DRIVE: a comprehensive resource deciphering drug-induced transcriptomic and splicing response in cancer cell

Abstract

Pharmacotherapy induces complex molecular reprogramming in cancer, driving transcriptome-wide alterations and widespread dysregulation of alternative splicing. Despite these profound changes, there remain limited resources characterizing drug-induced whole-transcriptomic responses in cancer. Furthermore, while aberrant splicing can generate immunogenic neoantigens, existing resources fail to systematically integrate drug perturbations, splicing dynamics, and neoantigen landscapes. To address this gap, the DRIVE database was constructed as a comprehensive resource detailing drug-induced transcriptomic and splicing responses. Utilizing the large language models for rigorous metadata curation and construct the standardized processing pipeline, thousands of publicly available raw transcriptomic datasets from drug-treated and control cancer cell lines were systematically processed. The resulting repository encompasses 3,911 samples, involving 278 drugs and 272 cell lines, enabling the precise quantification of differential gene expression, differential alternative splicing events, and the prediction of splicing-derived human leukocyte antigen-binding peptides. Analysis of the data revealed that drug-induced transcriptomic reprogramming is highly context-dependent and correlated with chemical structural similarity. We identified Osimertinib as a potential immunomodulatory agent associated with transcriptional signatures of an activated tumor microenvironment, while KB-0742 emerged as an unappreciated candidate global splicing modulator. Furthermore, our large-scale prediction of differential splicing-derived neoantigens uncovered several drugs that warrant further investigation as candidates for combination immunotherapy. DRIVE also provides a user-friendly interface to browse datasets, perform drug enrichment and connectivity analysis. (https://componclab.com/DRIVE). This database could improve our understanding of molecular reprogramming under pharmacotherapy, and serve as a valuable platform for deciphering drug mechanisms, promoting virtual cell modeling and discovering novel strategies of drug repurposing.

Introduction

Pharmacotherapy remains the cornerstone of cancer treatment, with transcriptome-wide alterations serving as the critical molecular foundation for both drug efficacy and the emergence of resistance. Systematic profiling of drug-perturbed gene expression is essential for elucidating mechanisms of action (MoA) [1], identifying novel biomarkers [2], and guiding rational drug combinations [3]. The emerging of virtual cell modeling also requires high-quality drug perturbation data as the resource for model training and evaluation. Large-scale efforts such as the LINCS L1000 project have catalogued gene expression signatures for thousands of perturbagens, yet this resource is limited to approximately 1,000 landmark genes, thereby restricting transcriptome-wide resolution [4]. While repositories of clinical transcriptomic response data, such as CTR-DB [5], provide valuable patient-derived insights but are confounded by heterogeneous cell type composition and often lack well-matched control samples. Conversely, public repositories like the Gene Expression Omnibus (GEO) house a vast amount of untapped drug-treatment transcriptomic data in cancer cell lines [6]. Systematic mining of these resources promises to substantially advance our understanding of drug-transcriptome relationships and hold immense potential to accelerate therapeutic discovery.

Drug-induced transcriptomic alterations extend beyond changes in gene expression levels to encompass widespread reprogramming of alternative splicing (AS). AS allows a single gene to generate diverse mRNA and protein products and is frequently dysregulated in cancer, contributing to oncogenic isoform switching, metabolic reprogramming, immune evasion, and therapeutic resistance [7]. Given the critical role of aberrant AS in cancer progression, targeting RNA splicing has emerged as a promising therapeutic for cancer [8]. Emerging evidence suggests that some small molecules, originally developed as general anti-cancer agents, actually function by targeting the splicing machinery [9]. A prominent example is Indisulam, which was later identified as a molecular glue that induces the degradation of the splicing factor RBM39, leading to widespread splicing defects [10]. Despite its importance, a comprehensive resource detailing drug-induced AS changes across diverse cancer types and therapeutic agents is currently lacking.

Beyond their impact on intrinsic cellular functions, aberrant isoforms generated by drug-induced AS may be translated, processed and presented on major histocompatibility complex (MHC) class I molecules as neoantigens recognizable by the immune system [11]. Tumor neoantigens play an important role in cancer immunotherapy, including immune checkpoint blockade and therapeutic cancer vaccines. Compared with neoantigens derived from somatic single-nucleotide variants (SNVs), AS-derived neoantigens offer several distinctive advantages, such as higher tissue specificity, greater immunogenic potential [12,13], and higher abundance in certain "cold" tumors with low mutational burdens [14]. Recent studies have demonstrated that pharmacological modulation of RNA splicing by agents such as indisulam and MS-023 [15] generates bona fide neoepitopes that elicit robust CD8+ T cell responses and synergize with anti-PD-1 therapy, transforming immunologically cold tumors into hot ones. Similarly, chemical induction of splice-neoantigens by the small molecule RECTAS suppressed tumor growth in a CD8+ T cell- and MHC class I-dependent manner and enhanced immune checkpoint blockade efficacy in preclinical models of colorectal cancer [16]. Nevertheless, despite these promising proof-of-concept studies, the landscape of drug-induced potential neoantigens has not been systematically decoded on a large scale.

