A dual-channel fusion cancer drug response prediction method

Abstract

The application discloses a double-channel fusion cancer drug response prediction method, and belongs to the technical field of biological information.The application integrates multi-omics data, drug molecular structure information and cancer cell line drug response data from databases such as CCLE, GDSC and PubChem, and forms unified input features after standardization and feature construction.In the first channel, a hierarchical double-attention conversion network is introduced to learn the representation of multi-modal features of drugs and cell lines.In the second channel, a high-order attention neighbor interaction graph convolution network is constructed to capture the graph structure information of cancer drug response.The output results of the two channels are adaptively weighted by a fusion module based on PPO, so that a dynamic optimal decision is realized.Finally, a classification predictor is used to generate the sensitivity prediction result of the cancer cell line to the drug.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics technology, specifically relating to a dual-channel fusion method for predicting cancer drug response. Background Technology

[0002] Cancer is a complex and heterogeneous disease that has become a leading cause of death and a serious threat to human health. A recent study has shown that each cancer patient has a unique genomic profile, making it impossible for a single treatment method to be effective for every type of cancer. Therefore, predicting drug responses based on each patient's genomic profile and providing personalized treatment plans has significant clinical implications.

[0003] Unlike other applications of machine learning, obtaining real-world clinical data on the response of the same patient to multiple anticancer drugs is challenging. Therefore, datasets for drug response prediction primarily rely on preclinical models such as cell lines. In recent years, studies such as the Cancer Cell Line Encyclopedia (CCLE) and Genomics of Drug Sensitivity in Cancer (GDSC) have yielded a wealth of genomic data, including mutation, copy number, and methylation data, as well as sensitivity and resistance data for hundreds of drugs in different cancer cell lines. By utilizing this valuable data, researchers can explore the mechanisms of drug response in cancer treatment and develop machine learning methods to predict cancer drug response (CDR).

[0004] With the development of precision medicine for cancer, how to efficiently and accurately predict cancer drug response has become an important direction for interdisciplinary research in bioinformatics and artificial intelligence. Early studies were mostly based on traditional machine learning methods, using the similarity between drugs and cell lines for modeling. For example, Guan et al. proposed a matrix factorization-based method to construct a potential response matrix using the similarity between drugs and cell lines; Liu et al. modeled drug response prediction as a recommendation problem, using an integrated multi-source similarity network for inference; Wang et al. used support vector machines to build a prediction model between cell lines and drugs; and Wan et al. used random forests to model multi-omics data. The advantages of these methods are their simplicity and strong interpretability, but they face problems such as high-dimensional noise and weak multimodal data fusion capabilities, making it difficult to effectively model the complex nonlinear relationship between drugs and cancer cells.

[0005] The rise of deep learning methods has significantly improved model expressiveness and nonlinear modeling capabilities. Liu et al. proposed the DeepCDR model, which uses graph neural networks to model drug molecule graphs and integrates multi-omics feature representations of cancer cells to better reflect drug-cell line interaction patterns. Chawla et al.'s model emphasizes the importance of gene expression in pathway activity estimation, while Jiang et al.'s DeepTTA model uses a Transformer structure to model drugs, further enhancing representation capabilities. Subsequently, models such as GraphDRP utilize GNNs to model drug molecule graphs and combine them with CNNs to extract cell line omics features, demonstrating superior performance compared to traditional methods. GraphCDR introduces contrastive learning techniques, constructing positive and negative sample pairs to improve model discriminative ability and enhance its robustness and generalization under conditions of sample scarcity.

[0006] In recent years, with the development of graph neural networks and multi-view learning techniques, cancer drug response prediction models have achieved significant breakthroughs in structural modeling and information fusion. TGSA introduces protein-protein interaction networks to model inter-cell line relationships, utilizes a dual GNN structure for parallel modeling of drugs and cell lines, and introduces a similarity enhancement strategy to improve performance in cold-start scenarios. SubCDR starts from the substructures or subgenes of drugs and cell lines, constructs interaction graphs to achieve finer-grained response prediction, and enhances the interpretability of the model through traceable paths. HMM-GDAN adopts a multi-view learning approach, combining protein interaction networks to construct multiple adjacency views, and fuses information from different sources through a multi-view self-attention mechanism to improve the model's ability to understand heterogeneous structures.

[0007] To enhance cell line representation capabilities, research is increasingly shifting towards multi-omics integration strategies. DeepFusionCDR combines contrastive learning with multi-omics fusion methods to enhance cell line representation at multiple omics levels, and integrates Transformer and molecular graph representations for drugs, making overall feature learning more comprehensive. RedCDR designs a bi-branch structure to model omics attributes and interaction information separately, achieving knowledge transfer and joint optimization through relation distillation, representing a further breakthrough in information flow modeling while also possessing some interpretability. BANDRP integrates multi-omics and drug fingerprint information, proposing a bilinear attention mechanism to achieve multi-layer information interaction and fusion, and combining it with MLP to improve expressive power, demonstrating new progress in fusion accuracy and model expressiveness. TransCDR introduces transfer learning techniques, utilizes pre-trained encoders such as ChemBERTa and GIN to improve the generalization of drug representation, and designs a multimodal fusion mechanism to enhance the collaborative modeling ability between cell lines and drugs.

[0008] While existing cancer drug response prediction methods have improved model performance to some extent, key shortcomings remain. First, most methods only extract deep features within a modality, ignoring potential interaction mechanisms between modalities and failing to fully capture the complex nonlinear synergistic relationship between drugs and cancer cells. Second, in terms of structural modeling, most models remain at the first-order or shallow neighbor modeling stage, failing to effectively utilize potential high-order semantic paths in biological networks and limiting the model's expression depth. Furthermore, current mainstream methods mostly employ a unified static fusion strategy, which cannot adapt to individual differences in the degree of multimodal information dependence among different cancer cells. Summary of the Invention

[0009] This invention aims to address the problems of existing CDR prediction methods, which are limited to intramodal feature extraction and have difficulty capturing the complex nonlinear synergistic relationship between drugs and cancer cells.

