Transfer learning with multiomics integration and deep neural networks reveals drug resistance mechanisms in cancer Authors: Semih Alpsoy, Osman Ugur Sezerman Published: 2025 ABSTRACT AbstractDrug resistance remains one of the primary challenges in effective cancer therapy. In this study, we employed a deep neural network (DNN)-based transfer learning (TL) approach to predict drug response and uncover drug resistance mechanisms. We integrated gene expression, somatic mutation, and copy number aberration (CNA) data with drug response profiles using multi-omics integration (MI). We used the Genomics of Drug Sensitivity in Cancer (GDSC) data for training and incorporated drugs with same pathways into the training models. We then evaluated drug response predictions on independent in-vivo PDX Encyclopedia (PDX) and ex-vivo the Cancer Genome Atlas (TCGA) datasets. In addition, we conducted pathway enrichment analyses to elucidate the mechanisms underlying drug resistance for paclitaxel, 5-fluorouracil (5-FU), gemcitabine, and cetuximab. We also applied Fisher’s exact test (FET) to assess potential associations between drug resistance and the presence of mutations or CNAs. Our pan-drug models outperformed other methods based on the area under the precision-recall curve (AUCPR). Our pathway enrichment analyses revealed LDHB-mediated pyruvate metabolism and FYN-mediated focal adhesion might have pivotal roles in paclitaxel resistance, while PINK1-mediated mitophagy might be critical in 5-FU resistance. In addition to transcriptional activation, FET suggested that CNAs in LDHB and PINK1 may also be associated with resistance to paclitaxel and 5-FU, respectively. Furthermore, enrichment results for paclitaxel and cetuximab indicated shared resistance mechanisms between the two drugs. Importantly, our findings are consistent with prior experimental studies, providing literature-based validation of our results. Overall, our DNN-based TL approach achieved strong predictive performance across PDX & TCGA datasets and enrichment analyses provided valuable biological insights into drug resistance mechanisms.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-025-23435-8.Keywords: Multi-omics integration, Transfer learning, Deep neural networks, Drug response prediction, Drug resistance in cancerSubject terms: Systems biology, Oncology, Cancer, Data integration, Data mining, Machine learning Introduction Cancer is a complex and heterogeneous disease characterized by uncontrolled cell growth and division1. Its development is driven by several hallmarks, which include sustaining proliferative signaling, evading growth suppressors, resisting programmed cell death, enabling replicative immortality, promoting angiogenesis, facilitating invasion and metastasis, deregulating cellular energetics, and avoiding immune destruction2. Various treatment approaches such as surgery, radiation therapy, chemotherapy, targeted therapy, immunotherapy, and hormone therapy are individualized according to the type and stage of cancer, underlying molecular and genetic alterations, and the overall health status of patients3,4. However, the success of treatments can be limited due to the development of several drug resistance mechanisms. Increased drug efflux, drug inactivation and detoxification, changes in drug targets, enhanced DNA repair, apoptotic evasion, immune evasion, activation of alternative signaling pathways, changes in the tumor microenvironment (TME), epigenetic changes, metabolic reprogramming, and epithelial-to-mesenchymal transition (EMT) are examples of such diverse and complex resistance mechanisms5,6. In addition to resistance mechanisms, tumor heterogeneity is another factor that might limit success in treatment. It poses significant challenges due to the varying responses to therapies and adaptive changes that drive resistance mechanisms7. There are two types of heterogeneity: inter-tumor and intra-tumor heterogeneity. Inter-tumor heterogeneity encompasses the varying genetic, molecular, and phenotypic changes in tumors across different individuals. Even variations observed in different tumor sites within the same individual are considered inter-tumor heterogeneity8. On the other hand, intra-tumor heterogeneity includes genetic, molecular, and phenotypic differences in different regions within a single tumor. The accumulation of distinct mutations, CNAs, or epigenetic changes in sub-clones within the tumor due to clonal evolution, differences in the TME, and treatment pressures can lead to variability in tumor progression and treatment responses8. Therefore, the variability in individual responses to treatments including cases where patients initially respond but later become non-responsive can be attributed to the combined effects of various drug resistance mechanisms and tumor heterogeneity9. The existence of diverse and complex resistance mechanisms as well as tumor heterogeneity necessitates the need for generalizable and robust computational models10. MI strategies with a TL approach have great potential to be applied for this purpose11. MI enables a comprehensive and holistic understanding of drug resistance mechanisms in cancer by simultaneously examining multiple data layers12. When combined with MI, TL offers the development of large, diverse, generalizable, and robust training models to better understand inter-tumor heterogeneity and resistance mechanisms13. Knowledge of heterogeneity and general resistance mechanisms gained from such trained models can be transferred to external datasets for validation by using TL even if they have a small sample size or are subject to distribution shifts14. TL approaches leveraging the advantages of DNNs have gained significant popularity in recent years toward this goal15. In this study, a TL approach utilizing DNNs with MI was applied to build pan-drug models, i.e., models trained across multiple drugs within the same pathway class rather than restricted to a single drug, to predict the response of various drugs in PDX & TCGA datasets16,17. Multi-omics datasets derived from hundreds of human cancer cell lines and response profiles of various drugs targeting mitosis, cytoskeleton, DNA replication, and EGFR signaling pathways in GDSC were separately trained for this purpose18. Drug response prediction for five drugs using only the PDX datasets and eight drugs using only the TCGA datasets was conducted successfully and demonstrated strong performance, surpassing that of similar methods applied to the same datasets13 Moreover, pathway enrichment analyses provided insights into drug resistance mechanisms following robust drug response predictions using four different strategies for paclitaxel, 5-FU, gemcitabine, and cetuximab, which are shared between PDX & TCGA datasets. FET further revealed that somatic mutations and CNAs can contribute to drug resistance in addition to transcriptional changes. These findings demonstrated the effectiveness of the TL approach with DNNs and MI in predicting drug responses and uncovering key drug resistance mechanisms, thereby paving the way for advancements in personalized and precision medicine approaches to cancer chemotherapy. Materials and methods Datasets GDSC: Gene expression, somatic mutation, CNA, and drug response datasets of human cancer cell lines from the GDSC were used for training. The GDSC datasets were generated using the Affymetrix Human Genome U219 Array for gene expression, whole-exome sequencing (WES) (Agilent SureSelectXT Human All Exon 50 Mb) for somatic mutations, and the Affymetrix SNP6.0 Array for CNA18. The gene expression, somatic mutation, and CNA datasets were downloaded from the GDSC website (https://www.cancerrxgene.org/downloads/bulk_download). Binarized drug response dataset was obtained from the supplementary table of Iorio et al.19. Responses of various drugs targeting only mitosis, cytoskeleton, DNA replication, and EGFR signaling pathways were used. The sample sizes categorized by tissue types with their respective counts and the names of the drugs are presented in Fig. 1A,B. GDSC: Gene expression, somatic mutation, CNA, and drug response datasets of human cancer cell lines from the GDSC were used for training. The GDSC datasets were generated using the Affymetrix Human Genome U219 Array for gene expression, whole-exome sequencing (WES) (Agilent SureSelectXT Human All Exon 50 Mb) for somatic mutations, and the Affymetrix SNP6.0 Array for CNA18. The gene expression, somatic mutation, and CNA datasets were downloaded from the GDSC website (https://www.cancerrxgene.org/downloads/bulk_download). Binarized drug response dataset was obtained from the supplementary table of Iorio et al.19. Responses of various drugs targeting only mitosis, cytoskeleton, DNA replication, and EGFR signaling pathways were used. The sample sizes categorized by tissue types with their respective counts and the names of the drugs are presented in Fig. 1A,B. Overview of samples by tissue types, cohorts, and response profiles across the GDSC, PDX, and TCGA datasets (A) Sample size distribution across tissue types for mitotic, cytoskeleton, DNA replication, and EGFR signaling inhibitors based on drug response profiles in the GDSC. Each panel displays tissue types on the left and their corresponding sample counts on the right. For each inhibitor class, the number of sensitive and resistant samples is indicated along with the percentage of sensitive samples. The total number of samples (N) is provided separately for each inhibitor class in parentheses. (B) Each of the four panels displays the drugs included in the training phase, grouped by their respective inhibitor classes: mitotic, cytoskeletal, DNA replication, and EGFR signaling inhibitors. Bars show the number of GDSC cell lines screened per drug as indicated to the right of each bar, with drug names on the left. Drugs highlighted in red were used for prediction in the PDX & TCGA datasets. The panel titles indicate the number of drugs included in