Although drugs, alternative splicing, and neoantigens are intimately interconnected, existing resources fail to integrate these three dimensions. To bridge this gap, we constructed DRIVE (Drug-induced RNA profile, Isoform Variability and Epitopes), a comprehensive resource integrating, drug induced transcriptomic responses, alternative splicing and neoantigen landscapes. We systematically retrieved drug-treated and matched control cancer cell line transcriptomic datasets from the GEO database, utilizing Large Language Models (LLMs) combined with rigorous manual inspection to ensure high-fidelity metadata. By implementing a standardized pipeline for raw data processing, we quantified drug-induced differential gene expression, AS alterations, and predicted splice-neoantigens across thousands of conditions. Leveraging DRIVE, we identified Osimertinib as a potential modulator of the tumor microenvironment and KB-0742 as a potential modulator of inducing global AS reprogramming. We also found several drugs with high potential for combination immunotherapy. Furthermore, we developed a user-friendly web interface (available at https://componclab.com/DRIVE) that enables drug enrichment, gene signature connectivity analysis, interactive downloading and visualization, thereby facilitating data access and exploration. DRIVE not only offers a omic perspective for understanding drug MoA but also serves as a hypothesis-generating platform for prioritizing drug-induced vaccine targets and informing the design of synergistic immunotherapeutic strategies.

Results

Curation and processing of transcriptome response data from drug-treated cancer cell lines

To systematically investigate the drug-induced transcriptomic response and their potential immunological implications, we collected and curated transcriptome data from drug-treated and control tumor cell lines from public repositories. As illustrated in Fig. 1A, to obtain consistent results, we applied uniform processing starting from the raw data. Notably, we combined LLMs techniques with meticulous manual curation to obtain high-confidence structural sample metadata. After stringent quality control and a series of standardized processing steps (detailed in Fig. 2 and Methods), we systematically identified transcriptomic changes, AS variations, and predicted neoantigens derived from differential alternative splicing. Based on these comprehensive analytical outcomes, we constructed the DRIVE database and web server, providing a user-friendly online platform to facilitate access and downstream analysis.

Fig. 1.

Overview of the DRIVE database and data statistics. (A) Schematic illustration of the DRIVE database. Raw transcriptome data from control and drug-treated tumor cell lines were initially collected. After a series of processing steps (detailed in Fig. 2 and Methods), transcriptomic changes, alternative splicing variations, and predicted neoantigens derived from differential alternative splicing were obtained. Based on these resources, we constructed the DRIVE database and web server to facilitate user access to relevant data and enable online, code-free analysis. (B) Flowchart of sample collection and screening. (C) Left: Distribution of cell line counts by tissue origin. Right: Distribution of sample counts by treated drug types. (D) Sample size distribution of drug-treated (top) and control groups (bottom). The number in the bar is un-log counts. (E) Sample size distribution across different drug types and tissue types. The size and color of the circles indicate the sample size, with blank spaces representing missing drug–cell line combinations.

Fig. 2.

Data processing workflow. The data processing is divided into four modules: raw data download and processing, metadata extraction, gene expression and alternative splicing quantification, and HLA-I binding prediction of AS-derived peptides. The left part of box in the workflow represents the processing content, and the right part indicates the tools used. Detailed data processing procedures are provided in the Methods section.

To ensure data reliability and comparability, we implemented a rigorous pipeline for sample collection and screening (Fig. 1B, detailed in Methods). Following the careful manual inspection, the retained high-confidence datasets were subjected to further analysis (The final data consist of 3,911 samples from 243 datasets, involving 931 drug-cell treatments from 272 cancer cell lines and 278 treated drugs). The statistics of the curated data reveal the broad landscape of our collection. Specifically, the distribution of cell lines categorized by their tissue origin demonstrates a diverse representation across multiple cancer types (Fig. 1C, left). Similarly, the distribution of sample counts according to the treated drug types highlights a wide coverage of chemotherapeutic and targeted agents (Fig. 1C, right).

Furthermore, we evaluated the sample size distribution for both the drug-treated and control groups (Fig. 1D). Most groups have three or more samples. The result indicates a robust volume of samples in both conditions, ensuring adequate statistical power for subsequent differential expression and splicing analyses. To provide a more granular view of the dataset composition, we also visualized the sample size distribution across the intersection of different drug classes and tissue types (Fig. 1E). Collectively, the systematic curation of these transcriptome response data lay a solid foundation for the DRIVE database, enabling comprehensive exploration of drug-induced molecular phenotypes.

Drug-induced transcriptomic rewiring and identification of potential immunomodulatory agents

To characterize the global transcriptomic alterations triggered by pharmacological interventions, we systematically analyzed the differential gene expression profiles between drug-treated and untreated control cancer cell lines. Following drug treatments, cell lines exhibited widespread transcriptomic changes, as evidenced by the substantial number of differentially expressed genes (DEGs) was identified. (|log2FC| > 1 and adjusted P < 0.05) (Fig. 3A, upper panel, median number of DEGs is 794). To elucidate the functional implications of these transcriptomic shifts, we performed Gene Set Enrichment Analysis (GSEA) using 10,490 ontology gene sets from the MSigDB database. The median number of significantly enriched pathways (adjusted P < 0.1) is 923 (Fig. 3A, lower panel). Together, these results demonstrate that drug exposures elicit broad and profound reprogramming of the transcriptome, affecting both individual gene expression and the activity of biological pathways.

