MDCDR: predicting cancer drug response via multimodal feature fusion and feature disentanglement Authors: Jinxiong Zhang, Hedong Xu, Chunyan Tang, Dandan Ni, Xuyang Shen Published: 2026 ABSTRACT AbstractDue to the heterogeneity in cancer patients, accurate prediction of cancer drug response is key to achieving personalized medicine. The biological differences between cell lines and tumor tissues, as well as the domain shift caused by diverse drug mechanisms of action, make it difficult for existing methods to effectively capture the complex relationships between multimodal features. In this study, we propose a multimodal feature fusion and feature disentanglement model, named MDCDR. The model effectively integrates various modalities of drugs and cell lines. By employing an interaction-level feature disentanglement framework that integrates multi-granularity contrastive learning and feature masking prediction, we successfully achieve effective synergy between shared common patterns and specific preserved patterns at the drug–cell line interaction level. Experiments show that MDCDR significantly outperforms existing methods on benchmarks and independent clinical test sets, and exhibits good generalization to new cell lines and drugs. Furthermore, case studies on Acute Myeloid Leukemia demonstrate MDCDR’s capacity to elucidate pathway–drug associations, offering biologically interpretable guidance for precision treatment.Keywords: cancer drug response, multimodal feature fusion, feature disentanglement, contrastive learning Introduction Cancer is a complex disease that is difficult to cure, and its incidence is increasing worldwide [1]. Due to heterogeneity among patients, the same anticancer drug often elicits different responses in patients with the same cancer type [2]. Therefore, accurately predicting the drug response of cancer patients is crucial for personalized medicine [3]. However, clinical patient data are limited and difficult to obtain. Advances in high-throughput sequencing technology have enabled the generation of many public large-scale preclinical datasets, such as the Cancer Cell Line Encyclopedia (CCLE) [4] and the Genomics of Drug Sensitivity in Cancer (GDSC) [5]. These datasets provide drug sensitivity data and cell line characteristics. Researchers have employed machine learning methods to integrate these resources for cancer drug response (CDR) prediction [6]. In the CDR prediction task, a core challenge centers on effectively characterizing drug and cell line entities. Traditional unimodal approaches typically represent each drug and cell line using a single feature modality for predictive modeling. tCNNs [7] utilizes simplified molecular input line entry system (SMILES) of drugs and cell line genetic features for drug response prediction. TGDRP [8] constructs a cell line graph based on protein–protein interaction networks and applies graph neural networks (GNNs) to representation learning. The resulting cell line representations are then integrated with drug features to predict drug response. GPDRP [9] predicts drug responses by combining drug molecular graphs with pathway activity scores, thereby enhancing model interpretability. However, unimodal feature representations fail to fully capture the complex inherent properties of drugs or cell lines, which limits the performance of predictive models based on such features. Therefore, integrating multimodal complementary features, such as multi-omics data from various biological perspectives or multiple molecular fingerprint representations, is a key research direction for improving the accuracy of CDR prediction. DeepCDR [10] uses a hybrid graph convolutional network to integrate multi-omics data from cell lines and drug molecular graphs to accurately identify drug responses. GraTransDRP [11] is a drug response prediction model that combines graph transformer networks and graph attention networks (GAT), integrating drug molecular graph features and multi-omics data. BANDRP [12] optimizes drug response prediction models by integrating multi-omics from cancer cell lines and multi-drug molecular fingerprint features. The above-mentioned studies have verified the significant performance gains of multimodal feature fusion in CDR prediction. Previous studies have achieved significant advances in drug response prediction. However, the absence of tumor microenvironmental regulation in cell lines leads to distributional discrepancies in gene expression between cell lines and tumors [16]. Many CDR models often exhibit limited generalization capability when applied to clinical data. To bridge the gap between preclinical and clinical data, recent studies have explored domain generalization techniques. Current approaches predominantly employ feature alignment within representation learning [17]. AITL [13] and PANCDR [15] align the feature distributions between cell lines and tumor samples through adversarial training. While these methods exhibit superior performance in enhancing cross-tissue-type generalization, their optimization process may eliminate specific