each inhibitor class, and the total number of drug-cell line responses is provided to reflect the dataset size (C) Sample size distribution across tissue types is shown for paclitaxel, 5-FU gemcitabine, and cetuximab, which are shared between PDX & TCGA datasets and have corresponding drug response profiles in the GDSC. The left side of each bar indicates the tissue type, and the right side shows the sample count. For each drug, the number of sensitive and resistant samples is given, along with the percentage of sensitive samples. The total number of samples (N) is provided separately for each drug in parentheses (D-E) Distribution of sensitive, resistant, and percentage of sensitive samples across cohorts for paclitaxel, 5-FU, gemcitabine, and cetuximab in the PDX & TCGA datasets. Venn diagrams illustrate sample classification per cohort, with the total number of samples (N) indicated in parentheses for each drug. To the right of each diagram, the total counts of sensitive and resistant samples, along with the corresponding percentage of sensitive samples, are shown for all samples of each drug. PDX: Gene expression data generated by RNA sequencing (RNA-seq), somatic mutation data obtained through WES, CNA data profiled using Affymetrix SNP6.0 array, and drug response profiles from immunocompromised mouse-derived tumor samples were obtained from the supplementary tables of Gao et al.16. Response profiles of only five drugs (Erlotinib, cetuximab, paclitaxel, 5-FU, and gemcitabine) were used for drug response prediction. PDX: Gene expression data generated by RNA sequencing (RNA-seq), somatic mutation data obtained through WES, CNA data profiled using Affymetrix SNP6.0 array, and drug response profiles from immunocompromised mouse-derived tumor samples were obtained from the supplementary tables of Gao et al.16. Response profiles of only five drugs (Erlotinib, cetuximab, paclitaxel, 5-FU, and gemcitabine) were used for drug response prediction. TCGA: Gene expression (RNA-seq), somatic mutation (WES), and CNA (Affymetrix SNP6.0 array) datasets of human patient samples were downloaded from Firehose Broad GDAC repository (https://gdac.broadinstitute.org/runs/stddata__2016_01_28/)17. Drug response profiles were obtained from the supplementary table of Ding et al.20. Response profiles of cetuximab, paclitaxel, docetaxel, 5-FU, gemcitabine, temozolomide, doxorubicin, and cisplatin were used for drug response prediction. TCGA: Gene expression (RNA-seq), somatic mutation (WES), and CNA (Affymetrix SNP6.0 array) datasets of human patient samples were downloaded from Firehose Broad GDAC repository (https://gdac.broadinstitute.org/runs/stddata__2016_01_28/)17. Drug response profiles were obtained from the supplementary table of Ding et al.20. Response profiles of cetuximab, paclitaxel, docetaxel, 5-FU, gemcitabine, temozolomide, doxorubicin, and cisplatin were used for drug response prediction. Preprocessing steps for gene expression, somatic mutation, CNA, and drug response datasets All preprocessing steps described in the Multi-Omics Late Integration (MOLI) pipeline were followed for the GDSC, PDX, and TCGA datasets13 The PDX expression dataset was converted from Fragments Per Kilobase of transcript per Million mapped reads (FPKM) to Transcripts Per Million (TPM), and log2-transformation was performed. The TCGA expression dataset was log2-transformed after TPM normalization. The GDSC gene expression dataset was Robust Multi-array Average (RMA)-normalized and log2-transformed, similar to the PDX & TCGA datasets. This approach ensured that all expression datasets were comparable. In addition, pairwise homogenization was performed between the datasets by removing batch effects using the sva R package21. For mutation datasets, genes with non-synonymous mutations affecting protein structure were binarized, assigning a value of 1 to genes carrying the mutations and 0 to those without them. For CNA datasets, the copy number of each gene was divided by the ploidy of the copy-neutral state in the sample after processing the genome segmentation files in TCGA, then log2-transformation was performed. In this way, the TCGA dataset was comparable with the gene-level estimates of total copy number data in GDSC & PDX. Insertions and deletions were binarized by assigning a value of 1 to insertions or deletions and 0 to copy-neutral regions. The GDSC drug response dataset was used in its binarized form, as described in Iorio et al.19. In contrast, the PDX & TCGA drug response datasets, which were reported based on RECIST criteria, were categorized by grouping partial and complete responses as sensitive, and progressive and stable diseases as resistant22. No imputation strategy was applied during preprocessing for the PDX & TCGA datasets as missing data did not pose a significant issue in the multi-omics layers. The primary criterion for sample selection was the availability of drug response data. For PDX, samples with drug response annotations were identified based on the supplementary material of Gao et al.16, while for TCGA, sample selection was based on the drug response annotations provided by Ding et al.20. Only samples with available multi-omics data were included in the analysis. If any omics layer was missing, that sample was excluded. This approach was also applied to the GDSC dataset as not all cell lines with drug response measurements have complete multi-omics profiles. Sensitive and resistant labels from Iorio et al.19 were used directly, and only the cell lines with available multi-omics data and drug response information were included in the final datasets. Finding and unionizing differentially expressed genes (DEGs) DEGs between sensitive and resistant samples were identified using microarray gene expression dataset and binary response profiles of each drug within the four pathways (mitosis, cytoskeleton, DNA replication, and EGFR signaling pathways) in GDSC. The limma R package was used for this purpose23. Genes with an adjusted p-value < 0.05 and |log2FC| > 1 were selected as DEGs. Moreover, the DEGs identified for at least one drug within the same pathway were combined into a unionized feature set for each corresponding pathway (Fig. 2). Distribution of DEGs identified for each drug within mitotic, cytoskeleton, DNA replication, and EGFR signaling inhibitors in the GDSC dataset. Bar plots in show DEGs for (A) mitotic inhibitors (B) cytoskeleton inhibitors (C) DNA replication inhibitors, and (D) EGFR signaling inhibitors. Red bars indicate drugs used in both training and prediction, while blue bars represent drugs used only in training. The total number of unionized DEGs for each inhibitor class is indicated on each plot. Drug response prediction using a TL approach with DNNs and MI TL was implemented using the GDSC cell line dataset as the source domain due to its large sample size and comprehensive pharmacogenomic profiling across multiple cancer types. In addition, the availability of multi-omics data across a large number of cell line samples and drug response information for multiple drugs targeting the same pathway enabled the construction of pan-drug datasets in the source domain. The PDX & TCGA datasets were selected as target domains due to their limited pharmacological profiling and smaller sample sizes compared to GDSC. Moreover, drug overlap and the inclusion of biologically relevant cancer types between the source and target datasets provided a consistent framework for assessing TL performance. Feed-forward subnetworks were independently applied to gene expression, somatic mutation, and CNA datasets using a similar architecture and the same hyperparameter ranges of the MOLI TL workflow to enable a fair performance comparison13 To optimize model performance and ensure generalizability, a random search was conducted over a predefined hyperparameter space for 50 iterations, accompanied by 5-fold cross-validation to prevent overfitting. The search space included variations in mini-batch size to balance gradient noise and computational efficiency with different input layer sizes to accommodate network capacity according to feature dimensionality. A diverse set of learning rates was employed to enable both rapid convergence and fine-grained weight updates. The number of training epochs was varied to account for different convergence behaviors across datasets. Moreover, multiple dropout rates were evaluated to prevent overfitting, particularly in data-scarce settings. Weight decay values were also explored to regularize model complexity and enhance generalization. The hyperparameter ranges used in the model training phase are summarized below. mini batch size = [8, 16, 32, 64]. input layer size= [1024, 512, 256, 128, 64, 32, 16]. learning rate = [0.5, 0.1, 0.05, 0.01, 0.001, 0.005, 0.0005, 0.0001, 0.00005, 0.00001]. epoch = [10, 15, 20, 30, 40, 50, 60, 70, 80, 90, 100]. dropout rate = [0.3, 0.4, 0.5, 0.6, 0.7, 0.8]. weight decay = [0.1, 0.01, 0.001, 0.0001]. This study introduces several key differences compared to the original MOLI framework. While MOLI utilizes all genes with more than 5% variation as input features and integrates all three omics datasets (gene expression, somatic mutation, and CNA) into a single representation for drug response prediction, it does not include a fine-tuning step. In contrast, this study uses a smaller set of DEGs as features, which reduces model complexity, potentially improves predictive performance, and yields more biologically meaningful results. Furthermore, MOLI applies TL only in the context of EGFR-targeting drugs by training on GDSC data and predicting responses of erlotinib and cetuximab in PDX dataset. In contrast, this study expands the TL framework by incorporating drugs that target four distinct pathways (Mitosis, cytoskeleton, DNA replication, and EGFR signaling) from the GDSC dataset, and predicting drug responses in both PDX & TCGA datasets. Rather than combining all omics data into a single integrated representation