Fig. 3.

Landscape of drug-induced transcriptome responses. (A) Distribution of the number of differentially expressed genes (upper panel) and significantly enriched pathways (lower panel). Differentially expressed genes are defined as those with an absolute log2 fold change (|log2FC|) > 1 and an adjusted P value < 0.05 from DESeq2 analysis. Significantly enriched pathways are defined as those with an adjusted P value < 0.1 from GSEA, using 10,490 ontology gene sets from the MSigDB database. (B) Comparison of drug-induced transcriptomic response similarities between cell lines derived from the same tissue origin (Y) versus different tissue origins (N), displayed separately for different drug types. The transcriptomic response similarity was calculated as the Kendall correlation coefficient between the DESeq2 statistics of different drug-cell line treatments. P values were determined by the Wilcoxon rank-sum test. (C) Distribution of transcriptomic response similarities induced by drugs applied to cell lines from the same tissue origin (upper panel) and different tissue origins (lower panel). (D) Difference in the proportions of samples treated with the same drug versus different drugs between the high and low transcriptomic response similarity groups. P values were obtained using the Chi-squared test. Pairs with a similarity > 0.5 were classified into the high-similarity group, and the rest into the low-similarity group. (E) Comparison of drug similarity between the high and low transcriptomic response similarity groups. The left panel shows the Tanimoto similarity based on the SMILES representations of the drug molecules, while the right panel shows the cosine similarity between drug molecular embeddings computed by a large-scale chemical language model (MoLFormer). (F) ROC curves evaluating the predictive power of drug structural similarity for classifying high versus low transcriptomic response similarity. The Tanimoto similarity based on SMILES representations (left) and the cosine similarity based on MoLFormer embeddings (right) were used as predictors. C-F was calculated on cell lines with the same type of tissue origin.

Given the ubiquity of these widespread transcriptomic changes, we next sought to evaluate the similarity of drug-induced transcriptomic responses across different treatment conditions. By calculating the Kendall correlation coefficient between the DESeq2 statistic vectors of any two drug-cell line treatments, we found that, overall, the transcriptomic response similarities were notably low across diverse drug-cell line pairs (Fig. 3B). This indicates that drug-induced transcriptional perturbations are highly context-dependent. When comparing cell lines from the same tissue origin versus those from different tissues, we observed that the drug-induced transcriptomic perturbation similarities were consistently higher within the same tissue type (Fig. 3B). This suggests that shared tissue-specific transcriptional programs impose a measurable constraint on the cellular response to pharmacological perturbations. Furthermore, the degree of response similarity was strongly influenced by the mechanism of action of the drug classes. For drug categories with diverse and dispersed targets, such as Natural_Product_Derived/TCM and Other_Therapeutic, the transcriptomic response similarities were particularly low (Fig. 3B). In contrast, drug classes with highly specific and similar targets, such as Kinase_inhibitors, exhibited notably higher transcriptomic response similarities (Fig. 3C). Similarly, samples treated with the same class of drug are enriched in groups with high transcriptional response similarity (Kendall correlation > 0.5, Fig. 3D).

While the aforementioned drug classifications are predominantly based on mechanisms of action or molecular targets, current phenotypic drug discovery [17] increasingly relies on the inherent structural characteristics of drug molecules themselves. Therefore, to directly assess how the chemical properties of drugs relate to their transcriptomic perturbations, we compared the molecular similarity of drugs between the high and low transcriptomic response similarity groups. We utilized two metrics: the classic Tanimoto similarity based on SMILES representations, and the cosine similarity of molecular embeddings derived from MoLFormer [18], a large-scale chemical language model. Both metrics consistently demonstrated that drug pairs within the high transcriptomic similarity group possessed significantly higher chemical structural similarity compared to those in the low-similarity group (Fig. 3E). To quantitatively evaluate the predictive power of structural similarity for transcriptomic convergence, we constructed Receiver Operating Characteristic (ROC) curves using the two structural metrics as predictors. Both features demonstrated robust predictive capabilities, with the MoLFormer embedding similarity achieving a higher AUC (0.814) compared to the Tanimoto similarity (0.791) (Fig. 3F). This finding indicates that structurally analogous compounds, especially those sharing deep chemical language representations, are more likely to induce convergent transcriptomic reprogramming.