information that is critically associated with particular cell lines or tumors. An alternative strategy involves explicitly disentangling representations into domain-shared and domain-private components via feature disentanglement [17] to mitigate this issue. CODE-AE [14] employs a self-supervised disentanglement mechanism to separately learn robust domain-shared features and domain-specific features. Nevertheless, existing methods, including CODE-AE, primarily concentrate on representation learning at the cell line level while generally overlooking the intrinsic heterogeneity of drugs themselves. Specifically, different drugs possess distinct mechanisms of action, and when interacting with specific cell lines, these drugs may generate highly specific response patterns that cannot be adequately captured by models that solely rely on cell line features. A comprehensive comparison of representative methods is summarized in Table 1. As observed, while recent works have explored either multimodal fusion or feature disentanglement, none simultaneously achieve comprehensive multimodal integration and interaction-level disentanglement. Table 1Comparative summary of MDCDR and state-of-the-art methods regarding feature modalities including profiles and type, and disentanglement granularity. Method Cell-line feature Drug feature Disentanglement granularity profiles type profiles type tCNNs [7]expCsmilesCN/ADeepCDR [10]exp,mut,methCgraphSN/ATGDRP [8]graphSgraphSN/AGPDRP [9]pathSgraphSN/AGraTransDRP [11]exp,mut,methCgraphSN/ABANDRP [12]exp,mut,meth,pathC+Secfp,espf,pubchemfpCN/AAITL [13]expCN/AN/ACell-levelCODE-AE [14]expCN/AN/ACell-levelPANCDR [15]expCgraphSCell-levelMDCDRexp,path,graphC+Secfp,espf,pubchemfp,graphC+SInteraction-levelC, contextual; S, structural; exp, expression; mut, mutation; meth, methylation. Comparative summary of MDCDR and state-of-the-art methods regarding feature modalities including profiles and type, and disentanglement granularity. C, contextual; S, structural; exp, expression; mut, mutation; meth, methylation. To address these challenges, we propose a drug response prediction model (MDCDR) based on multimodal fusion and feature disentanglement. MDCDR integrates molecular fingerprints and molecular graph structures of drugs, as well as the transcriptome data, pathway activity scores, and gene interaction networks of cell lines. Subsequently, drug and cell line information are fused through a bilinear attention network (BAN) [18] to generate contextual and structural CDR features. We design a feature disentanglement module that explicitly partitions the CDR feature space into shared and private subspaces. To effectively improve feature disentanglement quality, we introduce a multi-granularity contrastive learning strategy. Additionally, we incorporate feature masking prediction [19] to force the private feature encoder to capture task-relevant information. Through this feature disentanglement approach, MDCDR performs feature disentanglement directly at the drug–cell line interaction level, allowing simultaneous capture of common response patterns across domains. Experimental results show that MDCDR outperforms existing baseline methods in all settings. Ablation study further validates the necessity and effectiveness of the feature disentanglement component. Furthermore, MDCDR exhibits strong generalization capabilities when predicting responses to novel cell lines and drugs. Finally, a case study on Acute Myeloid Leukemia demonstrates the model’s interpretability by elucidating pathway–drug associations for precision treatment. Datasets Cancer cell line–drug response data In the binary drug response prediction task, we collect binary responses (sensitive or resistant) for training the classification model. For GDSC, we obtain the filtered values from the supplementary material of Iorio et al. [20] and binarized them. We remove cell lines without transcriptome data and drugs without PubChemId. After screening, we obtain 901 cell lines and 155 drugs, including 13 255 sensitive reactions and 97 068 resistant reactions. For TCGA, patient and clinical annotations are downloaded from the supplementary material provided by Ding et al. [21]. We classify the response measures into sensitive including “complete response” and “partial response” and resistant including “clinically progressive disease” and “stable disease.” After removing drugs without PubChemId, we obtain 1114 patients and 102 drugs, containing a total of 1430 sensitive responses and 812 resistant responses. In the regression drug response prediction task, we obtain the natural logarithm value of from GDSC2 [5] for training the regression model. Similar to the classification task, we obtain 934 cell lines, 184 drugs, and 155 174 response data. Cell line data We download the RNAseq gene expression files of GDSC and TCGA from DeMap [22] and GDC (https://xenabrowser.net/), respectively, and use the log-normalized TPM values as gene expression data. To mitigate the risk of model overfitting due to the high dimensionality of gene expression data [23], we select 