[0010] A dual-channel fusion cancer drug response prediction method is proposed, which uses a dual-channel fusion cancer drug response prediction model for prediction. The prediction process of the dual-channel fusion cancer drug response prediction model includes:

[0011] The steps involved in drug and cell line characterization and hierarchical dual attention shifting include:

[0012] S100, Order Represents the set of all m drugs, for each drug Obtain the extended connectivity fingerprint, the interpretable substructure partition fingerprint, and the PubChem fingerprint, represented as follows: , and ;based on , and Obtain the latent representation of the same dimension , and For the three-dimensional structure of drug molecules, the representations of atoms and chemical bonds are updated layer by layer based on graph isomorphic networks and geometrically enhanced graph neural networks, thereby obtaining a geometric graph-level representation. ; , , and The final potential representation of the drug is obtained by splicing together the components. ;

[0013] S200, Order Represents the set of all n cancer cell lines; for each cancer cell line... The feature vectors of gene expression data, DNA methylation data, and genome mutation data are obtained and represented as follows: , and ;against The cumulative sum of random walks for sample j was determined using normal quantile transformation and pathway gene set. Sample j refers to the j-th cancer cell line; based on Determine the pathway enrichment score of sample j After sample standardization, a feature vector is obtained with the pathway enrichment score of sample j as its elements. ;based on , , and Obtain the latent representation of the same dimension , , and The multi-omics potential representation of cancer cells was obtained by splicing them together. ;

[0014] S300, First, drug model Cancer cell modalities Multi-head self-attention mechanisms are executed separately to obtain enhanced representations. and This leads to the corresponding pooling features. ;based on and By utilizing drug-to-cell and cell-to-drug cross-modal attention mechanisms, the outputs of cross-modal attention mechanisms were obtained respectively. and This leads to the corresponding pooling features. ;based on Determine dynamic weights Used for The fusion yields the fused feature representation. ;based on and Obtain the final fusion features Then, a regression prediction of the cancer cell response to the drug is obtained through a predictor. ;

[0015] The construction of the sparse adjacency matrix and the processing steps of the multi-hop graph convolutional network include:

[0016] S400, Determine drug-cancer cell pairs based on cancer cell line-drug response data. Based on drug-cancer cell response Construct edge set The unique identifier for drugs and cancer cell lines and Mapped to an integer index, the resulting integer index is obtained after mapping. and , and respectively serve as row and column indices of the adjacency matrix A, and according to Does an edge generate an adjacency matrix? Based on adjacency matrix The node degree matrix D is obtained and normalized to obtain the final propagation matrix. ;

[0017] S500, in the normalized sparse adjacency matrix Based on this, K-hop graph convolution is performed to capture high-order neighbor features, resulting in the final enhanced node embedding. ;

[0018] S600, for a pair of indexes with prediction These correspond to drug nodes and cancer cell nodes, respectively. The features of the drug nodes are obtained by indexing from H. and cancer cell nodes ,based on Constructing differences and interaction features , For element-wise absolute values, ⊙ indicates concatenation, and ⊙ indicates element-wise multiplication; The regression prediction value of the cancer cell response to the drug is obtained through the decoder. ;

[0019] And the final fusion prediction steps, including:

[0020] Will and The states are concatenated to form the state input s of the fusion controller; a policy network with self-attention encoding is used to map the state s to the normalized weights of the two channels. And perform linear weighting on the two channel outputs to obtain Finally based on The final predicted probability of cancer cells responding to the drug is obtained.

[0021] Furthermore, a geometric representation is obtained. The process includes:

[0022] Based on graph isomorphic networks and geometrically enhanced graph neural networks, and combining the information transfer mechanism of atom-to-atom graphs and bond-to-bond graphs, the representations of atoms and chemical bonds are updated layer by layer:

[0023]

[0024] in, For a multilayer perceptron of the k-th layer; This represents the representation of atom u in the k-th layer. This represents the representation of atom v in the (k-1)th layer; Let be the set of neighbors of atom u; Let uv represent the bond between atoms u and v in the k-th layer;

[0025] Finally, pooling operations are performed on the representations of all atoms and bonds to obtain a geometric representation of the drug molecule:

[0026]

[0027] in, This represents all atoms from layer 0 to layer K. To splice; Perform global average pooling on the concatenated node representation.

[0028] Furthermore, targeting The cumulative sum of random walks for sample j was determined using normal quantile transformation and pathway gene set. The process includes:

[0029] The intersection of the pathway gene set provided by MSigDB and the gene set contained in the cancer cell line expression matrix is ​​taken, and pathways with values ​​below a threshold are filtered out to obtain a background pathway set of size P. , Represents the pathways in the background pathway set. ;against Perform a normal quantile transformation to obtain the normal quantile transformation result. ,Will The gene sequence is obtained by arranging the gene sequences from largest to smallest. G represents the total number of expressed genes; based on the gene sequence... Does it belong to The cumulative sum of random walks for sample j corresponding to the cumulative number of steps taken to obtain the k-th gene in the sorted sequence. .

[0030] Furthermore, the cumulative sum of random walks for sample j corresponding to the cumulative step count at the k-th gene in the sorted sequence. The process includes:

[0031] Based on genes in gene sequencing sequences Does it belong to Determine whether it belongs to an arbitrary path Binary indications:

[0032]

[0033] Constructing a Kolmogorov–Smirnov type random walk cumulative sum:

[0034]

[0035] Where m is the pathway index, representing the m-th pathway; k represents the cumulative number of steps when running to the k-th gene in the sorted sequence; This represents the t-th gene in the gene sequence. Indicates the number of pathway genes.

[0036] Furthermore, the pathway enrichment score of sample j .

[0037] Furthermore, the obtained cross-modal attention mechanism output and as follows:

[0038]

[0039]

[0040] in, This is a multi-head self-attention mechanism. The outputs of the cross-modal attention mechanism represent the drug-to-cell and cell-to-drug interactions, respectively.

[0041] Furthermore, on The fusion yields the fused feature representation. The process includes:

[0042]

[0043]

[0044] in, These are the parameters of two linear transformation layers in a gated network; ⊙ represents element-wise multiplication; This represents the dynamic weights of the drug→cell and cell→drug paths; This represents the feature representation after gating and fusion. and All of these are activation functions.