To demonstrate the utility of our database, we sought to identify drugs capable of transforming the tumor immune microenvironment (TME) into a "hot" state, which is generally associated with enhanced anti-tumor immunity and improved responsiveness to immunotherapy. We established a screening strategy based on three gene signatures reflecting distinct immune-modulatory features: antigen processing and presentation machinery from msigdb database, a CRISPR screen-derived gene set of cancer-intrinsic evasion of T cell killing [19], and genes up-regulated in immune-resistant cancer cell lines [20]. We hypothesized that an ideal immunostimulatory drug would positively enrich the first two gene sets while negatively enriching the immune evasion upregulated gene set. Applying GSEA analysis on our DRIVE data, we evaluated the drug-induced enrichment profiles across various treated cell lines. As a result, we identified multiple drugs that met these criteria (Fig. 4A). Notably, several of these candidate drugs have already been reported in previous literature to enhance the anti-tumor immune microenvironment, thereby validating the reliability of our screening strategy. For example, CCS1477 [21,22], Crizotinib [23], JQ1 [24,25], PLX51107 [26,27]. Among the identified candidates, We selected Osimertinib for further validation, due to the scarcity of matched pre- and post-treatment data for the other drugs in public repositories. To validate its potential to activate the TME, we leveraged publicly available single-cell RNA sequencing (scRNA-seq) data from [28], which included samples before drug treatment (n = 15) and after Osimertinib treatment (n = 10, Fig. 4B). Specifically, the proportion of T cells significantly increased in the post-treatment samples compared to the naive baseline (Fig. 4C), suggesting an enhanced immune infiltration. Furthermore, using bulk RNA-seq data of paired patient samples from [29], we assessed the functional state of the immune microenvironment. Although the limited sample size restricted statistical power, the GSVA score for the "T CELL MEDIATED CYTOTOXICITY" gene set exhibited a clear increasing trend following Osimertinib treatment (Fig. 4D). Consistent with this, the TIDE score [30], a metric indicative of immune evasion, showed a noticeable decreasing trend in post-treatment (Fig. 4E). Despite the lack of formal statistical significance, these consistent directional shifts across independent metrics further support the notion that Osimertinib treatment is associated with transcriptional shifts indicative of a hotter TME, underscoring the reliability of our database for drug repurposing and immunological research.

Fig. 4.

Case study: Screening for drugs that could potentially induce a hotter tumor immune microenvironment. (A) Drug enrichment results based on three gene signatures. The three gene sets are (from top to bottom): the antigen processing and presentation gene set (KEGG pathway from MSigDB), the Lawson's immune evasion downregulated gene set, and the immune evasion upregulated gene set (immune checkpoint molecule ligands and the gene set proposed by Lin et al.). Drugs that positively enrich the first two gene sets and negatively enrich the last gene set were identified via GSEA (see Methods for details). The color represents the normalized enrichment score (NES), and each small square represents a treated cell line. (B) UMAP visualization of single-cell RNA sequencing data from Maynard et al., keeping samples without drug treated and after Osimertinib treatment. (C) Comparison of T cell proportions between naive and post-treatment samples in the single-cell data from B. (D) Comparison of GSVA scores for the T CELL MEDIATED CYTOTOXICITY gene set between paired patient samples before and after Osimertinib treatment. (E) Comparison of TIDE scores (indicative of the immune evasion) between paired patient samples before and after Osimertinib treatment. The bulk RNA-seq data in D-E was obtained from Roper N et al. P-values in D and E were calculated using the paired Wilcoxon rank sum test.

Drug-induced shift of alternative splicing

To characterize the global impact of drug treatment on AS in cancer cells, we systematically identified differential AS events between drug-treated and untreated control groups. Using stringent criteria (absolute delta percent spliced in (PSI) > 0.1 and BH-adjusted P < 0.05), we quantified the number of significant differential AS events across cell lines. We observed widespread drug-induced AS alterations, with the distribution varying across cell lines of different tissue origins and drug categories (Fig. 5A). Skipped exon (SE) is the most prevalent differential AS type (Figure S1). To evaluate the global directional shift of AS induced by specific drugs, we compared the proportional differences between positive and negative delta PSI for each AS type across samples. To ensure statistical robustness, we focused on drugs tested in at least three cell lines. Among the top 5 drugs showing the most significant imbalances between percentages of AS events with positive and negative delta PSI (Fig. 5B, full drug list with at least one significant AS type are shown in Figure S2), Indisulam ranked first. This is consistent with its well-established role as a splicing modulator [10], thereby could serve as a positive control for our analytical pipeline. As expected, Indisulam treatment induced a pronounced imbalance between positive and negative delta PSI, particularly for the retained intron (RI) and SE types across five cell lines (Fig. 5C).

Fig. 5.

Landscape of drug-induced alternative splicing reveals KB-0742 as a potential splicing modulator. (A) Distribution of the number of significant differential alternative splicing (AS) events between drug-treated and control groups. The top panel shows cell lines grouped by tissue origin, and the bottom panel shows drug categories. Splicing events with an absolute delta PSI > 0.1 and a Benjamini–Hochberg (BH) adjusted P-value < 0.05 were defined as significant differential AS events. (B) Comparison of the proportional differences between positive and negative delta PSI for drug-induced differential AS types across samples for each drug, showing the top 5 drugs. P-values were calculated using the Wilcoxon rank-sum test. The analysis was restricted to drugs tested in at least three cell lines. Drugs are ranked by number of significant differential AS types and then max difference between percentages of AS events with positive and negative delta PSI. (C) Proportional differences between positive and negative delta PSI for retained intron (RI) and skipped exon (SE) AS types across five Indisulam-treated cell lines. (D) Proportional differences between positive and negative delta PSI for RI and SE AS types in KB-0742-treated cell lines. (E) GSEA results for gene expression between KB-0742-treated and untreated control samples using the KEGG_SPLICESOME pathway. All treatment conditions are shown. (F) KEGG pathway enrichment results based on the hypergeometric test for genes with differential RI-type AS events in KB-0742-treated samples. Pathways with a BH-adjusted P-value < 0.05 are shown.