as in MOLI, three separate integration strategies were employed: expression–mutation (EM), expression–CNA (EC), and expression–mutation–CNA (EMC). Combining mutation data with gene expression (EM) allows the model to account for driver alterations that may not be reflected at the transcriptional level. Similarly, integrating CNAs with expression data (EC) helps capture gene dosage effects that influence expression without changing sequence information. The full integration of expression, mutation, and CNA data (EMC) enables the model to leverage both regulatory and structural genomic information. Moreover, unlike the original MOLI workflow, which combines binary cross-entropy loss with triplet loss, this study uses only binary cross-entropy loss for classification. Finally, although MOLI reports the area under the receiver operating characteristic curve (AUC) as the primary evaluation metric despite substantial class imbalances in drug response data, this study adopts AUCPR as the main evaluation metric. AUCPR offers a more realistic assessment of model performance under imbalanced class distributions13. For drugs that belong to the same pathway, the identified DEGs were unionized as gene features to reduce feature size, allowing the training phase to be carried out with less noise and more meaningful features. TL was utilized not only to predict drug responses of five drugs in PDX and eight drugs in TCGA, but also of the four drugs shared between the two datasets (Paclitaxel, 5-FU, gemcitabine, and cetuximab). The total sample sizes and proportions of sensitive samples in GDSC, PDX, and TCGA for these drugs are visualized in Fig. 1C–E. Once the feature representations for all three omics layers were obtained, they were concatenated into a single representation. Only the genes common to all datasets were included during integration. In these representations, columns corresponded to genes, and rows to cell line samples for drugs within the same pathway (Fig. 3). Workflow for constructing EM, EC, and EMC multi-omics representations using the GDSC source dataset. Gene expression, somatic mutation, and CNA data from GDSC were first preprocessed, and only the genes common to all three omics layers were retained. DEGs identified in a unionized manner across drugs targeting the same pathway were then used to filter each omics dataset. Three separate autoencoders for each omics type were trained independently with hyperparameters optimized via random search. The resulting low-dimensional representations were concatenated to generate EM, EC, and EMC multi-omics representations. Drugs sharing the same pathway were used for representation learning with the number of drugs denoted by N. These N drug-specific samples were combined row-wise to form each representation matrix. In these matrices, gene expression features are continuous, whereas somatic mutation, CNA, and drug response features are binary. During the construction of multi-omics representations, hyperparameters for the gene expression, somatic mutation, and CNA auto-encoder layers (AutoencoderE for gene expression, AutoencoderM for somatic mutation, and AutoencoderC for CNA) were selected using random search within predefined ranges. Each omics-specific auto-encoder consisted of a linear layer followed by batch normalization, a ReLU activation function to introduce non-linearity and improve computational efficiency, and dropout for regularization. The model used dropout, weight decay, and batch normalization as regularization techniques; however, early stopping was not applied. The values for these regularization parameters were chosen from the same hyperparameter ranges outlined above. After constructing the multi-omics representations, a classifier was built as the final subnetwork. The classification layer constructed a linear layer with an input size equal to the sum of the hidden layer sizes from the multi-omics representations. This was followed by dropout and a sigmoid activation function, which was applied to enable binary drug response prediction. The Adagrad optimizer, similar to that in MOLI, was applied for model training. Fine-tuning was exclusively performed in the classification layer by freezing the expression, mutation, and CNA auto-encoders. 50 iterations with random search and the same hyperparameter ranges were used in fine-tuning as those employed during the training phase. Model performance was evaluated using AUCPR although AUC was also reported (Fig. 4). The model architecture is also illustrated in Fig. 5, where three separate auto-encoders (AutoencoderE, AutoencoderM, and AutoencoderC) independently transform their respective input features into low-dimensional representations. These representations are then concatenated and passed to a final classifier layer, which outputs drug response predictions in the PDX & TCGA datasets. Workflow for drug response prediction in PDX & TCGA datasets using multi-omics representations. After constructing EM, EC, and EMC multi-omics representations, each representation is passed through a separate final classifier layer for drug response prediction. The predictions are performed on PDX & TCGA datasets following appropriate preprocessing steps. With TL approach, the gene expression, somatic mutation, and CNA encoder layers are frozen, and only the classification layer is fine-tuned on the target PDX & TCGA datasets. This fine-tuning strategy allows for more robust and accurate drug response predictions across both datasets. Detailed model structure diagram including the autoencoders and hyperparameter components used in MI. AutoencoderE, AutoencoderM, and AutoencoderC are used for gene expression, somatic mutation, and CNA data, respectively. For each autoencoder, input size (I), hidden layer size (h), and dropout rate (r) are selected via random search. Each encoder consists of a linear layer followed by batch normalization, a ReLU activation function, and dropout. After concatenating these layers, low-dimensional multi-omics representations (EM, EC, and EMC) are generated. These representations are then passed to a final classifier layer, which includes a linear layer, dropout, and a sigmoid activation function to produce drug response predictions. Pathways with a large number of annotated drugs in the GDSC dataset, specifically mitosis, cytoskeleton, DNA replication, and EGFR signaling inhibitors, were focused on to enable TL. Among the test datasets, only paclitaxel, 5-FU, gemcitabine, and cetuximab were shared between PDX & TCGA with sufficient sample sizes for robust evaluation. Therefore, these four drugs were selected for systematic assessment of model performance and resistance mechanisms using four complementary evaluation strategies. First strategy: The training datasets included sample data from all cancer types and response profiles of all drugs targeting the corresponding pathways (Mitosis, cytoskeleton, DNA replication, and EGFR signaling pathways). First strategy: The training datasets included sample data from all cancer types and response profiles of all drugs targeting the corresponding pathways (Mitosis, cytoskeleton, DNA replication, and EGFR signaling pathways). Second strategy: The training datasets included sample data from all cancer types but excluded the unique features (drug specific DEGs) and response profiles of the predicted drugs targeting the corresponding pathways. Second strategy: The training datasets included sample data from all cancer types but excluded the unique features (drug specific DEGs) and response profiles of the predicted drugs targeting the corresponding pathways. Third strategy: The training datasets excluded data from overlapping cancer types, but the response profiles of all drugs targeting the corresponding pathways were retained. Third strategy: The training datasets excluded data from overlapping cancer types, but the response profiles of all drugs targeting the corresponding pathways were retained. Fourth strategy: The training datasets excluded the unique features and response profiles of the predicted drugs targeting the corresponding pathways, as well as data from overlapping cancer types. Fourth strategy: The training datasets excluded the unique features and response profiles of the predicted drugs targeting the corresponding pathways, as well as data from overlapping cancer types. This four-strategy approach effectively addressed potential biases arising from including tested drugs and overlapping cancer types in the training datasets. Elucidating drug resistance mechanisms via pathway enrichment analysis The pathfindR R package was used to investigate the resistance mechanisms of paclitaxel, 5-FU, gemcitabine, and cetuximab24. The models with the highest predictive performance among the EM, EC, and EMC configurations on the TCGA datasets were selected to identify clinically relevant enriched terms through enrichment analysis. For this purpose, only the gene features derived from these best-performing models were utilized. The corresponding p-values and |log₂FC| values of the gene features identified in GDSC were used as input for pathfindR. Enrichment analysis was conducted only under the first and second strategies. A comparison of the enrichment results was then performed to identify both unique and common terms. The common terms were considered indicative of general resistance mechanisms, while the unique terms were examined for their potential association with drug-specific resistance mechanisms. Association of mutations and CNAs with drug resistance using FET FET was applied to the GDSC samples to evaluate the associations of mutation and CNA statuses independently with drug resistance status. Odds ratios (ORs) and p-values were calculated to assess associations. Mutations or CNAs with OR > 1 and p-value < 0.05 were considered potentially associated with drug resistance. However, it was not performed on the PDX & TCGA datasets due to their low sample sizes and high class imbalances. The