711 genes from the COSMIC Cancer Gene Atlas [24]. We calculate pathway activity scores for the gene expression matrix with gene set variation analysis (GSVA) [25], based on the c2 pathway set containing 3917 gene sets from the Molecular Signature Database (MSigDB) [26], and filter out gene sets with fewer than five genes after mapping. Finally, we convert the gene expression matrix into a GSVA score matrix covering 1285 gene sets. Protein–protein interaction networks (PPIs) are obtained from the STRING database v11.5 [27]. Drug data We download SMILES strings for all drugs from PubChem [28]. SMILES strings are converted into molecular graphs and molecular fingerprints using RDKit [29]. The RDKit-derived molecular graph is composed of nodes representing atoms and edges denoting bonds. Similar to the method of Yang et al. [9], we present the features of the atomic node as a 78D one-hot vector. Methods The basic framework of MDCDR is shown in Fig. 1, which includes contextual feature encoder, structural feature encoder, feature disentanglement, and classifier. All symbols in this figure correspond to the definitions in Table 2. (a) Overview of MDCDR. According to feature disentanglement, MDCDR explicitly disentangles contextual feature and structural feature by using dual contextual feature encoders and dual structural feature encoders, respectively. The four contrastive loss functions work together to ensure the integrity of feature disentanglement. Private features are enhanced in terms of robustness and semantic discrimination through feature masking. Both the context branch and the structure branch add auxiliary prediction tasks to improve the task relevance of features. Finally, all features are fused for drug response prediction. (b) Contextual feature extraction module. (c) Structural feature extraction module. Table 2Notation and definitions. Symbol Definition contextual feature contextual shared encoder contextual private encoder contextual shared feature contextual private feature drug contextual feature cell line contextual feature structural feature structural private encoder structural shared encoder structural private feature structural shared feature drug structural feature cell line structural feature Notation and definitions. Feature encoder Drug representation To integrate comprehensive chemical information description, we collect the extended connectivity fingerprint (ECFP) [30], explainable substructure partition fingerprint (ESPF) [31], and PubChem fingerprint (PubChem) [32] for each drug. We design a dedicated feature encoder for each fingerprint to obtain the same latent representation: , , and . These representations are concatenated to form the contextual features of the drug . The molecular graph can intuitively represent the connection relationship between atoms in a molecule and the characteristics of the atoms, which can be expressed as , where and are the drug feature matrix and adjacency matrix, respectively. We use GAT [33] to capture the local topological structural features of molecules. Finally, we obtain the structural features of the drug . Cell line representation For gene expression features, we use a fully connected neural network as the feature encoder. We regard the extracted gene expression features as the contextual features of the cell line. The functions and characteristics of cells are regulated by gene expression, and the gene interaction network determines the dynamic changes of gene expression [34]. Therefore, we construct a gene interaction graph based on the TGSA [8] method. In addition, as a collection of genes that coordinate specific biological functions, the activity score of a biological pathway can effectively capture the high-level functional state of the cell. We then employ a GAT and a multi-layer perceptron (MLP) to learn for the gene interaction graph and for the pathway activity score matrix. By concatenating and , we can obtain the comprehensive feature that represents the structural information of the cell line. Feature fusion We use a BAN [18] to fuse contextual and structural features of cell lines and drugs: and , and , respectively. Specifically, given a cell line comprehensive feature and drug comprehensive feature , where represents the number of features of the cell line and represents the number of features of the drug, we construct a bilinear weight matrix , which is expressed as follows: where and are learnable weight matrices, is a learnable weight vector, is a unit vector, represents the activation function ReLU, and represents element-wise multiplication. Then, we extract different modal features through low-rank bilinear pooling: where is the th element of the fused representation , and on the right side of the equation represents the column index of the matrix. Next, we perform a pooling operation on the fused representation using a sum pooling layer to obtain the final representation F: where represents the pooling layer window size. Finally, we obtain the CDR contextual feature and the CDR structural feature . Cancer drug response feature