Interestingly, KB-0742 ranked second, suggesting its previously unappreciated potential as a splicing modulator. Similar to Indisulam, KB-0742 treatment led to a significant directional shift in RI and SE events (Fig. 5D). To further elucidate the transcriptional consequences of KB-0742 treatment, we performed GSEA comparing KB-0742-treated samples to untreated controls. Remarkably, the KEGG_SPLICESOME pathway was significantly enriched across all KB-0742 treatment conditions (Fig. 5E), providing transcriptomic evidence that KB-0742 globally disrupts splicing machinery. Furthermore, to investigate the functional impact of the AS alterations triggered by KB-0742, we extracted genes exhibiting differential RI-type AS events (which has the largest number, Figure S3) and subjected them to KEGG pathway enrichment analysis. The results revealed that these genes were significantly enriched in multiple genetic information processing pathways, including transcription, translation, protein folding, and degradation (Fig. 5F), indicating that KB-0742-induced splicing dysregulation broadly impacts central dogma processes. We further experimentally validated this finding by treating K562 cells with KB-0742, which confirmed the significant directional shift in RI and SE events and the enrichment of the KEGG_SPLICESOME pathway (Figure S4).These findings collectively suggest that KB-0742 may act as a global splicing modulator, and that its associated aberrant splicing correlates with widespread disruptions of essential cellular pathways.

Characterization of potential neoantigen derived from drug-induced differential alternative splicing

Following the observation of widespread drug-induced AS events, we next investigated their potential to generate novel immunogenic targets by predicting HLA-binding peptides derived from these differential AS events. Overall, we identified a substantial repertoire of drug-induced HLA-binding peptides across the dataset. The abundance of these peptides varied across samples depending on the tissue origin and the specific drug treatment (Fig. 6A). To identify the most potent modulators of splicing-derived antigenicity, we ranked the drug treatments by the number of differential AS-derived HLA-binding peptides (Fig. 6B, upper panel). Consistent with our previous findings, Indisulam, which induced the most profound global AS alterations, ranked first (Fig. 6B, upper panel). Notably, we observed a strong positive correlation (Pearson R = 0.86) between the number of differential AS events and the number of HLA-binding peptides across drug treatments (Figure S5), indicating that the raw count of binding peptides is largely driven by the overall splicing alteration magnitude. Given this inherent bias, we normalized the peptide counts by the total number of differential AS events per treatment to evaluate the splicing-to-antigen conversion efficiency. Strikingly, after normalization, onvansertib and its combinations emerged as top-ranking drug treatments (Fig. 6B, lower panel), despite its relatively modest impact on global AS shit in previous analyses. This suggests a distinct class of drugs with high splicing-to-antigen conversion rates, the antigenic potential of which is masked when relying solely on raw counts. To rigorously validate this higher conversion efficiency of onvansertib and rule out confounding effects from differing HLA genotypes across cell lines, we restricted our analysis to the shared HLA alleles between indisulam- and onvansertib-treated cell lines. By calculating the ratio of binding peptides to their originating differential AS events for these shared alleles, we found that onvansertib indeed exhibits a substantially higher splicing-to-antigen conversion efficiency compared to indisulam (Figure S6).

Fig. 6.

Overview of HLA-binding peptides derived from drug-induced differential alternative splicing. (A) Distribution of the number of HLA-binding peptides across samples from different tissue origins and treated with different drug types. (B) Top 10 drug treatments by the number of HLA-binding peptides derived from differential alternative splicing. The upper panel shows the raw counts, and the lower panel shows the peptide counts normalized by the number of differential alternative splicing events. Only drugs tested in at least three cell lines were considered. (C) Comparison of the distribution of HLA-binding peptides derived from different types of differential alternative splicing. P-values were calculated using the Wilcoxon rank-sum test. (D) Alluvial plots showing the proportions of peptides derived from different differential splicing types, the proportions of HLA-binding versus non-HLA-binding peptides, and the proportion of HLA-binding peptides within each splicing type. (E) Distribution of differentially spliced-derived HLA-binding peptides that bind to the three HLA alleles with the highest population frequencies across different drug treatments. Peptides derived from differentially expressed genes (with log2 fold change > 1) were exclusively included. Furthermore, only drugs tested in at least three cell lines were considered and the median number of binding peptides across cell lines is shown. Gray color indicates the absence of the specific HLA allele in the samples treated with the corresponding drug.

We then compared the HLA-binding peptide yield across different types of AS events. Our analysis revealed significant differences in the number of binding peptides generated from various differential splicing types (Fig. 6C). Specifically, SE events yielded the highest number of binding peptides, whereas alternative 5′ splice sites (A5SS) events yielded the fewest. Interestingly, while SE contributed the most peptides in total, RI events exhibited the highest proportion of binding peptides (2.12 % versus 2.09 % for SE, Fig. 6D). This observation parallels our drug-level findings, reinforcing that the AS event type with the greatest absolute output may differ from the one with the highest antigenic conversion potential.