test was limited to the four drugs used in the enrichment analysis. Results Predictive performances of pan-drug models Drug response predictions yielded strong performance for the five drugs in PDX and eight drugs in TCGA, not only across all samples but also specifically for cohort-level samples (Supplementary Table S1). Table 1 presents the best performing models for these drugs. Predictions were significantly superior to random predictions despite the presence of high class imbalances in drug response profiles. In general, the EC and EMC models achieved the highest performance across both datasets, but EMC models showed greater dominance in TCGA. In PDX, the EC and EM models yielded the highest AUC and AUCPR scores (1.0) for cetuximab in the Non-Small Cell Lung Cancer (NSCLC) cohort and 5-FU in the Colorectal Cancer (CRC) cohort, respectively. However, a reliable assessment of predictive power could not be made for these cases as each included only one sensitive sample. The EM model was the most predictive for erlotinib in the NSCLC cohort, achieving an AUC of 0.93 and an AUCPR of 0.81. For gemcitabine in the Pancreatic Ductal Adenocarcinoma (PDAC) cohort, the EMC model achieved the highest performance, with an AUC of 0.74 and an AUCPR of 0.67. The EC model trained with mitotic inhibitors and the EMC model trained with cytoskeleton inhibitors predicted AUC scores close to 0.90 for all paclitaxel samples with AUCPR scores slightly above 0.40. Similarly, the EC model trained with mitotic inhibitors and the EMC model trained with cytoskeleton inhibitors achieved AUC scores approaching 0.90 for paclitaxel in the Breast Invasive Carcinoma (BRCA) cohort. However, the cytoskeleton-trained model outperformed the mitotic-trained model, achieving an AUCPR of 0.72 compared to 0.67. For all the five drugs, the AUC and AUCPR scores obtained from response predictions were significantly higher than those of random predictions (Table 1). Table 1Predictive performance of pan-drug models evaluated on PDX & TCGA datasets. Models were trained using drugs targeting the same pathways (Mitosis, cytoskeleton, DNA replication, and EGFR signaling).Drug classesSample size (drug classes)DrugsSample size (drugs)Test datasetsCohortsIntegrationIntegration (fine-tuned)AUC (test dataset)AUC (fine-tuned, test dataset)AUCPR (test dataset)AUCPR (fine-tuned, test dataset)AUCPR (random, test dataset)Mitosis6360Paclitaxel43PDXAllECEC0.670.88*0.220.44*0.12 (S:5, R:38)24BRCAEMCEC0.590.87*0.420.67*0.21 (S:5, R:19)Cytoskeleton7561Paclitaxel43AllEMCEMC0.89*0.870.42*0.42*0.12 (S:5, R:38)24BRCAEMCEMC0.89*0.880.690.72*0.21 (S:5, R:19)EGFR signaling5633Erlotinib21All (NSCLC)ECEM0.810.93*0.480.81*0.14 (S:3, R:18)EGFR signaling5633Cetuximab60AllECEC0.630.68*0.150.41*0.08 (S:5, R:55)36CRCEMCEC0.72*0.690.330.40*0.11 (S:4, R:32)24NSCLCECEC0.871.0*0.251.0*0.04 (S:1, R:23)DNA replication94355-Fluorouracil23All (CRC)EMEM1.0*1.0*1.0*1.0*0.04 (S:1, R:22)DNA replication9435Gemcitabine25All (PDAC)EMCEMC0.550.74*0.470.67*0.28 (S:7, R:18)Mitosis6360Paclitaxel35TCGAAllEMCEMC0.76*0.76*0.88*0.88*0.71 (S:25, R:10)27BRCAEMCEMC0.85*0.85*0.96*0.96*0.81 (S:22, R:5)Mitosis6360Docetaxel16AllEMCEC0.730.83*0.780.85*0.5 (S:8, R:8)8BRCAEMCEC0.670.92*0.90.98*0.75 (S:6, R:2)Cytoskeleton7561Paclitaxel35AllECEC0.660.77*0.780.87*0.71 (S:25, R:10)27BRCAECEC0.770.86*0.950.96*0.81 (S:22, R:5)Cytoskeleton7561Docetaxel16AllECEMC0.640.73*0.670.77*0.50 (S:8, R:8)8BRCAEMCEMC0.92*0.92*0.98*0.98*0.75 (S:6, R:2)EGFR signaling4876Cetuximab9All (HNSC)EMCEMC0.610.89*0.720.95*0.67 (S:6, R:3)DNA replication9435Cisplatin64AllEMCEC0.730.81*0.960.98*0.91 (S:58, R:6)37HNSCECEC0.490.67*0.940.97*0.92 (S:34, R:3)DNA replication9435Doxorubicin18AllEMCEMC0.861.0*0.671.0*0.22 (S:4, R:14)15SARCEMCEMC0.851.0*0.671.0*0.13 (S:2, R:13)DNA replication94355-Fluorouracil33AllEMCEMC0.620.65*0.810.84*0.73 (S:24, R:9)18STADEMCEMC0.820.88*0.930.96*0.72 (S:13, R:5)6PAADEMEM0.50.75*0.670.75*0.33 (S:2, R:4)DNA replication9435Gemcitabine55AllEMEM0.580.61*0.430.47*0.36 (S:20, R:35)48PAADECEC0.670.71*0.630.69*0.42 (S:20, R:28)DNA replication9435Temozolomide90All (LGG)ECEC0.58*0.540.220.25*0.12 (S:11, R:79)The table reports performance across different drug classes and test datasets using three MI strategies: EM, EC, and EMC. Fine-tuned results indicate performance after retraining only the classification layer while keeping the expression, mutation, and CNA layers frozen. AUC and AUCPR values are presented for each test dataset along with random baseline AUCPR values for comparison. Bold /asterisk notation is used to indicate the best-performing models. Predictive performance of pan-drug models evaluated on PDX & TCGA datasets. Models were trained using drugs targeting the same pathways (Mitosis, cytoskeleton, DNA replication, and EGFR signaling). The table reports performance across different drug classes and test datasets using three MI strategies: EM, EC, and EMC. Fine-tuned results indicate performance after retraining only the classification layer while keeping the expression, mutation, and CNA layers frozen. AUC and AUCPR values are presented for each test dataset along with random baseline AUCPR values for comparison. Bold /asterisk notation is used to indicate the best-performing models. In TCGA, several models achieved both AUC and AUCPR scores greater than 0.80 for some of the drugs. Overall, the EMC models were strongly predictive across several drugs. High performance with the EMC models was observed for doxorubicin in both the Sarcoma (SARC) cohort and all TCGA samples (AUC = 1.0, AUCPR = 1.0), 5-FU in the Stomach Adenocarcinoma (STAD) cohort (AUC = 0.88, AUCPR = 0.96), cetuximab in the Head and Neck Squamous Cell Carcinoma (HNSC) cohort (AUC = 0.89, AUCPR = 0.95), docetaxel in the BRCA cohort trained with cytoskeleton inhibitors (AUC = 0.92, AUCPR = 0.98), and paclitaxel in the BRCA cohort trained with mitotic inhibitors (AUC = 0.85, AUCPR = 0.96). Similarly, the EC models demonstrated strong predictive performance in multiple cases: paclitaxel in the BRCA cohort trained with cytoskeleton inhibitors (AUC = 0.86, AUCPR = 0.96), docetaxel in all TCGA samples trained with mitotic inhibitors (AUC = 0.83, AUCPR = 0.85), and docetaxel in the BRCA cohort trained with mitotic inhibitors (AUC = 0.92, AUCPR = 0.98). As in PDX, the AUC and AUCPR scores for all the eight drugs in TCGA were significantly higher than those of random predictions (Table 1). The performance of pan-drug models was compared with the predictions of MOLI and other implemented methods using PDX & TCGA datasets13 The pan-drug models outperformed all other methods in drug response prediction for all the six drugs predicted with these methods. In particular, the EM, EC, and EMC models all outperformed MOLI and other implemented methods by a significant margin for erlotinib in PDX and docetaxel in TCGA. Using TL, MOLI predicted an AUCPR of 0.28 for erlotinib and 0.20 for cetuximab. In contrast, the EM model for erlotinib (AUCPR = 0.81) and the EC model for cetuximab (AUCPR = 0.41) outperformed the TL predictions of MOLI by a significant margin in this study (Table 2). Table 2Comparison of the predictive performances of pan-drug models with similar methods implemented on the same PDX & TCGA datasets. This table compares AUCPR values of our pan-drug models using EM, EC, and EMC integration strategies. Results are compared against Elastic Net, early integration via NMF and DNNs, MOLI (complete and pan-drug), and a random classifier. Test datasets include PDX & TCGA. Fine-tuned results are obtained by retraining only the classification layer. Expression, mutation, and CNA layers are kept frozen. Bolded AUCPR values indicate improvement over the best AUCPR scores obtained among all compared models for each test case. “NA” indicates results not applicable.MethodsTest datasetsDrug classesDrugsIntegrationAUCPRAUCPR (fine-tuned)ImprovementElastic netEarly integration via NMFEarly integration via DNNsMOLI completeMOLI pan-drugRandom classifierPDXMitosisPaclitaxelEMC0.190.42+0.180.100.21-0.24NA0.12EC0.220.44+0.20EM0.110.15-0.09CytoskeletonPaclitaxelEMC0.420.42+0.180.100.21-0.24NA0.12EC0.180.34+0.10EM0.140.22-0.02DNA replicationGemcitabineEMC0.470.67+0.180.280.350.350.49NA0.28EC0.470.490EM0.320.40-0.09EGFR signalingCetuximabEMC0.150.19-0.010.060.07-0.110.200.08EC0.150.41+0.21EM0.110.21+0.01EGFR signalingErlotinibEMC0.450.46+0.180.110.28-0.330.280.14EC0.480.61+0.33EM0.360.81+0.53TCGAMitosisDocetaxelEMC0.780.80+0.290.510.510.450.49NA0.50EC0.570.85+0.34EM0.580.75+0.24CytoskeletonDocetaxelEMC0.500.77+0.260.510.510.450.49NA0.50EC0.670.72+0.21EM0.570.67+0.16DNA replicationCisplatinEMC0.960.97+0.040.850.93-0.93NA0.91EC0.960.98+0.05EM0.870.87-0.06DNA replicationGemcitabineEMC0.400.43-0.020.380.370.460.45NA0.37EC0.410.43-0.02EM0.430.47+0.02This table compares AUCPR values of our pan-drug models using EM, EC, and EMC integration strategies. Results are compared against Elastic Net, early integration via NMF and DNNs, MOLI (complete and pan-drug), and a random classifier. Test datasets include PDX & TCGA. Fine-tuned results are obtained by retraining only the classification layer. Expression, mutation, and CNA layers are kept frozen. Bolded AUCPR values indicate improvement over the best AUCPR scores obtained among all compared models for each test case. “NA” indicates results not applicable Comparison of the predictive performances of pan-drug models with similar methods implemented on the same PDX & TCGA datasets. This table compares AUCPR values of our pan-drug models using EM, EC, and EMC integration strategies. Results are compared against Elastic Net, early integration via NMF and DNNs, MOLI (complete and pan-drug), and a random classifier. Test datasets include PDX & TCGA. Fine-tuned results are obtained by retraining only the classification layer. Expression, mutation, and CNA layers are kept frozen. Bolded AUCPR values indicate improvement over the best AUCPR scores obtained among all compared models for each test