[0045] Furthermore, based on the adjacency matrix The process of obtaining the node degree matrix D includes:

[0046] For adjacency matrix Perform the following processing:

[0047]

[0048]

[0049] Where ◎ represents an element-by-element selection operation based on a Boolean mask;

[0050] Introducing the identity matrix I yields Next, calculate and normalize the node degree matrix D:

[0051]

[0052] in, This is the final propagation matrix, used for feature aggregation in higher-order graph convolutions; These are the elements on the main diagonal of the node degree matrix D.

[0053] Furthermore, the final enhanced node embedding is obtained. The process includes:

[0054] The representation of the k-th hop graph convolution in the k-hop graph convolution process:

[0055]

[0056] in, , Initial characteristics of drugs and cell nodes This is a pre-constructed identity matrix based on all unique entity nodes; It is the weight matrix of the k-th hop graph convolution. The bias of the k-th hop graph convolution Indicates the dimension of the output feature; It is the representation of each node under the k-th hop graph convolution;

[0057] Introducing attention weights Weighting:

[0058]

[0059] in, Let be the attention weight vector of the k-th hop feature; Let be the attention score of a node to the features of its k-th hop neighbor; Embedded for the final enhanced node; This indicates the inner product operation.

[0060] Furthermore, the dual-channel fusion cancer drug response prediction model is pre-trained and trained using the reinforcement learning PPO algorithm; during the training process using the reinforcement learning PPO algorithm, a binary cross-entropy loss is used. Define instant rewards in negative form y represents the supervision label, which is a binary label, where a value of 1 indicates sensitivity and a value of 0 indicates drug resistance;

[0061] Value Network Used to estimate the expected return of state s under the current policy. , Represents a value network; These are the parameters of the value network; thus, the advantages in the PPO algorithm are obtained. ;

[0062] During the policy update process, the old policy weights are obtained from the current policy network, and the new policy, in the same state s, is forward-propagated to obtain the fused weights of the policy network. And replace objective and entropy regularization optimization with pruning The weights based on the old strategy are: The weights under the new strategy are Obtain the probability ratio in the PPO algorithm And used to calculate the trimming substitution loss. Determining the entropy regularization term based on the weights under the new strategy. And combined with the loss of cutting substitution Determine the final strategy loss ;

[0063] During training, first minimize Only update the policy network parameters Minimize again Only update the value network parameters Subsequently, binary cross-entropy was used. The parameters of the dual-channel fusion cancer drug response prediction model are fed back and updated to enable training of the dual-channel fusion cancer drug response prediction model.

[0064] The beneficial effects of this invention are:

[0065] This invention effectively integrates the multi-omics features of cancer cell lines and the multimodal structural representation of drugs by introducing a dual-channel fusion framework based on reinforcement learning. On the one hand, it achieves joint modeling of intramodal feature consistency and deep intermodal interactions through a hierarchical dual-attention transformation network, enhancing the fusion capability of multi-source heterogeneous data. On the other hand, it fully captures the high-order structural dependencies between drugs and cancer cells using a high-order attention neighbor interaction graph convolutional network, enhancing the expressive power of interaction modeling. This invention can effectively capture the complex nonlinear collaboration between drugs and cancer cells, improving prediction performance. Furthermore, this invention employs a PPO-based reinforcement learning strategy network to dynamically adjust the dual-channel fusion weights, thereby achieving adaptive optimization under different data distributions and task conditions, balancing prediction performance and model interpretability. Experimental results show that this invention outperforms existing methods on various benchmark datasets and under different experimental environments, significantly improving the accuracy and robustness of cancer drug response prediction, and possesses significant clinical application value and promotion potential. Attached Figure Description

[0066] Figure 1 This is a schematic diagram of the overall process of the method of the present invention.

[0067] Figure 2 This is a comparison of the ROC curves of the method of the present invention with other comparative methods under a hot-start experimental environment.

[0068] Figure 3 This is a comparison of ROC curves between the method of this invention and other comparative methods under the cold start experimental environment of cell lines.

[0069] Figure 4 This is a comparison of ROC curves between the method of this invention and other comparative methods under a drug cold start experimental environment.

[0070] Figure 5 This is a comparison chart of the PR curves of the method of the present invention and other comparative methods under a hot-start experimental environment.

[0071] Figure 6 This is a comparison of the PR curves of the method of the present invention with other comparative methods under the cold start experimental environment of cell lines.

[0072] Figure 7 This is a comparison of the PR curves of the method of the present invention with other comparative methods under the experimental environment of drug cold start.

[0073] Figure 8 This is a comparison of lollipop images between the method of this invention and other comparative methods under an independent testing experimental environment. Detailed Implementation

[0074] To address the shortcomings of existing CDR prediction methods, such as limitations in intra-modal feature extraction, difficulty in capturing complex nonlinear synergistic relationships between drugs and cancer cells, reliance on shallow neighbor-based structural modeling, failure to fully utilize higher-order semantic paths in biological networks, and the prevalence of uniform static strategies in multimodal fusion that struggle to adapt to individual differences among cancer cells, this invention proposes a dual-channel fusion cancer drug response prediction method. This method enables dynamic interactive modeling of cross-modal features, adaptively fusing different information sources, and improving the model's accuracy, robustness, and personalized adaptability in drug response prediction. The implementation details are provided below.

[0075] Specific implementation method one: Combining Figure 1 This implementation method is described below.

[0076] This embodiment presents a dual-channel fusion cancer drug response prediction method, comprising a data and preprocessing module, a channel 1 drug and cell line feature construction module, a channel 1 hierarchical dual-attention transformation network construction module, a channel 2 data preparation module, a channel 2 high-order attention neighbor interaction graph convolutional network construction module, a reinforcement learning PPO fusion module, and a classification predictor construction module. Since the policy network that outputs the normalized weights of the two channels during processing is determined based on reinforcement learning, this invention is essentially a reinforcement learning-based dual-channel fusion cancer drug response prediction method.