Ultimately, to screen for drugs with the greatest potential to enhance immunotherapy, we restricted our analysis to peptides binding to the three HLA alleles with the highest population frequencies [31], thereby focusing on candidates with broad patient coverage (Fig. 6E). Notably, the most effective drug varied drastically depending on the specific HLA allele: the PLX4720+PD0325901 combination treatment yielded the most binding peptides for HLA-A02:01, decitabine for HLA-A24:02, and doxorubicin for HLA-C*04:01. To further validate the immunomodulatory potential of these prioritized drugs in human tumors, we focused on decitabine. Analysis of single-cell RNA-seq data from an AML patient (scCT-DB database [32]) revealed that decitabine treatment increased the proportion of cytotoxic CD8+ T cells and significantly elevated T cell cytotoxicity scores (Figure S7A-D). This in vivo immunostimulatory effect was further corroborated by bulk RNA-seq data from seven patient-derived gliomaspheres [33], which demonstrated higher cytolytic activity (CYT [34]) and T CELL MEDIATED CYTOTOXICITY scores in the decitabine-treated group compared to controls (Figure S7E-F). Finally, to provide direct evidence of drug-induced antigen presentation, we leveraged mass spectrometry-based HLA immunopeptidome data from three glioblastoma cell lines [35]. The results revealed a significant increase in the number of presented HLA-binding peptides following decitabine treatment (Figure S7G). These validations indicate that these drugs may serve as prioritized candidates for future experimental validation of combination treatment strategies with immunotherapy.

Development of user-friendly database and web server

To facilitate the access and analysis of the data within the DRIVE database, we developed a user-friendly web interface. The interface primarily consists of four functional modules: Transcriptome Response, Drug Enrichment, Connectivity Analysis, and AS-derived Neoantigen. In the Transcriptome Response module, users can select a specific drug and cell line to display the information of the corresponding drug-treated samples (Fig. 7A). This module presents three aspects of data: Metadata, Gene Expression, and Alternative Splicing. The Metadata section displays detailed information regarding the cell lines and drug treatments. The Gene Expression section presents the results of differential expression analysis, including a gene list and a volcano plot. The Alternative Splicing section illustrates the results of differential AS analysis, which includes a list of differential AS events (FDR < 0.05), a statistical summary of the number of significantly differential AS types, the top 10 genes with the highest number of significant differential AS events, and the PSI difference between the treatment and control groups for a user-selected gene. In the Drug Enrichment module, users can upload a gene list (either by direct text input or by uploading a gmt/txt file). We perform GSEA based on the user-uploaded gene set to identify drug-cell line samples that significantly enrich the given gene set (Fig. 7B). This module displays a table of the enrichment results alongside the top 10 positively and negatively enriched drugs.

Fig. 7.

Functional modules of the DRIVE database web interface. (A) Transcriptome Response module. This module includes Metadata, which displays detailed information on cell lines and drug treatments; Gene Expression, which presents the results of differential expression analysis; and Alternative Splicing, which shows the results of differential alternative splicing analysis. (B) Drug Enrichment module. This module performs GSEA based on a user-uploaded gene set against the database and displays the corresponding results. (C) Connectivity Analysis module. This module calculates the similarity between a user-provided query signature and all drug-induced gene expression signatures in the database, and displays the corresponding results. (D) AS-derived Neoantigen module. This module presents information on drug induced significantly differentially AS-derived HLA-binding peptides.

Gene signature connectivity analysis connects disease-related gene expression profiles (signatures) with perturbation signatures (drugs, genetic reagents) to uncover MoA, drug repurposing opportunities, or disease pathways [36]. Positive correlations between two signatures indicate there are likely shared MoA, whereas negative correlations imply that the compound of the reference signature can be used to reverse the query signature, and hence, can be considered as candidate drugs to treat the latter. In the Connectivity Analysis module, users can provide a gene signature file (containing gene names and a fold change column). The system then screens the database for drug-induced gene expression signatures that exhibit high similarity to the user-provided query signature (Fig. 7C). We calculated six similarity metrics including Xsum (eXtreme Sum) [37], CSS (Connection Strength Score) [38], WTCS (Weighted Connectivity Score) [39], Spearman correlation, Pearson correlation, and Cosine similarity (Methods). This module presents a table of these similarity metrics between all drug-induced signatures in database and the query signature, as well as radar charts for the top-ranked positively and negatively correlated drug-induced signatures.

In the AS-derived Neoantigen module, users can select a drug and a cell line to retrieve HLA binding peptides information derived from significantly differential AS events (FDR < 0.05 and absolute delta PSI > 0.1) in the corresponding drug-treated samples (Fig. 7D). It displays a table of predicted %Rank scores for binding peptides across HLA alleles, along with a detailed information table for each binding peptide (including the source rMATS ID, rMATS type, and gene). Furthermore, we provide a ternary relationship plot among alternative splicing, differential expression, and binding peptides, allowing users to efficiently filter and identify differential AS events of interest and their resulting peptides based on both the plot and the tables. Collectively, the DRIVE web interface provides an interactive and comprehensive platform for the scientific community to systematically explore drug-induced transcriptomic alterations, alternative splicing dynamics, and the subsequent emergence of AS-derived potential neoantigens.

Discussion

Pharmacotherapy remains the cornerstone of cancer treatment, and a comprehensive understanding of drug-induced molecular reprogramming is essential for improving therapeutic efficacy and overcoming resistance. In this study, we constructed DRIVE, a comprehensive resource integrating drug-induced transcriptomic responses, AS alterations, and splicing-derived neoantigens in cancer cell lines. By systematically mining and uniformly processing publicly available GEO data, DRIVE bridges a critical gap by linking transcriptomic shifts, splicing dynamics, and immunogenic potential, thereby providing a multi-omic perspective for drug repurposing and the design of synergistic immunotherapeutic strategies.