case. “NA” indicates results not applicable. This table compares AUCPR values of our pan-drug models using EM, EC, and EMC integration strategies. Results are compared against Elastic Net, early integration via NMF and DNNs, MOLI (complete and pan-drug), and a random classifier. Test datasets include PDX & TCGA. Fine-tuned results are obtained by retraining only the classification layer. Expression, mutation, and CNA layers are kept frozen. Bolded AUCPR values indicate improvement over the best AUCPR scores obtained among all compared models for each test case. “NA” indicates results not applicable The pan-drug models were trained using four different strategies to ensure a robust evaluation of predictions for the four drugs shared between PDX & TCGA (Paclitaxel, 5-FU, gemcitabine, and cetuximab). The results showed that the AUCPR scores of the top-performing pan-drug models varied across strategies; nevertheless, all models outperformed random predictions. In TCGA, the AUCPR scores were generally comparable to or higher than those obtained under the first strategy. In contrast, AUCPR scores were mostly higher under the first strategy in PDX (Fig. 6). Predictive performance of pan-drug models (EM, EC, and EMC) evaluated by AUCPR scores across PDX & TCGA datasets. Each bar plot in (A–E) displays the performance of models trained on mitotic, cytoskeleton, DNA replication, and EGFR signaling inhibitors from the GDSC dataset and tested on paclitaxel, 5-FU, gemcitabine, and cetuximab in the PDX & TCGA datasets using DNN based TL with four different strategies. Legends are color-coded by strategy, and the random AUCPR baseline is included for reference. Bar plots show (A) paclitaxel predicted using models trained on mitotic inhibitors, (B) paclitaxel predicted using models trained on cytoskeleton inhibitors, (C) 5-FU predicted using models trained on DNA replication inhibitors, (D) gemcitabine predicted using models trained on DNA replication inhibitors, and (E) cetuximab predicted using models trained on EGFR signaling inhibitors. Models trained with mitotic inhibitors yielded comparable AUCPR scores across all four strategies for paclitaxel in TCGA in both the BRCA cohort and all TCGA samples. However, AUCPR scores in PDX varied across strategies. For all PDX samples, the first and second strategies produced similar results, while the strongest performance was achieved with the EC model under the fourth strategy. In the BRCA cohort, the EMC model achieved the strongest performance under the first strategy, but AUCPR scores showed variability across the other strategies (Fig. 6A). Similarly, models trained with cytoskeleton inhibitors performed comparably across all four strategies for paclitaxel in TCGA in both the BRCA cohort and all TCGA samples. However, the EC model was the strongest for all PDX samples under the first strategy. AUCPR scores decreased under other strategies. In the BRCA cohort, the EC model was again the top performer under the second strategy (Fig. 6B). Models trained with DNA replication inhibitors achieved the highest AUCPR with the EM model under the third strategy for 5-FU in both the STAD cohort and all TCGA samples although similar AUCPR scores were observed under the first and second strategies. In the Pancreatic Adenocarcinoma (PAAD) cohort, where only six samples were available, the strongest predictions were achieved under the first and third strategies. Despite having only one sensitive sample in PDX, the EMC model showed the strongest performance under the first strategy (Fig. 6C). Similarly, models trained with DNA replication inhibitors yielded comparable AUCPR scores across all strategies in TCGA for all gemcitabine samples. However, in the PAAD cohort, the AUCPR obtained under the third strategy was slightly lower. For all PDX samples consisting of only the CRC cohort, AUCPR scores were comparable across strategies except for the fourth strategy, which showed a substantial decrease (Fig. 6D). Models trained with EGFR signaling inhibitors exhibited strategy-dependent AUCPR variations for cetuximab in the HNSC cohort in TCGA, where AUCPR scores were comparable across strategies except for the fourth strategy. For all PDX samples, the EMC model demonstrated the strongest performance under the first strategy. In the CRC cohort, AUCPR scores varied although the first and fourth strategies yielded comparable results. In the NSCLC cohort, where only one sensitive sample was available, AUCPR scores were identical and highest across all strategies except for the second strategy, which showed a substantial decrease (Fig. 6E). The number and list of all features from the GDSC dataset used to predict drug response in the PDX & TCGA datasets are provided in Supplementary Table S2. Drug resistance mechanisms with pathway enrichment results Paclitaxel resistance mechanisms The EMC model was the strongest predictor in TCGA when trained with mitotic inhibitors (Fig. 6A). Pathway enrichment analysis based on its gene features identified 102 enriched terms with the top 10 shown in Fig. 7A (Supplementary Table S3). The top three were “Focal adhesion”, “ECM–receptor interaction”, and “p53 signaling pathway”. Among them, only FYN-mediated “Focal adhesion” remained partially active (Fig. 7B–D). Pathway enrichment results for paclitaxel trained using mitotic inhibitors in GDSC (A) Fold enrichment plot showing the top 10 enriched KEGG pathways. The right-side legend indicates –log₁₀(p-values) and the number of genes contributing to each pathway. Each pathway is represented by a colored circle, with color and size scaled according to these parameters (B) Scores per sample plot for resistant (Res) and sensitive (Sen) GDSC cell lines. The left side lists selected enriched pathways that are likely associated with paclitaxel resistance, not the full set of enriched pathways. The right side includes a color legend indicating the agglomerated z-scores per sample, where values greater than 0 indicate activation and values less than 0 indicate repression (C) Term-gene heatmap showing expression levels of genes within the selected enriched pathways The right-side legend indicates red for down-regulated genes, green for up-regulated genes, and white for genes not included in any of the selected pathways (D) Term-gene network graph displaying connections between enriched pathways and contributing genes. Green nodes represent up-regulated genes, red nodes represent down-regulated genes, and yellow nodes represent enriched terms. The legend shows pathway sizes by number of connected genes. Pathway information was obtained from the KEGG database (Kanehisa Laboratories, 2025) based on data from Kanehisa et al.25–27. Most enriched terms were either partially active or fully suppressed. In particular, survival-related signaling pathways such as PI3K/Akt, MAPK, Ras, and NF-κB were suppressed, whereas compensatory pathways like Wnt and Hippo remained partially active with consistent TCF7L2 upregulation28–32. TCF7L2 was upregulated across multiple terms including “Adherens junction”, which was co-activated with FYN. In contrast, “Tight junction” and “Gap junction” were suppressed. Partial activity was also observed in “Regulation of actin cytoskeleton” mediated by CYFIP2 and CXCR4 upregulation, whereas “Cytoskeleton in muscle cells” was suppressed. LDHB-mediated terms such as “Pyruvate metabolism”, “Central carbon metabolism in cancer”, and “HIF-1 signaling pathway” also showed partial activation (Fig. 7B-D). Similar to mitotic inhibitors, the EMC model was also the strongest predictor in TCGA when trained with cytoskeleton inhibitors (Fig. 6B). Enrichment analysis revealed 94 enriched terms with the top 10 visualized in Fig. 8A (Supplementary Table S4). The leading terms included “Focal adhesion”, “Estrogen signaling pathway”, and “ECM–receptor interaction”. As in the mitotic inhibitor results, only FYN-mediated “Focal adhesion” remained partially active (Fig. 8B–D). Among the 94 enriched terms, 76 overlapped with those found in mitotic inhibitors (Supplementary Table S5). A Spearman rank correlation showed a strong concordance in term rankings between both drug classes (ρ = 0.93, p-value = 6.92 × 10–33). The top three shared terms were “Focal adhesion”, “ECM–receptor interaction”, and “p53 signaling pathway” (Supplementary Table S5). Pathway enrichment results for paclitaxel trained using cytoskeleton inhibitors in GDSC (A) Fold enrichment plot showing the top 10 enriched KEGG pathways. The right-side legend indicates –log₁₀(p-values) and the number of genes contributing to each pathway. Each pathway is represented by a colored circle, with color and size scaled according to these parameters (B) Scores per sample plot for resistant (Res) and sensitive (Sen) GDSC cell lines. The left side lists selected enriched pathways that are likely associated with paclitaxel resistance, not the full set of enriched pathways. The right side includes a color legend indicating the agglomerated z-scores per sample, where values greater than 0 indicate activation and values less than 0 indicate repression (C) Term-gene heatmap showing expression levels of genes within the selected enriched pathways. The right-side legend indicates red for down-regulated genes, green for up-regulated genes, and white for genes not included in any of the selected terms (D) Term-gene network graph combining enrichment results for paclitaxel based on gene features from the best-performing pan-drug models trained with mitotic and cytoskeleton inhibitors. Green nodes represent up-regulated genes, red nodes represent down-regulated genes, and yellow nodes represent enriched pathways common to both models. The legend on the right indicates the size of each pathway based on the number of connected genes. Pathway information was obtained from the KEGG database (Kanehisa Laboratories, 2025) based on data from Kanehisa et al.25–27. The enrichment results from the first and second strategies were also compared. For