[0077] The data and preprocessing module specifically performs the following steps:

[0078] Step a1: Obtain multi-omics data of cancer cell lines from the CCLE database, including genomic mutation data (MUT), gene expression data (EXP), and DNA methylation data (METH). Among them, the mutation data covers 697 genes in the COSMIC cancer gene survey, containing 34,673 unique mutation sites. The gene expression data is log normalized to obtain TPM values ​​and represented as a 697-dimensional feature vector. The DNA methylation data is based on the upstream TSS region extracted from the 1 kb bisulfite sequencing results of the promoter, and is finally represented as an 808-dimensional feature vector.

[0079] Step a2: Download the SMILES strings of the drug molecule from the PubChem database as the basic data source for the structural features of the drug molecule, which will be used for subsequent molecular fingerprint extraction and graph structure modeling.

[0080] Step a3: Collect cancer cell line-drug response data from the GDSC database to construct a dataset containing 104,464 responses from 561 cell lines and 222 drugs; binarize the IC50 values ​​based on the maximum screening concentration threshold to obtain 11,479 sensitive responses and 92,985 drug-resistant responses.

[0081] Step a4: Process the drug response activity area values ​​in the CCLE database, use z-score normalization and set a threshold (effective area > 0.8 is sensitive) to obtain an independent test set for independent testing in the experiment; the test set contains 7307 responses from 317 cell lines and 24 drugs, including 1375 sensitive cells and 5932 drug-resistant cells.

[0082] Step a5: Standardize and format the multi-omics data, drug molecule data, and drug response data obtained above to form a unified input dataset required for model training and testing.

[0083] The channel-one drug and cell line feature construction module specifically performs the following steps:

[0084] Step b1: Characterization of drug molecules.

[0085] make Let m represent the set of all m drugs. The string representing the SMILES of a drug molecule. For example, RDKit is used to calculate the extended connectivity fingerprint (ECFP), interpretable substructure partition fingerprint (ESPF), and PubChem fingerprint (PubchemFP) of a drug, thereby enabling the drug to... The three molecular fingerprints are represented as , and ;

[0086] Then, they are fed into different MLPs to obtain potential representations of the same dimension: , and The process can be formalized as follows:

[0087]

[0088]

[0089] in, It is a linear transformation matrix; It is bias; BN is batch normalization. For activation functions; It is the hyperbolic tangent function; and The subscript mlp represents , , .

[0090] Step b2: First, the 3D structure of the drug molecule is inferred and optimized from SMILES using RDKit. Then, the 3D structure of the drug molecule is modeled based on a graph neural network. A graph isomorphism network (GIN) and geometry-enhanced graph neural network (GeoGNN) methods are used, combining the information transfer mechanism of atom-to-atom (A2A) and bond-to-bond (B2B) graphs to update the representation of atoms and chemical bonds layer by layer. Specifically, in GIN, the representation of atom u in the k-th layer (atom u is the u-th atom in the RDKit molecule) is obtained by aggregating neighboring nodes and bond features, and its initial representation... It includes nine categories of discrete atomic characteristics (atomic number, chiral label, total degree, formal charge, total number of hydrogens, number of free radical electrons, hybridization type, aromaticity, and whether it is in a ring), and each bond The original characteristics include bond type, stereochemistry, whether it is conjugated, and bond length. The formula is as follows:

[0091]

[0092] in, For a multilayer perceptron of the k-th layer; This represents the representation of atom u in the k-th layer. This represents the representation of atom v in the (k-1)th layer; Let be the set of neighbors of atom u.

[0093] In GeoGNN, to incorporate the atom-bond-angular geometry of drugs, it is necessary to first define atomic triples. This means that atoms u and v are connected by a bond, and atom w is a neighbor of v, thus forming an angle. That is, with v as the vertex. and Two keys form an angle. The resulting angle feature is used as the edge feature in the B2B graph to update the key representation.

[0094] Initialization key representation At level k, the representation of key uv is updated as follows:

[0095]

[0096] in, and They represent atomic triplet The different bond angle features formed.

[0097] Through a multi-layered iterative aggregation and update process, the alternating fusion of atomic and bond representations is achieved. The final graph-level readout result is based on the alternately updated atomic representations; the corresponding bond graph serves only as an auxiliary pathway. Its representation is not directly used for graph-level readout but plays a role as an edge feature of the atomic graph during each layer's update process, thus forming an alternating coupling mechanism between the atomic and bond graphs. Under this mechanism, after the bond representation is updated, the representation of atomic node u is redefined to reflect the richer structural and geometric information obtained after the coupled update. The formula is as follows:

[0098]

[0099] Finally, pooling operations are performed on the representations of all atoms and bonds to obtain a geometric representation of the drug molecule:

[0100]

[0101] in, This represents all atoms from layer 0 to layer K. To splice; Perform global average pooling on the concatenated node representation.

[0102] Step b3: The final potential representation of the drug It is obtained by concatenating four sub-representations, where D is the dimension of the latent representation. The formula is as follows:

[0103]

[0104] in, This indicates a splicing operation.

[0105] Step b4: Perform multi-omics feature modeling on cancer cell lines. Let Let represent the set of all n cancer cell lines. For example, the feature vectors of the three omics data—gene expression data, DNA methylation data, and genome mutation data—are respectively represented as follows: , and .

[0106] Step b5: Gene set acquisition and marker. Obtain the pathway set from the c2 gene set of the Molecular Signatures Database (MSigDB). Let M represent the total number of pathways selected from the MSigDBc2 gene set. Let the gene expression vector of the j-th cancer cell line be:

[0107]

[0108] Where G represents the total number of expressed genes. This represents the expression vector of gene g in sample j, where sample j is the j-th cancer cell line.

[0109] Cross-filtering and background set. The intersection of the pathway gene set provided by MSigDB and the gene set contained in the sample expression matrix is ​​used to filter out... The pathways are identified, resulting in the background pathway set:

[0110]

[0111] The sample expression matrix consists of the expression levels of G genes and n cancer cell lines, with rows corresponding to genes and columns corresponding to samples. The final size of the background pathway set is denoted as P.