fusion Contextual features and structural features usually contain different information. Merging contextual features and structural features can provide a more comprehensive description of biological systems and the mechanisms of drug action [35]. The fusion feature is defined as the sum of the contextual feature and the weighted structural feature , where the weight is given by the contextual feature after passing it through the sigmoid function. This allows the fused features to better incorporate the details and characteristics of the structural features while retaining the contextual features. The fusion process is as follows: where denotes the sigmoid function. Feature disentanglement Feature disentanglement explicitly decomposes the feature space to accurately model the commonalities and differences between different tasks or domains. To this end, we construct dual contextual feature encoders and dual structural feature encoders, generating shared features (, ), and private features (, ), respectively. Through contrastive learning, shared features focus on cross-modal invariance, while unique features retain modality-specific characteristics. To further enhance the robustness of the private features, we introduce a feature masking mechanism. Contrastive learning To achieve fine-grained disentanglement of contextual features and structural features, we design a contrastive learning framework that utilizes four jointly optimized loss functions to separate these features. This method constructs the loss function based on InfoNCE [36] and its variant—negative InfoNCE—which are defined as follows: where denotes anchor feature vector, denotes positive sample vector, denotes candidate sample vector. The cross-modal alignment loss employs InfoNCE loss to maximize similarity between shared features of the same entity while minimizing similarity between shared features of different entities, enabling the model to learn consistent cross-modal invariant representations. The inter-modal separation loss utilizes negative InfoNCE loss to minimize similarity between private features of the same entity, while encouraging high similarity between private features across different modalities, thereby enhancing modality-specific characteristics. The intra-modal disentanglement loss, based on negative InfoNCE loss, applies separate constraints to contextual and structural modalities to maximize the representation distance between intra-modal shared and private features, preventing feature redundancy within each modality. The overall contrastive learning loss function is: These four loss functions work together to ensure complete feature disentanglement, both across different modalities and within each individual modality. Feature mask Feature masking [19] randomly masks dimensions of private features and enforces reconstruction. The encoder must learn intrinsic relationships among feature dimensions to achieve accurate reconstruction. This reconstruction process inherently filters random noise and compels the model to capture nonlinear dependencies between features. The feature masking process is defined as follows: the input features are , where is the batch size, and is the feature dimension. Randomly generate a mask matrix , where each sample has positions set to 0 and the rest are set to 1. represents the masked feature. and are the encoder and decoder, respectively, and is the reconstructed feature. We use MSE to calculate the reconstruction loss on the mask location. The loss function only considers the masked locations. Let , that is, set the mask position to 1 and the rest to 0. is the total number of masked elements. By masking the private features of context and structure, we can obtain and . The sum of the two losses is taken as the total loss for feature mask reconstruction. Cancer drug response prediction The fusion feature can be constructed by concatenating contextual shared feature and private feature into , and concatenating structural shared feature and private feature into . In order to enable the feature encoder to extract more robust features, we use auxiliary prediction tasks, namely contextual feature prediction and structural feature prediction, to force the model to learn more meaningful representations in their respective feature spaces. Finally, three independent classifiers are used for CDR prediction, and the binary cross entropy is used as the loss function : where represents the label, sensitive is 1, and resistant is 0. represents the probability value predicted by the model as sensitive. For the regression model, we use the mean squared error as the loss function : where B represents the batch size. Finally, we get three CDR losses: contextual feature loss , structural feature loss , and fusion feature loss . Finally, we get the total loss : where each is a hyperparameter. Results Model training and evaluation To comprehensively evaluate the performance of MDCDR, we conduct extensive comparative experiments with multiple representative baseline models in the current field of drug response prediction. The