mitotic inhibitors, 86 terms were shared across both strategies, while 16 were unique to the first and two to the second (Fig. 9A and Supplementary Table S3). Unique terms from the first strategy included LDHB-mediated “Pyruvate metabolism”, CXCR4- and FYN-mediated “Axon guidance”, and the suppressed “Neuroactive ligand–receptor interaction”. For cytoskeleton inhibitors, 88 terms were shared between the two strategies. Among the six unique to the first strategy, LDHB-mediated “Pyruvate metabolism” and the suppressed “Neuroactive ligand–receptor interaction” also appeared in mitotic inhibitors. However, “Transcriptional misregulation in cancer” and “Motor proteins” were exclusive to cytoskeleton inhibitors. Among the seven terms unique to the second strategy, “IL-17 signaling pathway” and “Autophagy - animal” were suppressed (Fig. 9B and Supplementary Table S4). Venn diagrams showing the number of enriched terms identified using gene features from the best-performing pan-drug models under the first and second strategies. Each plot corresponds to a drug-inhibitor class pair: (A) paclitaxel trained with mitotic inhibitors (B) paclitaxel trained with cytoskeleton inhibitors (C) 5-FU trained with DNA replication inhibitors (D) gemcitabine trained with DNA replication inhibitors, and (E) cetuximab trained with EGFR signaling inhibitors. The diagrams display the overlap and uniqueness of enriched terms between the two enrichment results. To the right of each plot, drug-specific DEGs identified in enriched pathways are listed. No such genes were identified for gemcitabine and cetuximab. Among the 11 mitotic inhibitors, paclitaxel-unique DEGs were observed only in the enrichment results of the first strategy (Fig. 9A). A similar pattern was seen for the 11 cytoskeleton inhibitors (Fig. 9B). Paclitaxel-unique DEGs (FYN, LDHB, FLNB, CYP24A1, LYPD6B, and MMP7) were consistently identified in both enrichment analyses of the first strategy across both drug classes (Supplementary Tables S3–S5). 5-FU resistance mechanisms The EM model demonstrated the strongest predictive performance for 5-FU in TCGA among those trained with DNA replication inhibitors (Fig. 6C). Enrichment analysis using the gene features of this model revealed numerous active pathways. Upregulated genes were abundant across various terms, including all cancer-related survival pathways. Similar to the paclitaxel enrichment results, terms related to signaling, cell-cell interaction, adhesion, metabolism, cancer, immunity, and TME were enriched (Supplementary Table S6). The most significant terms (“ECM–receptor interaction”, “Focal adhesion”, and “PI3K/Akt signaling pathway”) were driven by a large number of upregulated genes (Fig. 10A,B). Pathway enrichment results for 5-FU and gemcitabine using gene features from pan-drug models trained with DNA replication inhibitors in GDSC (A) Fold enrichment plot showing the top 10 enriched KEGG pathways for 5-FU. The right-side legend indicates –log₁₀(p-values) and the number of genes contributing to each pathway. Each pathway is represented by a colored circle, with color and size scaled according to these parameters (B) Scores per sample plot for resistant (Res) and sensitive (Sen) GDSC cell lines for 5-FU. The left side lists selected enriched pathways likely associated with 5-FU resistance, not the full set of enriched terms. The right side includes a color legend indicating agglomerated z-scores per sample, where values greater than 0 indicate activation and values less than 0 indicate repression (C) Fold enrichment plot showing the top 10 enriched KEGG pathways for gemcitabine, with visual representation and legend as described in A (D) Scores per sample plot for resistant (Res) and sensitive (Sen) GDSC cell lines for gemcitabine as described in B. Pathway information was obtained from the KEGG database (Kanehisa Laboratories, 2025) based on data from Kanehisa et al.25–27. Key pathways associated with 5-FU resistance were enriched, offering insights into its resistance mechanisms. Multiple signaling pathways such as PI3K/Akt, FoxO, Wnt, Estrogen, JAK/STAT, Ras, MAPK, NF-κB, p53, Hippo, and VEGF were either partially or highly active. Cytoskeleton-related terms such as “Cytoskeleton in muscle cells” and “Regulation of actin cytoskeleton” also showed high activity. Adhesion and cell–cell interaction pathways like “ECM–receptor interaction”, “Focal adhesion”, “Gap junction”, “Tight junction”, and “Adherens junction” were similarly active. Drug metabolism terms such as “Drug metabolism - other enzymes”, “Drug metabolism - cytochrome P450”, and “Metabolism of xenobiotics by cytochrome P450” were composed entirely of upregulated genes. Immunity-related terms such as “TNF signaling pathway”, “B cell receptor signaling pathway”, “T cell receptor signaling pathway”, “IL-17 signaling pathway”, and “PD-L1 expression and PD-1 checkpoint pathway in cancer” were partially active. In particular, AP-1 transcription factor (TF) components FOS and JUN were co-upregulated in these immune-related pathways and in the partially active “Apoptosis”33. Autophagy-related terms like “Lysosome” and “Mitophagy - animal” were active. Additional metabolic pathways including “Porphyrin metabolism”, “Ether-lipid metabolism”, “Glycerolipid metabolism”, and “Retinol metabolism” were either partially or fully active (Fig. 10B and Supplementary Table S6). Pathway enrichment analysis identified 199 enriched terms under the first strategy and 201 under the second with 195 shared between them. Four terms were unique to the first strategy (“Renal cell carcinoma”, “Adipocytokine signaling pathway”, “Longevity regulating pathway”, and “Neurotrophin signaling pathway”), all of which were partially or fully active. The six unique to the second strategy (“Glutathione metabolism”, “TGF-β signaling pathway”, “Platinum drug resistance”, “Signaling pathways regulating pluripotency of stem cells”, “Autophagy - animal”, and “Carbohydrate digestion and absorption”) were also at least partially active (Fig. 9C and Supplementary Table S6). Moreover, 5-FU-specific DEGs (PINK1, MAGED1, LYN, PREX1, TNFSF13B, CASP1, MYO1F, and KLK1) contributed to enriched terms uniquely identified under the first strategy among the 13 DNA replication inhibitors (Fig. 9C). Gemcitabine resistance mechanisms The EM model was the most predictive for gemcitabine in TCGA among those trained with DNA replication inhibitors (Fig. 6D). The most significant enriched terms included “Glutathione metabolism”, “TGF-β signaling pathway”, and “Signaling pathways regulating pluripotency of stem cells”, all of which were either partially or fully active. Moreover, FZD6-mediated terms such as “Hippo signaling pathway”, “Wnt signaling pathway”, and “Melanogenesis” along with CDH1-mediated “Adherens junction” were active. Terms like “Drug metabolism - cytochrome P450”, “Chemical carcinogenesis - DNA adducts”, and “Metabolism of xenobiotics by cytochrome P450” showed exclusive upregulation of MGST2. “Cytokine–cytokine receptor interaction” was also active, whereas “Cytoskeleton in muscle cells” and “Protein digestion and absorption” were suppressed (Fig. 10C-D). Enrichment analyses under the first and second strategies identified 26 shared terms. No unique term appeared in the first strategy, while “Focal adhesion” and “Proteoglycans in cancer” were unique to the second (Fig. 9D and Supplementary Table S7). Cetuximab resistance mechanisms The EM model was the most predictive for cetuximab in TCGA among those trained with EGFR signaling inhibitors (Fig. 6E). Enrichment analysis identified numerous terms related to cell–cell interaction, cellular adhesion, survival-related signaling pathways, cancer, immunity, metabolism, and TME (Supplementary Table S8). The most significant terms were “ECM–receptor interaction”, “Focal adhesion”, and “PI3K/Akt signaling pathway”, all of which were partially active (Fig. 11A,B). A total of 167 enriched terms were shared between the first and second strategies. Four terms were unique to the first strategy, while three were unique to the second (Fig. 9E). Pathway enrichment results for cetuxinab trained using EGFR signaling inhibitors in GDSC (A) Fold enrichment plot showing the top 10 enriched KEGG pathways. The right-side legend indicates –log₁₀(p-values) and the number of genes contributing to each pathway. Each pathway is represented by a colored circle, with color and size scaled according to these parameters (B) Term-gene heatmap showing expression levels of genes within the selected enriched pathways that are likely associated with cetuximab resistance, not the full set of enriched pathways. The right-side legend indicates red for down-regulated genes, green for up-regulated genes, and white for genes not included in any of the selected pathways (C) Term-gene network graph displaying connections between enriched pathways and contributing genes. Green nodes represent up-regulated genes, red nodes represent down-regulated genes, and yellow nodes represent enriched terms. The legend shows pathway sizes by number of connected genes. Pathway information was obtained from the KEGG database (Kanehisa Laboratories, 2025) based on data from Kanehisa et al.25–27. PI3K/Akt, Wnt, and Hippo signaling pathways remained active although most survival-associated signaling pathways were suppressed. Terms related to cell–cell interaction and cellular adhesion were either inactive or exhibited limited activity. In contrast, “Cell adhesion molecules” showed increased activity, while “EGFR tyrosine kinase inhibitor resistance” was suppressed. Cytoskeleton-related terms such as “Regulation of actin cytoskeleton” and “Cytoskeleton in muscle cells” were partially active (Fig. 11B,C and Supplementary Table S8). Numerous immune-related terms were also enriched. Among them, “IL-17 signaling pathway” and “TNF signaling pathway” were suppressed, whereas “B cell receptor signaling pathway”, “T cell receptor signaling pathway”, and “PD-L1 expression and PD-1 