have the same meaning as above, except that the dimension is 256.

[0030] Residual fusion is as follows:

[0031] in This is the residual mapping matrix, used to adjust the fusion ratio of features at different scales.

[0032] In implementation, the gated fusion mechanism adaptively fuses multi-scale features of each omics, and its gate coefficients and fusion rules are as follows:

[0033] in For the Sigmoid function, , For the gate layer weights and biases, This represents element-wise multiplication (Hadamard product).

[0034] Modal weighted summation:

[0035] And spliced ​​together to form an embedded representation of cancer cell lines: .

[0036] S4: A dual-projection module is used to interactively map and fuse drug and cell features, and a fused representation is obtained through global average pooling. This fused representation is then input into a fully connected neural network to predict the half-maximal inhibitory concentration (IC50) of the drug against cancer cells. In practice, the dual-projection layer maps drug and cell features to the same-dimensional space and allows them to interact. Let the batch size be... The drug and cell characteristic sequences are respectively , Define the output dimension The projection is:

[0037] in , , .

[0038] Preferably, to align the drug with cell characteristics, mean pooling is first performed on the cell sequences:

[0039] MeanPool represents average pooling. This is element-wise multiplication.

[0040] Then, global average pooling and Dropout operations are performed:

[0041] GAP stands for Global Average Pooling, and Dropout is used to prevent overfitting.

[0042] In implementation, the fully connected neural network comprises a three-layer structure: hidden layers with dimensions of 512 and 128, using the ReLU activation function and batch normalization (BatchNorm), and an output layer that is a single-neuron regression layer used to output IC. 50 .

[0043] In practice, the mean squared error (MSE) is used as the loss function during model training.

[0044] in For the sample size, For the true value, These are the predicted values. Parameters are updated via the Adam optimizer.

[0045] Example 2 Referring to Figure 2, this embodiment of the invention provides an anticancer drug response prediction system based on feature fusion and gating mechanism, including a feature acquisition module, a drug embedding representation module, a cancer cell line multi-scale characterization module, and a dual-projection fusion prediction module.

[0046] This module extracts drug features from the extended connectivity fingerprint (ECFP), interpretable substructure partition fingerprint (ESPF), and PubChem molecular fingerprint of drug molecules; simultaneously, it extracts multi-omics features of cancer cell lines from gene expression, gene mutation, DNA methylation, and pathway enrichment scores; these features are used as the initial input features for both drugs and cancer cell lines. The construction method is described in Example 1 and will not be repeated here.

[0047] The module inputs the three drug fingerprints into independent two-layer multilayer perceptrons (MLPs) for nonlinear mapping and unified dimensionality transformation to generate drug multimodal feature sequences. Then, a single-layer unidirectional gated recurrent unit (GRU) is used to capture the complex relationships between the fingerprints, and finally, the drug embedding representation is output. The construction method can be referred to in Example 1, and will not be repeated here.

[0048] Two-layer MLP branches with two scales, 128-dimensional and 256-dimensional, were constructed for the initial input features of the cancer cell line. Overfitting was suppressed by residual connections and Dropout regularization, and adaptive weighted fusion of multi-scale features was achieved using a gated fusion mechanism to obtain the embedded representation of the cancer cells. The construction method can be found in Example 1 and will not be repeated here.

[0049] The dual-projection fusion module inputs the drug embedding representation and the cancer cell line embedding representation into the dual-projection layer and the pooling layer. It calculates the drug-cancer cell line interaction mapping matrix through projection and feature alignment, and generates a fused representation of the drug and cancer cell lines through a GlobalAvgPool operation. This fused representation is then input into a three-layer fully connected MLP, outputting the IC50 prediction value of the drug for the cancer cell lines. The model training uses mean squared error loss, and the Adam optimizer determines the optimal hyperparameters through grid search. Its construction method can be found in Example 1, and will not be repeated here.

[0050] Performance Validation: To evaluate the model's performance, this invention uses two comprehensive metrics: RMSE and MAE. The prediction accuracy of the cancer drug response prediction method is evaluated using the following method: First, the dataset is randomly divided into training, validation, and test sets, with 80% of the data used as the training set, 10% as the validation set, and 10% as the test set. Experimental results are shown in Table 1: Table 1: Experimental Results of this Invention's MGDRP Compared with Other Methods on a Benchmark Dataset

[0051] As can be seen from Table 1, the cancer drug response prediction method of the present invention achieves 91.00% RSME and 66.81% MAE. All indicators of the MGBRP model of the present invention are superior to the BANDRP model.