selected baseline models include TGDRP [8], TGSA [8], GPDRP [9], GraTransDRP [11], PANCDR [15], and BANDRP [12]. In terms of model performance evaluation, the classification task uses AUC, AUPR, Recall, F1 and Accuracy as evaluation indicators; the regression task uses MSE, MAE, , PCC, and SCC for evaluation. The MDCDR model is implemented using Python 3.9 and PyTorch 2.0.1 framework, and all experiments are performed on the Sugon7000A high-performance computer system of the Supercomputing Center of Guangxi University. The network is optimized using the AdamW optimizer [37]. Hyperparameters are determined via grid search on the training set, and detailed search results are provided in Supplementary Table S1. The final configuration for all hyperparameters is listed in Supplementary Table S2. Performance comparison To evaluate classification prediction performance, the models are trained and tested in GDSC and TCGA datasets, and all models adopt a unified data partitioning strategy. Specifically, all 118 921 drug-cell line pairs contained in GDSC dataset are used as the training set. The 2242 drug–patient pairs in the TCGA dataset are randomly divided into five equal subsets with balanced positive and negative class ratios for stratified five-fold cross-validation: in each fold, four subsets (1794 pairs) are used as the test set, and the remaining subset (448 pairs) is equally split into training (224 pairs) and validation (224 pairs) sets. The final performance metric is reported as the average of the cross-validation results, with 95% confidence intervals detailed in Supplementary Table S3. As shown in Table 3, MDCDR outperforms other baseline models in all evaluation metrics, with statistical significance confirmed by paired t-test (Supplementary Table S4). Compared with the second-best performing baseline model, MDCDR improves AUC, AUPR, Recall, F1, and ACC by 4.93%, 2.56%, 46.2%, 36%, and 20.45%, respectively. Table 3Classification performance comparison of MDCDR and baseline methods in TCGA dataset. Model AUC (SD) AUPR (SD) Recall (SD) F1 (SD) ACC (SD) TGDRP0.6590 (0.0637)0.7628 (0.0347)0.5047 (0.2733)0.5553 (0.2171)0.5778 (0.1089)TGSA0.6684 (0.0518)0.7683 (0.0386)0.3051 (0.3053)0.3487 (0.3190)0.5014 (0.1371)GPDPRP0.6012 (0.0136)0.7124 (0.0093)0.2584 (0.0950)0.3761 (0.1019)0.4759 (0.0395)GraTransDRP0.6667 (0.0814)0.7552 (0.0593)0.5849 (0.2953)0.5891 (0.2949)0.6059 (0.1242)PANCDR0.7239 (0.0132)0.8005 (0.0125)0.4335 (0.0786)0.5631 (0.0651)0.5791 (0.0356)BANDRP0.7135 (0.0166)0.7860 (0.0189)0.3356 (0.1775)0.4440 (0.1932)0.5257 (0.0794)MDCDR 0.7596 (0.0099) 0.8210 (0.0092) 0.8551 (0.0308) 0.8012 (0.0033) 0.7298 (0.0070) Bold values indicate the best performance. Classification performance comparison of MDCDR and baseline methods in TCGA dataset. Bold values indicate the best performance. For the regression task, we only use GDSC2 data and adopt five-fold cross-validation for evaluation. Specifically, in each fold, one subset is designated as the test set, while the remaining four subsets are combined and divided into training and validation sets at a 9:1 ratio. The results are shown in Table 4, with 95% confidence intervals detailed in Supplementary Table S5. MDCDR achieves the best performance on all regression metrics, with statistical significance confirmed by paired t-test (Supplementary Table S6). Compared with the second-best model, MDCDR significantly reduces MSE by 6.51% and MAE by 4.18%, while improving by 0.85%, PCC by 0.44%, and SCC by 0.21%. We further evaluate the performance of MDCDR using the dataset provided by BANDRP (Supplementary Table S7). Since TCGA dataset lacks values, we use the trained regression model to predict the LN IC50 of its sensitive and resistant groups. As shown in the Supplementary Fig. S1, the median LN IC50 of the sensitive group is significantly lower than that of the resistant group (-value =.0004). Table 4Regression performance comparison of MDCDR and baseline methods in GDSC2 dataset. Model MSE (SD) MAE (SD) (SD) PCC (SD) SCC (SD) TGDRP0.8373 (0.0198)0.6763 (0.0096)0.8843 (0.0023)0.9404 (0.0012)0.9171 (0.0018)TGSA0.8285 (0.0160)0.6718 (0.0085)0.8855 (0.0019)0.9412 (0.0011)0.9181 (0.0018)GPDPRP0.8546 (0.0064)0.6858 (0.0031)0.8819 (0.0009)0.9393 (0.0003)0.9150 (0.0004)GraTransDRP0.8673 (0.0066)0.6845 (0.0010)0.8801 (0.0014)0.9382 (0.0007)0.9161 (0.0005)PANCDR0.8882 (0.0065)0.6998 (0.0015)0.8773 (0.0010)0.9367 (0.0004)0.9120 (0.0007)BANDRP0.8877 (0.0175)0.6829 (0.0039)0.8855 (0.0010)0.9412 (0.0005)0.9251 (0.0011)MDCDR 0.7746 (0.0114) 0.6437 (0.0041) 0.8930 (0.0013) 0.9453 (0.0007) 0.9232 (0.0010) Bold values indicate the best performance. Regression performance comparison of MDCDR and baseline methods in GDSC2 dataset. Bold values indicate the best performance. To further evaluate model stability, we expand the test set ratio from 30% to 70% in increments. As detailed in Supplementary Table S8, although overall metrics decreased as the training pool shrank, our method consistently achieves higher scores compared to competing models throughout the evaluation. Moreover, we verify the model