checkpoint pathway in cancer” remained partially active. In particular, AP-1 TF components JUN and FOS were downregulated across several immune-related terms33. While “Apoptosis” was suppressed, E2F2-mediated “Cellular senescence” remained active. The “MicroRNAs in cancer” term mediated by E2F2, ZEB1, and VIM was also active. Moreover, metabolic terms such as “Arginine biosynthesis”, “Central carbon metabolism in cancer”, and “Choline metabolism in cancer” were suppressed, whereas “Inositol phosphate metabolism” was active with PIK3CG upregulation. “Chemical carcinogenesis – reactive oxygen species” was another suppressed term (Fig. 11B,C and Supplementary Table S8). Association of mutation and CNA statuses with drug resistance FET did not identify any significant association between mutation status and resistance for the gene features used in the models for paclitaxel and cetuximab. However, significant associations were observed for 5-FU with TJP1 (OR = 6.83, p-value = 0.02) and CNTRL (OR = 6.41, p-value = 0.03), and for gemcitabine with NUP210 (OR = ∞, p-value = 0.04) and COL3A1 (OR = ∞, p-value = 0.04) (Supplementary Table S9). FET identified significant associations between CNA status and resistance for gene features used in the models for paclitaxel, 5-FU, and gemcitabine. Among paclitaxel-unique genes, LDHB showed a potentially significant association (OR = 2.87, p-value = 0.05). In addition, several genes within enriched pathways demonstrated significant associations including PLAU (OR = 5.87, p-value = 0.007), TCF7L2 (OR = 4.11, p-value = 0.02), and DKK1 (OR = 3.51, p-value = 0.04). For 5-FU, more than 500 genes exhibited significant associations including several 5-FU-unique genes such as PINK1 (OR = 2.34, p-value = 0.003), PREX1 (OR = 2.68, p-value = 5.7 × 10⁻⁵), TNFSF13B (OR = 2.13, p-value = 0.002), and MYO1F (OR = 2.72, p-value = 0.0005). Significant associations were also identified for gemcitabine, with over 100 genes showing CNA-related resistance links (Supplementary Table S9). Discussion Our enrichment analyses revealed that LDHB-mediated “Pyruvate metabolism” was uniquely active for paclitaxel. LDHB was upregulated exclusively for paclitaxel among all mitotic, cytoskeleton, DNA replication, and EGFR signaling inhibitors, and “Pyruvate metabolism” was uniquely enriched for paclitaxel. Furthermore, FET for LDHB indicated a possible association between paclitaxel resistance and CNA status. These findings suggest that transcriptional upregulation and CNAs in LDHB may be associated with paclitaxel resistance. Supporting this, previous studies reported that resistant cancer cells may upregulate LDHB through metabolic reprogramming by converting lactate to pyruvate and utilizing it via oxidative phosphorylation under hypoxic conditions34,35. In addition, LDHB upregulation was linked to paclitaxel resistance in oral squamous cell carcinoma36. Moreover, LDHB knockdown in Merkel cell carcinoma cell lines was shown to increase paclitaxel-induced apoptosis, which was reversed upon LDHB reintroduction37. Paclitaxel-resistant cell lines are likely in a drug-tolerant persister (DTP) state, which is associated with the development of drug-resistant phenotypes38. In DTP, cancer cells prioritize survival over proliferation in a state of reversible quiescence in response to drug stress, and Wnt signaling pathway is commonly reported to be active39. In this state, stress response and compensatory pathways are activated, epigenetic and metabolic reprogramming occurs, stem-like properties are exhibited, and pathway activities are either partially active with only a few genes mediating them or completely suppressed39. In our paclitaxel enrichment results, most terms were suppressed or partially active with upregulation of several genes across many terms. Although many survival-related pathways were suppressed, TCF7L2-mediated stemness-associated compensatory Wnt and Hippo signaling pathways maintained survival31,32. FYN-mediated “Focal adhesion” might also contribute to apoptotic evasion and promote survival40. Beyond its role in promoting survival, FYN upregulation was an important finding since its activation was reported as a general characteristic of the DTP phenotype41. In addition, LDHB-mediated “Pyruvate metabolism” and ODC1-mediated “Glutathione metabolism” indicated metabolic reprogramming, while the EZH2-mediated “MicroRNAs in cancer” pointed to epigenetic reprogramming42. In our enrichment results for paclitaxel, terms related to cellular adhesion and cell-cell interaction such as “ECM-receptor interaction” and “Tight junction” were suppressed. In contrast, FYN-mediated “Focal adhesion” and TCF7L2- and FYN-mediated “Adherens junction” were partially active. In addition, survival-related signaling pathways such as p53 and PI3K/Akt were suppressed. These results imply anoikis resistance, in which FYN upregulation may represent a paclitaxel-specific survival mechanism. FYN was identified as a DEG exclusively for paclitaxel among 21 mitotic and cytoskeletal inhibitors. Our finding that FYN was the only upregulated gene in the most significant term, “Focal adhesion”, may indicate that paclitaxel-resistant cell lines utilize FYN-driven adhesion signaling to evade apoptosis. Supporting this, Chou et al.43 showed that FYN was upregulated in paclitaxel-resistant basal-like triple-negative breast cancer (TNBC). Similarly, Egeland et al.44 reported FYN upregulation in resistant TNBC patients and in rapidly progressing metastatic lesions. However, FLNB and MYL9, two other paclitaxel-specific DEGs we found in “Focal adhesion” term, were downregulated. These genes are known to be involved in actin cytoskeleton remodeling and contraction45,46. Their downregulation may help cells survive mitotic arrest by disrupting cytoskeletal tension. FYN upregulation and MYL9 downregulation were also observed in “Adherens junction” and “Axon guidance” terms. Our enrichment results showed that TCF7L2 upregulation may also contribute to paclitaxel resistance, potentially acting in concert with FYN. TCF7L2 was found to be upregulated not only in “Adherens junction”, but also exclusively in the Wnt and Hippo signaling pathways, as well as in various cancer-related terms. Unlike FYN, however, FET indicated that CNAs in TCF7L2 may also play a role in paclitaxel resistance in addition to its transcriptional upregulation. Mechanistically, FYN can activate the Wnt signaling pathway both directly by phosphorylating β-catenin or indirectly by reducing β-catenin degradation through interaction with GSK3β in the β-catenin destruction complex47–49. Thus, TCF7L2 upregulation might be triggered by FYN activation. FYN may be a critical player in paclitaxel resistance not only by upregulating TCF7L2 and by being activated through phosphorylation by CXCR4, as well50. CXCR4, which was co-upregulated with FYN in “Axon guidance”, was also co-upregulated with TCF7L2 in “Pathways in cancer”. CXCR4 is known to be upregulated in hypoxic niches, where it may promote chemotactic migration to CXCL12-rich adaptive areas in TME, enabling cells evade paclitaxel exposure51. CXCR4 was also upregulated in “Regulation of actin cytoskeleton”. This indicates that alterations in the actin cytoskeleton may increase invasiveness, motility, metastasis, and EMT in paclitaxel-resistant cell lines, supporting resistance through apoptotic evasion52. In particular, CXCR4 had the highest logFC among all genes upregulated by paclitaxel in this study. These findings suggest that CXCR4, FYN, and TCF7L2 might act together in an interconnected mechanism contributing to paclitaxel resistance. Similar to the enrichment results for paclitaxel, CXCR4, FYN, and TCF7L2 were mostly upregulated in the same terms in the enrichment results for cetuximab. This suggests that both drugs may share similar resistance mechanisms. Clinical trial results showing improved treatment response with the combination of paclitaxel and cetuximab in metastatic or recurrent HNSC support the hypothesis that these genes may play a pivotal role in resistance and that co-targeting them could enhance treatment success in combination strategies53,54. In our paclitaxel results, FYN was a unique DEG among mitotic and cytoskeleton inhibitors, while it was upregulated for cetuximab but did not meet the DEG threshold by logFC. However, FYN was a DEG for other EGFR inhibitors such as gefitinib, afatinib, and lapatinib, suggesting a potential broader role in resistance to EGFR-targeted therapies. This aligns with previous studies reporting that FYN can be activated by EGFR pathway mutations and may play a role in resistance and cancer progression49,55. Moreover, while CNA status of TCF7L2 was significantly associated with paclitaxel resistance, no such association was observed for cetuximab, indicating that TCF7L2 may contribute to cetuximab resistance mainly through transcriptional upregulation. Among the genes upregulated for cetuximab, CXCR4 had the highest logFC, consistent with its expression pattern in paclitaxel. Our enrichment results for cetuximab indicate that cetuximab-resistant cell lines may be in a DTP state analogous to that observed in paclitaxel-resistant cell lines. In the enrichment results, most terms were either partially active or suppressed, with a higher number of downregulated genes. Key survival-related pathways such as PI3K/Akt, MAPK, NF-κB, Wnt, and Hippo signaling were all partially active, indicating limited survival signaling. In addition, PIK3CG-mediated inositol phosphate metabolism, along with ZEB1-, E2F2-, and VIM- mediated “MicroRNAs in cancer” term pointed to both metabolic and epigenetic reprogramming. Most importantly, FYN upregulation, a general characteristic of the DTP phenotype, was also observed41. Our enrichment results for cetuximab also revealed key findings about its resistance mechanisms. The suppression of “EGFR tyrosine kinase inhibitor resistance” suggested that canonical EGFR-dependent mechanisms may not be the main