[0112] Kernel density transformation. The expression levels of all genes in sample j are concatenated to form a sample expression vector, and Gaussian kernel smoothing distribution mapping is performed on this vector to obtain continuous quantile mapping. :

[0113]

[0114]

[0115] Where x' is the independent variable, representing the value of gene expression level; t represents the variable in the convolution kernel, used to calculate the value of the Gaussian kernel; This represents the bandwidth parameter of the kernel function, which controls the smoothness.

[0116] The expression of each gene Mapped to Smooth rank on Furthermore, normal quantile transformation can be performed:

[0117]

[0118]

[0119] in, Representing gene expression data The expression of the g-th gene; The cumulative distribution function (CDF) represents the standard normal distribution. Therefore, the transformation means normalizing the distribution to a normal distribution with a mean of 0 and a variance of 1.

[0120] In-sample ranking and random walk statistic (running-sum). Transformed Arrange from largest to smallest to obtain the gene sequence. , Representing gene expression data The corresponding gene in the gene sequence is the t-th gene. .

[0121] For any path Define binary indicators:

[0122]

[0123] Construct the Kolmogorov–Smirnov type random walk cumulative sum (KS running-sum):

[0124]

[0125] Where m is the pathway index, representing the m-th pathway; k represents the cumulative number of steps when running to the k-th gene in the sorted sequence; This represents the t-th gene in the gene sequence. Indicates the number of pathway genes.

[0126] Pathway enrichment scores and normalization. The enrichment score for sample j is defined as:

[0127]

[0128] After in-sample or cross-sample standardization (z-score of all pathway ES values ​​in the same sample), the interval-controlled pathway enrichment score is obtained.

[0129] Construction of the pathway enrichment vector. All background pathway scores are stacked in a fixed order to form the pathway enrichment score feature vector for sample j:

[0130]

[0131] in, Indicates cancer cell lineage A vector of pathway characteristics.

[0132] Step b6: For , and Using different MLPs to compute The latent representations of the same dimension are obtained: , and This process is the same as the MLP process for drug characterization in b1.

[0133] Gene mutation characteristics First, convolution and flattening are performed, then batch normalization and linear mapping are used to obtain the latent representation:

[0134]

[0135] in, This represents two convolutional layers. This represents the output of the flattened convolution. This indicates batch normalization of the flattened vector, where W and b are the weights and biases, respectively.

[0136] Step b7: Multi-omics potential representation of cancer cells It is composed of four types of features:

[0137]

[0138] The channel-one hierarchical dual attention conversion network construction module specifically performs the following steps:

[0139] Step c1: Intramodal self-attention modeling. First, for the drug modality... Cancer cell modalities By executing multi-head self-attention mechanisms separately, local dependencies within the modality are captured, resulting in enhanced representations:

[0140]

[0141]

[0142]

[0143]

[0144]

[0145] in, The four-way feature input represents the drug and cancer cell modalities, where D is the feature dimension; This represents the output after in-modal self-attention; These represent query, key, and value, respectively. This represents the linear transformation matrix corresponding to the k-th attention head; H represents the vector dimension of a single attention head; H represents the number of attention heads. This represents the linear transformation matrix after multiple heads are spliced ​​together.

[0146] Step c2: Bidirectional cross-modal semantic interaction. (via drug) cell and cell A cross-modal attention mechanism for drugs, modeling the semantic-level bidirectional dependency between drugs and cancer cells:

[0147]

[0148]

[0149] in, The output of the cross-modal attention mechanism represents the drug. cell and cell The interaction results of drug.

[0150] Step c3: Gated-guided interaction fusion. Pooling is performed on the intramodal representation and cross-attention results, and a gated network is used to generate dynamic weights for the drug→cell and cell→drug paths:

[0151]

[0152]

[0153] in, The average pooling result for the drug and cancer cell modal representations, with dimension 1. ; The average pooled representation of the output of drug→cell and cell→drug interactions; These are the parameters of two linear transformation layers in a gated network; ⊙ represents element-wise multiplication; This represents the dynamic weights of the drug→cell and cell→drug paths; This represents the feature representation after gating and fusion. and All of these are activation functions.

[0154] Step c4: Residual-guided fusion enhancement. This involves enhancing the fused features by gating them (…). ) and original modal features ( and The summation operation helps enhance the model's feature representation capabilities while preserving the original input modal information. Residual connections can be viewed as an information flow enhancement strategy; by retaining the influence of the original modalities in the fused features, it effectively avoids excessive information loss, thereby improving the expressive power of the final output features.

[0155] The final fused features are obtained through layer normalization. It is represented as:

[0156]

[0157] in, For layer normalization operation; This is the final fusion feature.

[0158] Step c5: Fuse feature regression predictions. Final fused representation. Regression prediction of cancer cell response to drug values ​​obtained through the predictor. .

[0159]

[0160] in, These are the weights and biases corresponding to the first-level linear transformation; The weights and biases are for the second-level linear transformation. To hide the dimension.

[0161] The data preparation module for channel two specifically performs the following steps: Step d1: Construction of the feature matrix. Let Let N represent the initial characteristics of all drug and cell nodes, where N represents the total number of all unique entity (drug and cell) nodes, constructed as an identity matrix:

[0162]

[0163] in, This represents an N-order identity matrix; each row in the matrix represents the feature vector of a drug node or cell node.

[0164] Step d2: Constructing the index mapping. To establish the correspondence between all nodes in the graph and all original entities (drugs or cells), an index dictionary is constructed. This ensures a consistent data structure. The original IDs for drugs originate from the PubChem database, specifically the corresponding ids within that database; the original IDs for cells originate from the DepMap database, specifically the corresponding ids within that database.

[0165] Step d3: Construction of the sparse adjacency matrix. First, drug-cancer cell pairs are determined based on cancer cell line-drug response data. Based on drug-cancer cell response Construct edge set The original ID is then mapped to an integer index, as shown in the following formula:

[0166]

[0167] in, and The unique identifiers corresponding to the drug and the cancer cell line are mapped to integer indices. and , and respectively serve as row and column indices of the adjacency matrix A.

[0168] Then generate the adjacency matrix. :

[0169]

[0170] To construct a symmetric graph structure (undirected graph), perform the following symmetry operation:

[0171]

[0172]

[0173] Where ◎ represents the element-by-element selection operation based on Boolean mask.