robustness through multiple independent runs (Supplementary Table S9) and under different random seeds (Supplementary Table S10). Additionally, we compare time overhead and memory usage in Supplementary Table S11. Although MDCDR requires higher computational resources than lightweight baselines, the resulting performance improvements justify this overhead. Inductive capability study To evaluate the generalization ability of MDCDR to new cell lines and drugs, we design Leave-Drug-Out (LDO) and Leave-Cell-Out (LCO) experiments for classification and regression tasks [38]. In the classification task, we select cell lines or drugs from TCGA dataset that are absent in GDSC, and randomly use 80% of them as the test set, with the remaining 20% equally split into the training and validation sets. In the regression task, we select 20% of the cell lines or drugs from GDSC2 as the test set, and the rest are divided into the training and validation sets at a ratio of 9:1. For the selected cell lines or drugs, we remove their corresponding drug response data to ensure that these samples remain unseen during model training. Experiments are repeated with five different random seeds for data partitioning, and averaged results are reported. For the classification task, average performance, statistical significance, and confidence intervals are provided in Supplementary Tables S12–S14. Similarly, results for the regression task are detailed in Supplementary Tables S15–S17. The average results demonstrate that MDCDR outperforms all baseline models across all metrics. Although the performance of all models decreases when predicting unseen drugs and cell lines, MDCDR maintains superior prediction performance compared with baseline methods. Ablation study We conduct ablation experiments in GDSC2 dataset and select the first fold of five-fold cross-validation for evaluation. The results are shown in Table 5. Experimental results demonstrate that removing any component from the feature disentanglement module results in performance degradation, underscoring the critical importance of each component to the module’s overall effectiveness. Notably, removing auxiliary prediction tasks causes the most significant performance loss, even worse than fully removing feature disentanglement, as it deprives the model of guidance to learn drug response-related features, leading the disentangled shared, and private features to be irrelevant to the task requirements of drug response prediction. Meanwhile, the BAN is critical for capturing cross-modal complementary information, and its removal impairs multimodal feature fusion. Table 5Results of ablation study. Model MSE MAE PCC SCC MDCDR 0.7811 0.6463 0.8929 0.9455 0.9219w/o Feature mask0.79120.64950.89150.9452 0.9223 w/o Contrastive learning0.78970.65190.89170.94500.9216w/o Feature disentanglement0.79910.65750.89040.94400.9201w/o Dual encoders0.80450.66110.88840.94290.9194w/o Auxiliary prediction0.82700.66450.88510.94120.9203w/o BAN0.79850.65930.89050.94370.9188Bold values indicate the best performance. Results of ablation study. Bold values indicate the best performance. Disentanglement effectiveness analysis To validate the effectiveness of the MDCDR model in feature disentanglement, we randomly select 1000 samples from the test set of GDSC2, extract the CDR fusion features generated under three conditions, and apply t-SNE to reduce the dimensionality to 2D for visualization. Without contrastive learning (Fig. 2a), when feature disentanglement is removed during training, significant overlap occurs between contextual shared features and contextual private features in the feature space. For MDCDR without auxiliary prediction (Fig. 2b), features are scattered in the space with blurred boundaries between shared and private features, which makes the model unable to effectively capture the task-relevant feature information for cancer drug response prediction. After complete model training (Fig. 2c), shared features from different modalities exhibit tight clustering, confirming the effectiveness of cross-modal alignment contrastive learning in extracting universal response patterns. Conversely, private features occupy distinct subregions, demonstrating that the intra-modal disentanglement and inter-modal separation mechanisms successfully guide private features to encode modality-specific information. Visualization of CDR fusion features generated by MDCDR in three cases: (a) MDCDR without contrastive learning: Overlap emerges between shared and private contextual features. (b) MDCDR without auxiliary prediction: Features are scattered with blurred boundaries between shared and private features. (c) MDCDR: Cross-modal alignment contrastive learning captures the shared information between contextual and structural features. Intra-modal disentanglement and inter-modal separation guarantee the independence of private features. Beyond visual inspection, quantitative disentanglement metrics are summarized in Supplementary Table S18. Low mutual information (MI) and limited linear predictability (Ridge ) confirm effective