[0052] The predictive accuracy of the cancer drug response prediction method was further evaluated using a case validation method, as follows: Acute myeloid leukemia (AML) is a malignant hematological disease characterized by abnormal proliferation of bone marrow hematopoietic cells, clinically often presenting with strong treatment resistance, high relapse rate, and difficulty in cure. Although some drugs have shown certain efficacy in treatment, improving patient survival still requires exploring new treatment strategies. Based on this, this study used a benchmark dataset to evaluate the response of AML cell lines to new drugs: nine AML cell lines were selected as independent test sets, and the remaining drug response data were used for model training. The model was ultimately used to predict the half-maximal inhibitory concentration (IC50) of the above nine cell lines to 169 drugs. 50 The value is shown in Figure 3.

[0053] By ranking the predicted drug values ​​in AML cell lines by mean, the four drugs with the highest sensitivity were epirubicin, gemcitabine, Telomerase Inhibitor IX, and tanespimycin. A literature search was conducted based on these drugs, and notably, existing experimental studies have confirmed their effectiveness in AML treatment. This evidence supports the ability of MGDRP to accurately predict cancer drug response.

[0054] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting anticancer drug response based on feature fusion and gating mechanisms, characterized in that, include: Initial feature representations of drugs are extracted based on extended connectivity fingerprints of drug molecules, interpretable substructure partitioning fingerprints, and molecular fingerprints. Initial feature representations of cancer cell lines were extracted based on multi-omics data, including gene expression, mutations, DNA methylation, and pathway enrichment scores. Based on the initial feature representations of the drug, a drug embedding representation was constructed using a combination of a multilayer perceptron and a gated recurrent unit. Based on the initial feature representations of the cancer cell lines, a dual-scale gating fusion mechanism was used to obtain the cancer cell line embedding representation. The drug embedding representation and the cancer cell line embedding representation were interactively mapped and fused using a dual-projection module, and a fused representation was obtained through global average pooling. The fused representation was then input into a fully connected neural network to predict the half-maximal inhibitory concentration (IC50) of the drug against cancer cells. 50 value.

2. The method for predicting anticancer drug response based on feature fusion and gating mechanism according to claim 1, characterized in that, The construction of the embedded representation of the drug includes: mapping the extended connectivity fingerprint, interpretable substructure partition fingerprint, and molecular fingerprint in the initial features of the drug into 128-dimensional embedding vectors through a multilayer perceptron (MLP), and concatenating them to form a drug pseudo-sequence tensor. The concatenation formula satisfies the following relationship: In the formula, 、 、 These represent the 128-dimensional embedding vectors of the extended connectivity fingerprint, the interpretable substructure partitioning fingerprint, and the molecular fingerprint after MLP mapping, respectively. Concat represents the vector concatenation operation, and the result forms the drug pseudo-sequence tensor. ; the drug pseudo-sequence tensor Input a single-layer unidirectional gated recurrent unit (GRU), and recursively calculate the hidden state step by step, satisfying the following relationship: In the formula, express The Middle Input features at each time step, This is the hidden state from the previous time step. The current hidden state is given; GRU is used to capture the temporal dependencies among the three fingerprint features of the drug; by concatenating or pooling the hidden states at all time steps, the overall embedding representation of the drug is obtained, which satisfies the following relationship: In the formula, For drug embedding representation, The connection operation is represented by h1, h2, ..., hT, where h1, h2, ..., hT all represent the hidden states obtained by GRU at each time step.

3. The method for predicting anticancer drug response based on feature fusion and gating mechanism according to claim 1, characterized in that, The method of obtaining the cancer cell line embedding representation using a dual-scale gating fusion mechanism specifically includes: for each omics modality Input features We construct two-level MLP branches in parallel, one with 128 dimensions and the other with 256 dimensions. The two branches are as follows: ； In the formula, For activation function, 、 It is a 128-dimensional learnable weight matrix. 、 For a 128-dimensional bias term, 、 It is a 256-dimensional learnable weight matrix. 、 The bias term is 256 dimensions; modal features are obtained through residual fusion, where the modal features satisfy the following relationship: In the formula, The residual mapping matrix is... The modal features are obtained through residual fusion; the features of each omics modality are adaptively fused through a gating fusion mechanism, and the gating coefficients are calculated, which satisfy the following relationship: ； In the formula, For the Sigmoid function, 、 For the gate layer weights and biases, This indicates element-wise multiplication. It is the gating factor. The features are fused after adjustment by gating coefficients; the weighted features of all omics modalities are summed, and the summation satisfies the following relationship: In the formula, u is a weighted feature representation for each omics modality; the summation results of different omics modalities are concatenated to form the cancer cell line embedding representation, which satisfies the following relationship: In the formula, For cancer cell embedding representation, 、 、 and Concatg( represents the weighted sum of gene expression, gene mutation, DNA methylation, and biological pathway characteristics, respectively). ) represents the splicing operation.