[0174] Step d4: Construction of the normalized propagation matrix. To improve the stability of convolution propagation, an identity matrix I is introduced, resulting in:

[0175]

[0176] Next, calculate and normalize the node degree matrix D:

[0177]

[0178] in, This is the final propagation matrix, used for feature aggregation in higher-order graph convolutions; These are the elements on the main diagonal of the node degree matrix D.

[0179] The channel 2 high-order attention neighbor interaction graph convolutional network construction module specifically performs the following steps:

[0180] Step e1: Construction of the multi-hop graph convolutional network. This involves normalizing the sparse adjacency matrix... Based on the initial feature matrix X, a K-hop graph convolution is performed to capture higher-order neighbor features. The update formula for the k-th hop is:

[0181]

[0182] in, , This is the initial node feature matrix; It is the weight matrix of the k-th hop graph convolution. The bias of the k-th hop graph convolution Indicates the dimension of the output feature; It is the representation of each node under the k-th hop graph convolution.

[0183] Step e2: Neighbor interaction enhancement mechanism. To integrate convolution results with different hop counts, attention weights are introduced. Weighting:

[0184]

[0185] in, Let be the attention weight vector of the k-th hop feature; Let be the attention score of a node to the features of its k-th hop neighbor; Embedded for the final enhanced node; This indicates the inner product operation.

[0186] This mechanism adaptively adjusts the importance of different hop count features, thereby improving multi-scale expressive power.

[0187] Step e3: Pairwise interactive prediction. For a pair of indices with a prediction... (Corresponding to drug nodes and cancer cell nodes respectively), drug node features are obtained by indexing from H. and cancer cell nodes :

[0188]

[0189] That is, select the corresponding features in H based on the index of the drug and cancer cells.

[0190] Constructing differences and interaction features :

[0191]

[0192] in, It is the absolute value of each element; This is the activation function.

[0193] Finally, the input to the decoder yields the regression prediction value of the cancer cell response to the drug. :

[0194]

[0195] in, These are the weights and biases corresponding to the first-level linear transformation; The weights and biases are for the second-level linear transformation. To hide the dimension.

[0196] The reinforcement learning PPO fusion module specifically performs the following steps:

[0197] Step f1: Input State Construction. The prediction scalars from Channel 1 and Channel 2 are concatenated to form the state input for the fusion controller.

[0198]

[0199] in, and These are the current prediction outputs for the two channels, respectively.

[0200] Step f2: The policy network computes the fusion weights. This is done using a policy network with self-attention encoding. The mapping state s represents the normalized weights (fusion strategy) of the two channels:

[0201]

[0202] in, ; This indicates that self-attention encoding is performed on s; The policy network is represented by a multi-layer fully connected layer structure. The fusion weights represent the output of the policy network; They are respectively and The fusion weights; This is the set of policy network parameters.

[0203] Step f3: Weighted fusion and probabilistic output. The policy weights are linearly weighted over the two channel outputs:

[0204]

[0205] Finally, the data is fed into the classification predictor building module for prediction. After sigmoid activation, the final predicted probability of cancer cells responding to the drug is obtained:

[0206]

[0207] Step f4: Define the immediate reward. Using binary cross-entropy loss. Define instant rewards in negative form:

[0208]

[0209] Where y represents the supervision label, which is a binary label, where a value of 1 indicates sensitivity and a value of 0 indicates drug resistance.

[0210] Step f5: Value Assessment and Benchmarking. Value Network Estimate the expected return of state s under the current policy:

[0211]

[0212] in The value network is represented by a multi-layer fully connected network. These are the parameters of the value network.

[0213] Step f6: Calculate the advantage function. Subtract the baseline from the immediate reward to obtain the advantage:

[0214]

[0215] Step f7: Policy Update. The old policy weights are obtained from the current policy network (the policy network at the beginning of each training iteration), and are saved by disconnecting gradients (actually, a function is used to save the current old policy weights). The new policy is obtained by repeatedly forward passing through the same state s. And replace objective and entropy regularization optimization with pruning Let the weights under the old strategy be... In fact, it is obtained from the current policy network. The weights under the new strategy are Define an approximate ratio:

[0216]

[0217] And use trimming to replace the loss:

[0218]

[0219]

[0220] Where A is the advantage function, which measures how much better the current action is than the average strategy; This represents the ratio of the probability of the new strategy to the probability of the old strategy. The pruning parameter (usually set to 0.1-0.3) controls the magnitude of the strategy update.

[0221] To encourage exploration and avoid premature convergence to a fixed fusion ratio, entropy regularization is introduced:

[0222]

[0223] in It is the i-th dimension (i=1,2) of the fusion weights output by the policy network, which corresponds to the new policy weights. It is used to encourage exploration and avoid prematurely collapsing the weights to a certain channel, such as [1,0] or [0,1].

[0224] The final strategy loss is:

[0225]

[0226] in, This is the entropy coefficient.

[0227] Step f8: The value network approximates the return with mean squared error:

[0228]

[0229] Step f9: Joint optimization and backpropagation of the main model. In each training epoch, the three types of networks are updated independently with their respective losses and optimizers, without summing the total losses.

[0230] First, follow the steps (F7) to update the policy network in multiple steps: minimize... (Prune and replace + entropy regularization) Only update policy network parameters Then update the value network using step F8: minimize Only update the value network parameters Then update the main model (two-channel prediction model): finally, use binary cross-entropy. Return and update the main model parameters.

[0231] Among them, binary cross-entropy Not with or Instead of simply adding them together, it indirectly influences the policy network or value network updates through r, thereby achieving decoupled collaboration between supervised learning and policy optimization.

[0232] It should be noted that f4-f9 are the steps in the training process: first, calculate the binary cross-entropy loss using the label y → reward r=-L → value network / advantage / pruning replacement loss / entropy regularization → update the policy network and value network respectively, and then update the two-channel prediction model (backpropagation of the binary cross-entropy gradient). No optimizations are performed after f4 during the prediction phase; however, f2 is still necessary, but "without gradients and without updates".

[0233] Supervised learning objectives and training. The classification predictor is trained using binary cross-entropy loss and serves as the source of reward and the main model update objective.