separation in the linear subspace. Notably, the higher non-linear predictability (Random Forest ) indicates a functional complementarity: shared features establish linear cross-modal alignment, while private features refine non-linear modality-specific details, collectively enhancing representation capacity. Case study Acute myeloid leukemia (AML) represents a heterogeneous malignancy characterized by the abnormal accumulation of myeloid blasts within the bone marrow [39]. Although current therapeutic regimens exist, the high rates of relapse and drug resistance necessitate novel strategic approaches to boost patient prognosis. To evaluate the generalizability of our model within a specific cancer context, we adopt the experimental framework established by Zhao et al. [40]. for this case study. Specifically, we select 10 AML cell lines from GDSC2 as independent test objects, with detailed information provided in Supplementary Table S19. During the training phase, we mask drug response data associated with these cell lines and utilize the remaining data to construct the training set, thereby establishing a leave-cell-line-out prediction scenario. Subsequently, we predict LN IC50 values for 184 compounds against these cell lines, and the detailed prediction matrix appears as a heatmap in Supplementary Fig. S2. Supplementary Fig. S3 shows the top six drugs with the highest and lowest predicted LN IC50 values. LN IC50 is the logarithm of the half-maximalinhibitory concentration (IC50) of the drug. The lower the value, the more effective the drug is in inhibiting cell growth at a lower concentration. Among these, Vinblastine [41], Bortezomib [42], Docetaxel [43], and SN-38 [44] are predicted as sensitive, aligning with prior experimental evidence of their efficacy in AML. Conversely, EPZ-5676 [45] and Temozolomide [46] show high values, consistent with documented resistance or limited utility. These results underscore the capacity of MDCDR to identify novel therapeutic effects and propose potential treatment options for cancer. Our model exhibits the highest prediction error for the AML cell line GDM-1. This discrepancy is likely attributable to its unique biological lineage. Unlike the typical myeloblastic AML models, GDM-1 is an acute erythroid leukemia (FAB M6) cell line [47]. Compared with myeloblasts, erythroid lineage cells show remarkable pathobiological heterogeneity, a characteristic confirmed by the latest WHO AML classification criteria [48]. Consequently, models trained predominantly on common myeloid or solid tumor profiles may fail to capture the specific pharmacological mechanisms governing erythroid leukemia, leading to reduced accuracy for this subtype [49]. Drugs typically exert their effects by modulating relevant biological pathways rather than targeting individual proteins [50]. To investigate the association between drugs and signaling pathways, we calculate the Pearson correlation coefficient between the pathway activity scores and the predicted drug responses (Supplementary Fig. S4). Analysis shows that most cytotoxic drugs (e.g. taxanes, topoisomerase inhibitors) exhibit a significant negative correlation with the p53 transcriptional regulation and DNA damage checkpoint pathways, indicating that the active state of these pathways acts as a potential marker for drug sensitivity [51]. Notably, the apoptotic execution phase is positively correlated with all tested drugs, implying that high activity of the basal apoptotic machinery may instead be associated with a drug-resistant phenotype [52]. These results demonstrate that our model can effectively identify clinically relevant pathway–drug associations and provide actionable guidance for precision treatment and rational drug selection in AML. It prioritizes effective first-line therapies for new diagnoses and screens potential agents for relapsed cases. Conclusion This study proposes a cancer drug response prediction model, MDCDR, based on multimodal feature fusion and feature disentanglement to address the dual challenges of drug heterogeneity and domain shift between preclinical and clinical data in cancer drug response prediction. MDCDR comprehensively captures the complex patterns of drug–cell line interactions by integrating molecular fingerprints and molecular graphs of drugs with transcriptome data, pathway activity scores, and gene interaction networks, followed by deep feature fusion using a BAN. Within the feature disentanglement framework, the shared feature space effectively extracts common information across modalities through a multi-granularity contrastive learning mechanism, while the private feature space retains specific information that is critical to specific drug–cell line combinations through feature masking prediction mechanism. Conceptually, this shifts the research focus from common cell-line-level alignment in previous works to interaction-level modeling. Systematic evaluations in multiple experimental settings demonstrate that MDCDR outperforms existing baseline methods in predicting clinical drug responses. Ablation experiments