A key contribution of our work lies in the rigorous and standardized curation of drug perturbation data specifically in cancer cell lines. The rationale for focusing on cell lines, rather than primary patient samples, is rooted in their relatively clean genetic background and homogeneous nature, which minimizes the confounding effects of tumor stromal heterogeneity and diverse immune cell infiltrations. More importantly, in vitro cell line models provide well-matched drug-treated and untreated control groups, enabling precise isolation and quantification of direct drug-induced transcriptomic and splicing changes, a prerequisite that is often unmet in clinical cohorts lacking matched pre- and post-treatment biopsies. Nevertheless, we acknowledge that the simplicity of cell line models inherently limits the investigation of complex tumor TME dynamics. As such, in future updates, we plan to incorporate clinical transcriptomic drug response data from patient cohorts, which will be invaluable for deciphering drug-induced TME remodeling and immune cell interactions in vivo.

Recently, the PharmGEO database [40] was published to archive drug-induced gene expression profiles. However, DRIVE offers several distinct and significant advantages over PharmGEO. First, PharmGEO lacks a specific focus on cancer cell lines, which are the most fundamental models for preclinical pharmacological research and mechanism elucidation. Second, PharmGEO suffers from incomplete meta-information curation, which severely hampers the reliable grouping of samples for downstream differential analysis. In contrast, we leveraged LLMs combined with meticulous manual inspection to extract high-fidelity structural metadata. Third, and most critically, PharmGEO directly aggregates previously processed data without standardizing the raw data from scratch, inevitably introducing computational biases. DRIVE overcomes this by implementing a uniform pipeline starting from raw FASTQ files. Finally, PharmGEO is limited to gene expression changes, completely overlooking the widespread reprogramming of alternative splicing and the subsequent emergence of splicing-derived neoantigens, two dimensions that DRIVE uniquely integrates.

The systematic profiling of drug perturbation data in DRIVE also holds significant implications for the emerging field of virtual cell modeling. Simulating cellular responses to pharmacological interventions is a grand challenge and a central theme in the development of virtual cells [41]. However, current computational models for drug perturbation are predominantly trained on the LINCS L1000 dataset [42,43]. A major limitation of L1000 is that it only measures the expression of approximately 1,000 landmark genes, failing to capture transcriptome-wide resolution, let alone the complex isoform-level dynamics. The high-quality, transcriptome-wide drug perturbation data provided by DRIVE can be utilized to benchmark existing drug perturbation models, evaluating their ability to predict not only gene expression changes but also splicing alterations under pharmacological stress. Integrating DRIVE into the training and evaluation of virtual cell models to achieve transcriptome-wide and isoform-level predictive simulations represents a promising direction for future research.

Leveraging the DRIVE database, we uncovered several novel biological insights. We demonstrated that structurally analogous compounds, especially those sharing deep chemical language representations (MoLFormer embeddings), are more likely to induce convergent transcriptomic reprogramming, providing computational evidence for the interplay between chemical space and transcriptomic space. More importantly, we generated testable hypotheses identifying potential immunomodulatory agents (e.g., Osimertinib) and novel splicing modulators (e.g., KB-0742). Additionally, our HLA-binding peptides analysis revealed several drugs could induce the production of potential neoantigens derived from differential alternative splicing and bound to high-frequency HLA, showing significant clinical application prospects.

Despite the robust bioinformatic evidence and the use of independent public datasets for validation, our study has notable limitations. The most significant limitation is the lack of experimental validation for the novel drug candidates identified through our pipeline. For instance, the potential of Osimertinib to transform the tumor microenvironment into a "hot" state, the mechanism of KB-0742 as a global splicing modulator, and the immunogenicity of drug-induced splicing-derived neoantigens require rigorous wet-lab validation. Future studies should employ in vitro and in vivo models, such as co-culture systems and syngeneic mouse models, to confirm the immunomodulatory effects and the bona fide presentation of these predicted neoantigens by MHC class I molecules. In addition, in the Jcast prediction step, we selected the peptide isoform based on the transcript-level inclusion and skipping counts, which is essentially an approximate estimation rather than a direct reflection of protein abundance. To accurately measure the protein abundance of the isoforms will require quantification using mass spectrometry (MS). Furthermore, while we propose that DRIVE can identify potential vaccine targets and inform synergistic immunotherapeutic strategies, we acknowledge that this claim is currently based solely on in silico predictions. The pipeline relies on HLA-binding affinity predictions, which does not guarantee actual peptide processing, MHC presentation, or the elicitation of CD8+ T cell responses in vivo. Therefore, at its current stage, DRIVE should be viewed as a hypothesis-generating and prioritization tool rather than a definitive platform for clinical vaccine design. The drug-induced splicing-derived neoantigens identified here require rigorous experimental validation, such as MS-based immunopeptidomics and T cell activation assays. Nevertheless, our analysis still provides a rich resource for subsequent experimental validation of alternative splicing events of interest and the resulting peptides.