[0234]

[0235] During training, the policy and value network (f7 and f8) are updated first, followed by... The parameters of the merged master model are updated in reverse.

[0236] On the GDSC dataset, the proposed method was tested and evaluated through various experiments, including five-fold cross-validation, hot start, cell line cold start, and drug cold start. Independent tests were also conducted on the CCLE dataset to further verify the generalization performance of the model.

[0237] To verify and evaluate the performance of this application, it was evaluated on a dataset, using the area under the ROC curve (AUC), area under the precision / recall curve (AUPR), accuracy (ACC), F1 score, precision, and Matthews correlation coefficient (MCC) as evaluation metrics. The specific calculation formulas are as follows:

[0238]

[0239]

[0240]

[0241]

[0242] Among them, TP (True Positive) represents cancer cell samples that are actually drug-sensitive but were correctly predicted as sensitive by the model; TN (True Negative) represents cancer cell samples that are actually drug-resistant but were correctly predicted as drug-resistant by the model; FP (False Positive) represents cancer cell samples that are actually drug-resistant but were incorrectly predicted as sensitive by the model; and FN (False Negative) represents cancer cell samples that are actually drug-sensitive but were incorrectly predicted as drug-resistant by the model.

[0243] To verify the robustness and generalization ability of the proposed method under different experimental environments, we conducted comparative experiments on a constructed dataset, systematically comparing the proposed RLDCF method with existing state-of-the-art methods (including GraphCDR, TGSA, DeepTTA, HMM-GDAN, RedCDR, DeepFusionCDR, TransCDR, BANDRP, and SubCDR). The experimental environments included five-fold cross-validation, independent testing, warm start (WS), cell-line cold start (CCS), and drug cold start (DCS), with five different settings. Detailed experimental results are shown in Tables 1-6. Figures 2-8 .

[0244] Table 1. Performance comparison results of five-fold cross-validation on the constructed dataset.

[0245]

[0246] Table 2. Performance comparison results of independent tests on the constructed dataset.

[0247]

[0248] Table 3. Comparison of warm-start performance on the constructed dataset

[0249]

[0250] Table 4. Comparison results of cold start performance of cell lines on the constructed dataset.

[0251]

[0252] Table 5. Comparison results of drug cold start performance on the constructed dataset.

[0253]

[0254] Table 6 Ablation Experiment Results

[0255]

[0256] The results show that the method of the present invention exhibits excellent performance in all evaluation metrics, especially in core metrics such as AUC, AUPR and F1-Score, which significantly outperform other methods, further verifying the robustness and predictive advantages of the present invention in multiple environmental scenarios.

[0257] To further explore the contributions of each module and feature to the overall architecture, this invention conducted ablation experiments. The multi-omics modal features of drugs and cell lines in channel one were removed sequentially; residual connections and intra-modal self-attention mechanisms were removed; neighbor interaction enhancement mechanisms in channel two were removed; reinforcement learning fusion strategies were removed; or only a single-channel structure was retained. The experimental results are detailed in Table 6. The comparison results show that the complete RLDCF outperforms the case of removing any component in all evaluation metrics, demonstrating the necessity and effectiveness of multi-modal features, dual-channel design, neighbor interaction enhancement mechanisms, and reinforcement learning-driven fusion strategies in improving performance, thus further demonstrating the innovation and advancement of this invention.

[0258] The above examples of the present invention are merely illustrative of the computational model and process of the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is impossible to exhaustively list all possible implementations here. Any obvious variations or modifications derived from the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A dual-channel fusion method for predicting cancer drug response, characterized in that, The prediction process of the dual-channel fusion cancer drug response prediction model includes: The steps involved in drug and cell line characterization and hierarchical dual attention shifting include: S100, Order Let m be the set of all m drugs, and let m be the set of all drugs. Obtain the extended connectivity fingerprint, the interpretable substructure partition fingerprint, and the PubChem fingerprint, represented as follows: , and ;based on , and Obtain the latent representation of the same dimension , and For the three-dimensional structure of drug molecules, the representations of atoms and chemical bonds are updated layer by layer based on graph isomorphic networks and geometrically enhanced graph neural networks, thereby obtaining a geometric graph-level representation. ; , , and The final potential representation of the drug is obtained by splicing together the components. ; S200, Order Represents the set of all n cancer cell lines; for each cancer cell line... The feature vectors of gene expression data, DNA methylation data, and genome mutation data are obtained and represented as follows: , and ;against The cumulative sum of random walks for sample j was determined using normal quantile transformation and pathway gene set. Sample j refers to the j-th cancer cell line; based on Determine the pathway enrichment score of sample j After sample standardization, a feature vector is obtained with the pathway enrichment score of sample j as its elements. ;based on , , and Obtain the latent representation of the same dimension , , and The multi-omics potential representation of cancer cells was obtained by splicing them together. ; against The cumulative sum of random walks for sample j was determined using normal quantile transformation and pathway gene set. The process includes: taking the intersection of the pathway gene set provided by MSigDB and the gene set contained in the cancer cell line expression matrix, filtering out pathways with values ​​below a threshold, and obtaining a background pathway set of size P. , Represents the pathways in the background pathway set. ;against Perform a normal quantile transformation to obtain the normal quantile transformation result. ,Will The gene sequence is obtained by arranging the gene sequences from largest to smallest. G represents the total number of expressed genes; based on the gene sequence... Does it belong to The cumulative sum of random walks for sample j corresponding to the cumulative number of steps taken to obtain the k-th gene in the sorted sequence. ; S300, First, drug model Cancer cell modalities Multi-head self-attention mechanisms are executed separately to obtain enhanced representations. and This leads to the corresponding pooling features. ;based on and By utilizing drug-to-cell and cell-to-drug cross-modal attention mechanisms, the outputs of cross-modal attention mechanisms were obtained respectively. and This leads to the corresponding pooling features. ;based on Determine dynamic weights Used for The fusion yields the fused feature representation. ;based on and Obtain the final fusion features Then, a regression prediction of the cancer cell response to the drug is obtained through a predictor. ; The construction of the sparse adjacency matrix and the processing steps of the multi-hop graph convolutional network include: S400, Determine drug-cancer cell pairs based on cancer cell line-drug response data. Based on drug-cancer cell response Construct edge set The unique identifier for drugs and cancer cell lines and Mapped to an integer index, the resulting integer index is obtained after mapping. and , and respectively serve as row and column indices of the adjacency matrix A, and according to Does an edge generate an adjacency matrix? Based on adjacency matrix The node degree matrix D is obtained and normalized to obtain the final propagation matrix. ; S500, in the normalized sparse adjacency matrix Based on this, K-hop graph convolution is performed to capture high-order neighbor features, resulting in the final enhanced node embedding. ; S600, for a pair of indexes with prediction These correspond to drug nodes and cancer cell nodes, respectively. The features of the drug nodes are obtained by indexing from H. and cancer cell nodes ,based on Constructing differences and interaction features , For element-wise absolute values, ⊙ indicates concatenation, and ⊙ indicates element-wise multiplication; The regression prediction value of the cancer cell response to the drug is obtained through the decoder. ; And the final fusion prediction steps, including: Will and The states are concatenated to form the state input s of the fusion controller; a policy network with self-attention encoding is used to map the state s to the normalized weights of the two channels. And perform linear weighting on the two channel outputs to obtain Finally based on The final predicted probability of cancer cells responding to the drug is obtained.