drivers of resistance, and that resistant cells may survive by bypassing EGFR56. In support of this, our results indicated that PIK3CG-mediated “Inositol phosphate metabolism” may activate the PI3K/Akt pathway in an EGFR-independent manner through metabolic reprogramming, thereby promoting cell survival and resistance with immune suppression57,58. Moreover, the upregulation of VIM and ZEB1 and the downregulation of CDH1 across multiple enriched terms implied that EMT might play a role in cetuximab resistance59. Cetuximab might also contribute to resistance through immune suppression by blocking the EGFR signaling pathway60. Since EGFR inhibition may suppress the MAPK pathway, our findings align with this as we found only partial activation of MAP4K1-mediated “MAPK signaling pathway” with our enrichment results61. Several immune-related terms including “B-cell signaling pathways” and “T-cell signaling pathway” were partially active, while key components of the AP-1 TF complex (JUN, FOS, and FOSL1), associated with immune responses through the MAPK pathway, were downregulated within these terms62. However, LCK, a Src kinase, was upregulated in immune-related terms, either alone or in combination with FYN, another Src kinase. Both FYN and LCK are important in T-cell response63. This pattern suggests that, rather than triggering a full immune response, cetuximab-resistant cell lines might be in a partial immune-like signaling state by selectively activating Src kinases to support survival and cytoskeletal adaptation64. Our enrichment results for 5-FU showed an opposite trend compared to paclitaxel and cetuximab, with many terms being active and containing many upregulated genes. In addition, some cancer stem cell (CSC) markers such as ALDH1A1, ALDH1A3, SOX9, KLF5, and ITGB1 were upregulated in several enriched terms65–69. These findings suggest that 5-FU-resistant cell lines likely exhibit a stem-like phenotype associated with strong drug resistance70,71. Enrichment analyses for 5-FU revealed that genes uniquely associated with 5-FU among the unionized gene features of 13 DNA replication inhibitors were represented in the enriched terms. Among these, FET analysis showed a significant association between CNA status and resistance for PINK1, PREX1, TNFSF13B, and MYO1F, suggesting that both CNAs and transcriptional regulation of these genes might contribute to 5-FU-specific resistance. In our enrichment results, PINK1 was upregulated, while the other three genes were downregulated. The association between PINK1 upregulation and 5-FU resistance has been reported in previous studies. Liu et al.72 demonstrated that low expression of the lncRNA ABALON led to PINK1 upregulation, which promoted mitophagy and contributed to 5-FU resistance. Similarly, Yamashita et al.73 found that PINK1 upregulation was linked to mitophagy, 5-FU resistance, and poor prognosis in esophageal squamous cell carcinoma. However, to our knowledge, no prior studies have shown that CNAs in PINK1 may also contribute to 5-FU resistance. Therefore, this study is the first to suggest that both transcriptional upregulation and CNAs in PINK1 might be together associated with 5-FU resistance. The relatively weak results observed for gemcitabine in the enrichment analysis can be attributed to the limited number of DEGs identified for this drug. Specifically, only one DEG was detected following differential expression analysis, which was insufficient to conduct pathway enrichment. Despite this limitation, the prediction performance for gemcitabine in both the PDX & TCGA datasets consistently exceeded that of a random classifier. This is likely due to the contribution of additional informative genes included in the unionized feature space derived from other DNA replication inhibitors. These shared DEGs provided a meaningful signal to the model, compensating for the paucity of gemcitabine-specific features and enabling above-baseline performance. This suggests that gemcitabine may exhibit cancer-type-specific activity and involve distinct resistance mechanisms. Another plausible explanation is that resistance to gemcitabine may be more strongly influenced by epigenetic or post-transcriptional regulation, as well as other genomic alterations, rather than transcriptional activity alone74–76. Therefore, conclusions regarding gemcitabine resistance mechanisms should be interpreted with caution. Future studies focusing on specific cancer types in which gemcitabine is clinically effective are necessary to gain deeper insights into its resistance mechanisms. Although the GDSC dataset included a large number of cell line samples, the number of drug response–annotated samples in the PDX & TCGA datasets was unfortunately limited. In fact, the imbalance in drug response classes posed an additional challenge. When we further restricted the analysis to cohort-specific samples rather than using all available data, the sample size decreased even more. Nevertheless, our TL approach was able to generate strong predictions even with these small datasets. In particular, obtaining more meaningful and reliable predictions would require larger sample sizes, particularly from cohorts in which the drug is known to be effective. This would ultimately enable more robust and generalizable predictions. In this study, we integrated gene expression, somatic mutation, CNA, and response data from GDSC, PDX, and TCGA using a DNN based TL approach. We combined drugs targeting mitotic, cytoskeleton, DNA replication, and EGFR signaling pathways in the GDSC dataset during the training phase, and subsequently performed drug response predictions using independent PDX & TCGA test datasets. Our predictions outperformed MOLI on these two test datasets in terms of AUCPR for all the six drugs MOLI predicted. In particular, our models outperformed MOLI for erlotinib and cetuximab, where MOLI applied TL to predict responses in the PDX dataset, highlighting the effectiveness of training drugs within the same pathway using the TL approach. However, a larger number of samples is required to further enhance predictive power77. Integration of additional omics data such as metabolomics and proteomics could further enhance drug response prediction78. Nevertheless, our findings suggest that integrating all omics layers does not necessarily improve predictive performance. Integrating different omics data can lead to more complex models, and the presence of sparse data may hinder model learning; therefore, using all omics data together does not necessarily guarantee more powerful predictions79. Furthermore, the structural similarity and target profiles of drugs were not considered in this study. Incorporating such drug-based features into the model training phase could lead to the development of more powerful and generalizable models80. Conclusions In this study, we demonstrated that using a DNN-based TL approach with MI and drug pathway information enables strong predictive performance. Our approach not only improved predictive performance on in-vivo PDX and ex-vivo TCGA datasets, but also uncovered key mechanisms of resistance to paclitaxel, 5-FU, gemcitabine, and cetuximab. In particular, LDHB-mediated pyruvate metabolism emerged as a potential driver of paclitaxel resistance, whereas PINK1-mediated mitophagy may have a pivotal role in 5-FU resistance. In addition, we showed that integrating all omics layers does not necessarily yield the most powerful or generalizable models. We also revealed that utilizing DEG information of drugs within the same pathway can enhance predictive performance by prioritizing a reduced set of biologically meaningful features. These findings underscore the value of integrating omics-driven insights with pathway-informed modeling to enhance drug response prediction, elucidate mechanisms of drug resistance, and support personalized treatment strategies in precision oncology. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 2 Supplementary Material 3 Supplementary Material 4 Supplementary Material 5 Supplementary Material 6 Supplementary Material 7 Supplementary Material 8 Supplementary Material 9 Acknowledgements We acknowledge the use of computational resources provided by the National Center for High Performance Computing of Turkey under TUBITAK through the TRUBA infrastructure. Author contributions S.A. and O.U.S. contributed substantially to the study’s conception, design, data acquisition, analysis, and interpretation. S.A. led the drafting of the manuscript, and both authors critically revised it for significant intellectual content. Both authors have approved the final version for publication and accept full responsibility for the accuracy and integrity of the work, ensuring that any issues are properly investigated and resolved. Data availability All code used for data preprocessing, MI, TL, and pathway enrichment analysis is openly available on GitHub at https://github.com/alpsoysemih/multiomics-drug-response-resistance-tl. The datasets used in this study include processed gene expression, somatic mutation, and CNA profiles. Drug response annotations from GDSC, PDX, and TCGA cohorts are also included. All data can be reached from the Zenodo repository https://doi.org/10.5281/zenodo.15637912. Declaration Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1 Supplementary Material 2 Supplementary Material 3 Supplementary Material 4 Supplementary Material 5 Supplementary Material 6 Supplementary Material 7 Supplementary Material 8 Supplementary Material 9 Data Availability Statement All code used for data preprocessing, MI, TL, and pathway enrichment analysis is openly available on GitHub at https://github.com/alpsoysemih/multiomics-drug-response-resistance-tl. The datasets used in this study include processed gene expression, somatic mutation, and CNA profiles. Drug response annotations from GDSC, PDX, and TCGA cohorts are also included. All data can be reached from the Zenodo repository https://doi.org/10.5281/zenodo.15637912.