confirm the key role of feature disentanglement components, and removing any component leads to a decrease in model performance. Finally, the case study on Acute Myeloid Leukemia validates the model’s interpretability by elucidating pathway–drug associations, offering actionable insights for precision treatment. Although MDCDR has shown promising performance in anticancer drug response prediction, there remains room for improvement. First, our proposed model MDCDR is modeled using only gene expression data. We find that using multi-omics as signatures can more accurately discover key biomarkers associated with the disease and potential therapeutic targets [53]. Therefore, integrating multi-omics data into our model is our future research work. Second, while shared and private features demonstrate functional complementarity, the non-linear predictability indicates some information overlap remains. Future work will explore stronger disentanglement constraints to further increase feature independence. Key Points This study proposes a multimodal feature integration network, named MDCDR, which fuses the multimodal features of drugs and cell lines through a bilinear attention network to comprehensively capture the complex mechanisms of drug–cell line interactions. This study proposes a multimodal feature integration network, named MDCDR, which fuses the multimodal features of drugs and cell lines through a bilinear attention network to comprehensively capture the complex mechanisms of drug–cell line interactions. The proposed network includes a structured feature disentanglement module to separate cross-domain shared features from domain-specific features, and accurately model biological commonalities and differences. The proposed network includes a structured feature disentanglement module to separate cross-domain shared features from domain-specific features, and accurately model biological commonalities and differences. Experiments verify the superiority of the model in cross-domain scenarios, and case studies reveal the key pathway mechanisms. Experiments verify the superiority of the model in cross-domain scenarios, and case studies reveal the key pathway mechanisms. Supplementary Material Contributor Information Jinxiong Zhang, School of Computer, Electronics and Information, Guangxi University, No. 100, Daxue East Road, Xixiangtang District, Nanning 530004, Guangxi, China; Key Laboratory of Parallel, Distributed and Intelligent Computing in Guangxi Universities and Colleges, Guangxi Center of Technology Innovation for Intelligent Digital Services, No. 100, Daxue East Road, Xixiangtang District, Nanning 530004, Guangxi, China. Hedong Xu, School of Computer, Electronics and Information, Guangxi University, No. 100, Daxue East Road, Xixiangtang District, Nanning 530004, Guangxi, China; Key Laboratory of Parallel, Distributed and Intelligent Computing in Guangxi Universities and Colleges, Guangxi Center of Technology Innovation for Intelligent Digital Services, No. 100, Daxue East Road, Xixiangtang District, Nanning 530004, Guangxi, China. Chunyan Tang, School of Computer, Electronics and Information, Guangxi University, No. 100, Daxue East Road, Xixiangtang District, Nanning 530004, Guangxi, China; Key Laboratory of Parallel, Distributed and Intelligent Computing in Guangxi Universities and Colleges, Guangxi Center of Technology Innovation for Intelligent Digital Services, No. 100, Daxue East Road, Xixiangtang District, Nanning 530004, Guangxi, China. Dandan Ni, School of Computer, Electronics and Information, Guangxi University, No. 100, Daxue East Road, Xixiangtang District, Nanning 530004, Guangxi, China; Key Laboratory of Parallel, Distributed and Intelligent Computing in Guangxi Universities and Colleges, Guangxi Center of Technology Innovation for Intelligent Digital Services, No. 100, Daxue East Road, Xixiangtang District, Nanning 530004, Guangxi, China. Xuyang Shen, School of Computer, Electronics and Information, Guangxi University, No. 100, Daxue East Road, Xixiangtang District, Nanning 530004, Guangxi, China; Key Laboratory of Parallel, Distributed and Intelligent Computing in Guangxi Universities and Colleges, Guangxi Center of Technology Innovation for Intelligent Digital Services, No. 100, Daxue East Road, Xixiangtang District, Nanning 530004, Guangxi, China. Author contributions H.X. and D.N. conceived the project. J.Z. supervised the research process. H.X. and X.S. implemented the model and conducted the experiments. H.X. and C.T. led the drafting and revision of the initial manuscript. All authors reviewed the manuscript, discussed the experimental results, and provided constructive comments to refine the final version. Conflicts of interest None declared. Funding This work is supported by the high-performance computing platform of Guangxi University. Data availability The MDCDR code and dataset are available at https://github.com/xhd666666/MDCDR. References Associated Data Supplementary Materials Data Availability Statement The MDCDR code and dataset are available at https://github.com/xhd666666/MDCDR.