4. The method for predicting anticancer drug response based on feature fusion and gating mechanism according to claim 1, characterized in that, The step of interactively mapping and fusing drug embedding representation and cancer cell line embedding representation through a dual projection module includes: mapping and fusing the drug embedding representation... Embedded representation of cancer cells Projected onto the same output dimension In the space, its projection satisfies the following relationship: ； In the formula, 、 ， ， The projection results of drug features and cell features are shown respectively; mean pooling is performed on the cell feature sequences, and the pooling satisfies the following relationship: In the formula, MeanPool represents average pooling. Represent the cell feature sequence after average pooling; calculate the interaction mapping matrix between the drug and cell features, satisfying the following relationship: In the formula, The interaction is element-wise multiplication, where F represents the result of the interaction between the drug and cellular features. Global average pooling is applied to the interaction mapping matrix, and the pooling satisfies the following relationship: ； In the formula, GAP represents Global Average Pooling, and Dropout is used to prevent overfitting; the final fusion representation is obtained. 。 5. The method for predicting anticancer drug response based on feature fusion and gating mechanism according to claim 1, characterized in that, The fully connected neural network includes a first hidden layer, a second hidden layer, and an output layer; the first hidden layer has a dimension of 512, the second hidden layer has a dimension of 128, the activation function is ReLU, and it is combined with batch normalization (BatchNorm); the output layer is a single-neuron regression layer used to output IC. 50 value.

6. The method for predicting anticancer drug response based on feature fusion and gating mechanism according to claim 1, characterized in that, During training, the mean squared error (MSE) is used as the loss function, which satisfies the following relationship: In the formula, For the sample size, For the true value, These are the predicted values; the parameters are updated using the Adam optimizer.

7. A system for predicting the response of anticancer drugs based on feature fusion and gating mechanisms, characterized in that, include: The feature acquisition module is used to extract the initial feature representation of drugs based on extended connectivity fingerprints, interpretable substructure partition fingerprints, and molecular fingerprints of drug molecules, as well as to extract the initial feature representation of cancer cell lines based on multi-omics data of cancer cell lines, including gene expression, mutations, DNA methylation, and pathway enrichment scores. The drug embedding representation module is used to construct an embedding representation of the drug based on the initial features of the drug, using a structure that combines a multilayer perceptron (MLP) and a gated recurrent unit (GRU). A multi-scale characterization module for cancer cell lines is used to obtain the cancer cell line embedding representation based on the initial features of the cancer cell lines using a dual-scale gating fusion mechanism. A dual-projection fusion prediction module is used to interactively map and fuse the drug embedding representation and the cancer cell line embedding representation through a dual-projection module, obtain the fused representation through global average pooling, and input the fused representation into a fully connected neural network to predict the half-maximal inhibitory concentration (IC50) of the drug on cancer cells. 50 value.

8. The anticancer drug response prediction system based on feature fusion and gating mechanism according to claim 7, characterized in that, The drug embedding representation module is used to: map the extended connectivity fingerprint, interpretable substructure partitioning fingerprint, and molecular fingerprint into 128-dimensional embedding vectors via a multilayer perceptron (MLP), and concatenate them to form a drug pseudo-sequence tensor. In the formula, Concat represents a vector concatenation operation; the drug pseudo-sequence tensor is... Input a single-layer unidirectional gated recurrent unit (GRU), recursively calculate the hidden state step by step, and obtain the overall embedding representation of the drug through splicing or pooling operations. ， 、 and These represent the three fingerprint embedding vectors after MLP mapping.

9. The system according to claim 7, characterized in that, The cancer cell line multi-scale characterization module is used for: for each omics modality Input features We construct 128-dimensional and 256-dimensional two-layer MLP branches in parallel to compute feature representations; Modal features are obtained through residual fusion: ;in, This is the residual mapping matrix, used to adjust the fusion ratio between 128-dimensional and 256-dimensional features; the gating coefficients and weighted features are calculated through a gating fusion mechanism. ； ;in, For the Sigmoid function, 、 For the gate layer weights and biases, This indicates element-wise multiplication. It is the gating factor. The features are fused after adjustment by gating coefficients; the weighted features of all omics modalities are summed and concatenated to form the cancer cell line embedding representation: In the formula, For cancer cell embedding representation, 、 、 and Concatg( represents the weighted sum of gene expression, gene mutation, DNA methylation, and biological pathway characteristics, respectively). ) represents the splicing operation.

10. The system according to claim 7, characterized in that, The dual-projection fusion prediction module is used to: project the drug embedding representation and the cancer cell line embedding representation onto the same dimensional space; perform mean pooling on the cell feature sequence; calculate the interaction mapping matrix between the drug and cell features; perform global average pooling on the interaction mapping matrix, and optionally apply Dropout operation to obtain the fused representation; and input the fused representation into a fully connected neural network to predict IC. 50 value.

Images