2. The dual-channel fusion cancer drug response prediction method according to claim 1, characterized in that, Obtain geometric graph-level representation The process includes: Based on graph isomorphic networks and geometrically enhanced graph neural networks, and combining the information transfer mechanism of atom-to-atom graphs and bond-to-bond graphs, the representations of atoms and chemical bonds are updated layer by layer: ， in, For a multilayer perceptron of the k-th layer; This represents the representation of atom u in the k-th layer. This represents the representation of atom v in the (k-1)th layer; Let be the set of neighbors of atom u; Let uv represent the bond between atoms u and v in the k-th layer; Finally, pooling operations are performed on the representations of all atoms and bonds to obtain a geometric representation of the drug molecule: ， in, This represents all atoms from layer 0 to layer K. To splice; Perform global average pooling on the concatenated node representation.

3. The dual-channel fusion cancer drug response prediction method according to claim 2, characterized in that, The cumulative sum of random walks for sample j corresponding to the cumulative number of steps taken when sorting the k-th gene in the sequence. The process includes: Based on genes in gene sequencing sequences Does it belong to Determine whether it belongs to an arbitrary path Binary indications: ， Constructing a Kolmogorov–Smirnov type random walk cumulative sum: ， Where m is the pathway index, representing the m-th pathway; k represents the cumulative number of steps when running to the k-th gene in the sorted sequence; This represents the t-th gene in the gene sequence. Indicates the number of pathway genes.

4. The dual-channel fusion cancer drug response prediction method according to claim 3, characterized in that, path enrichment score of sample j .

5. The dual-channel fusion cancer drug response prediction method according to claim 4, characterized in that, The obtained cross-modal attention mechanism output and as follows: ， ， in, This is a multi-head self-attention mechanism. The outputs of the cross-modal attention mechanism represent the drug-to-cell and cell-to-drug interactions, respectively.

6. The dual-channel fusion cancer drug response prediction method according to claim 1, characterized in that, right The fusion yields the fused feature representation. The process includes: ， in, These are the parameters of two linear transformation layers in a gated network; ⊙ represents element-wise multiplication; This represents the dynamic weights of the drug→cell and cell→drug paths; This represents the feature representation after gating and fusion. and All of these are activation functions.

7. The dual-channel fusion cancer drug response prediction method according to claim 1, characterized in that, Based on adjacency matrix The process of obtaining the node degree matrix D includes: For adjacency matrix Perform the following processing: ， ， Where ◎ represents an element-by-element selection operation based on a Boolean mask; Introducing the identity matrix I yields Next, calculate and normalize the node degree matrix D: ， ， in, This is the final propagation matrix, used for feature aggregation in higher-order graph convolutions; These are the elements on the main diagonal of the node degree matrix D.

8. The dual-channel fusion cancer drug response prediction method according to claim 1, characterized in that, The final enhanced node embedding is obtained. The process includes: The representation of the k-th hop graph convolution in the k-hop graph convolution process: ， in, , Initial characteristics of drugs and cellular nodes This is a pre-constructed identity matrix based on all unique entity nodes; It is the weight matrix of the k-th hop graph convolution. The bias of the k-th hop graph convolution Indicates the dimension of the output feature; It is the representation of each node under the k-th hop graph convolution; Introducing attention weights Weighting: ， in, Let be the attention weight vector for the k-th hop feature; Let be the attention score of a node to the features of its k-th hop neighbor; Embedded for the final enhanced node; This indicates the inner product operation.

9. A dual-channel fusion cancer drug response prediction method according to any one of claims 1 to 8, characterized in that, The dual-channel fusion cancer drug response prediction model is pre-trained and trained using the reinforcement learning PPO algorithm; during the training process using the reinforcement learning PPO algorithm, binary cross-entropy loss is employed. Define instant rewards in negative form y represents the supervision label, which is a binary label, where a value of 1 indicates sensitivity and a value of 0 indicates drug resistance; Value Network Used to estimate the expected return of state s under the current policy. , Represents a value network; These are the parameters of the value network; thus, the advantages in the PPO algorithm are obtained. ; During the policy update process, the old policy weights are obtained from the current policy network, and the new policy, in the same state s, is forward-propagated to obtain the fused weights of the policy network. And replace objective and entropy regularization optimization with pruning The weights based on the old strategy are: The weights under the new strategy are Obtain the probability ratio in the PPO algorithm And used to calculate the trimming substitution loss. Determining the entropy regularization term based on the weights under the new strategy. And combined with the loss of cutting substitution Determine the final strategy loss ; During training, first minimize Only update the policy network parameters Minimize again Only update the value network parameters Subsequently, binary cross-entropy was used. The parameters of the dual-channel fusion cancer drug response prediction model are fed back and updated to enable training of the dual-channel fusion cancer drug response prediction model.

Images