Conclusions

In conclusion, DRIVE represents a valuable and comprehensive resource that deciphers the intricate connections between drug perturbations, transcriptomic reprogramming, alternative splicing, and neoantigen emergence. By providing a standardized dataset and a user-friendly web server, DRIVE not only advances our understanding of drug mechanisms of action but also provides a computational framework for prioritizing novel vaccine targets, generating testable hypotheses for rational combination immunotherapies and provides a solid data infrastructure for building more powerful virtual cell models.

Methods

Data collection and filtering

We initially queried the Gene Expression Omnibus (GEO) database using the keywords "RNA seq AND drug AND cancer", which yielded 3,224 relevant results. These datasets were subjected to manual curation based on the following criteria: (1) explicit drug treatment information; (2) the use of cancer cell lines; (3) a minimum of two biological replicates; and (4) the availability of raw next-generation sequencing (NGS) data. Following this initial screening, we downloaded the raw data, obtaining 6,512 SRR files representing 5,047 samples, with a total data volume of approximately 12 TB. Subsequently, a more stringent filtering step was performed to exclude non-human cell lines and cell lines with non-wild-type genetic backgrounds (including those subjected to gene knockdown/knockout and drug-resistant cell lines). This final curation resulted in a high-quality dataset comprising 3,911 samples, derived from 243 GEO datasets, encompassing 272 distinct cancer cell lines and 278 unique drug treatments.

To validate the potential immuno-microenvironment modulating drugs identified from our screening, we collected scRNA-seq data from [28]. This dataset comprises 7,169 cells from 15 untreated and 10 Osimertinib-treated samples. Cell type annotations provided by the original authors were utilized, and dimensionality reduction was performed using the RunUMAP function in the Seurat package. Furthermore, we obtained bulk RNA-seq data from [29], which included paired pre- and post-Osimertinib treatment samples from 10 patients.

Meta-information extraction and transcriptome data processing

The data processing procedure consists of four modules: raw data download and processing, metadata extraction, gene expression and alternative splicing quantification, and HLA-I binding prediction of AS-derived peptides (Fig. 2).

Raw data download and processing

After identifying the target GEO datasets, we first obtained their SRA Study accessions and downloaded the SRA files using the prefetch tool from the SRA Toolkit. Subsequently, fasterq-dump was utilized to convert the SRA files into FASTQ format. The raw FASTQ files were then pre-processed using fastp, which included quality control, removal of low-quality reads, adapter trimming, and read trimming. The cleaned reads were aligned to the human reference genome (GRCh38) using STAR [44], generating BAM files.

Metadata extraction

The metadata required for our database included dataset IDs, cell line information, drug information, and specific drug treatment details (e.g., treatment duration and concentration). Since these details were not available as structured data, we leveraged the information extraction capabilities of LLMs combined with meticulous manual curation. Initially, the pysradb package [45] was used to retrieve basic metadata, such as GEO sample records, species, and sequencing strategy (single-end or paired-end). Next, we employed Python scripts to batch-download the textual descriptions of the samples from the GEO database. The Qwen3 [46] LLM (version Qwen3-30B-A3B-Instruct-2507), guided by prompt engineering, was utilized to extract key information from these texts, including cell line names, drug treatments, treatment concentrations, and durations (The prompt is available at the https://github.com/ComOncLab/DRIVE). Finally, manual curation was performed to correct potential errors, yielding the structured metadata for the samples in our database.

Gene expression and alternative splicing quantification

Salmon [47] was employed to quantify transcript-level abundance from the RNA-seq data, using a GRCh38-based index. Subsequently, tximport [48] was used to summarize the transcript-level abundance estimates for downstream transcript- and gene-level analyses. Differential gene expression analysis between the drug treatment and control groups was performed using DESeq2 [49] based on the tximport output, thereby capturing the drug-induced transcriptomic responses. In addition, rMATS [50] was utilized to detect differential AS events between the drug treatment and control groups, generating a comprehensive profile of drug-induced AS alterations.

HLA-I binding prediction of AS-derived peptides

To obtain the Class I HLA allele information for each cell line, we first retrieved HLA annotation from the Cellosaurus database [51]. If HLA information was unavailable for a specific cell line in the database, we performed HLA typing directly from the RNA-seq data using arcas-hla [52]. The HLA-I alleles (focusing on HLA-A, -B, and -C) obtained from both sources were then manually reviewed and curated. Next, based on the rMATS output (keep events with FDR < 0.05 and absolute delta PSI > 0.1), Jcast [53] was used to predict the protein sequences of AS isoforms that exhibited significant differential changes between the drug treatment and control groups. For each AS type, Jcast produces two protein sequence variants (Type 1 and Type 2) representing the inclusion and skipping isoforms, respectively. Thus, to determine the predominant isoform for each differential AS event, Type 1 is selected when the mean inclusion count is elevated in the treatment group compared to the control group; conversely, Type 2 sequence is selected if the mean skipping count shows an increase upon drug treatment. These protein sequences were subsequently digested in silico into 8–11 mer peptides. Finally, NetMHCpan4.1 [54] was applied to predict the binding affinity between HLA-I molecules and the peptides. We predicted binding scores for both Binding Affinity (BA) and Eluted Ligand (EL) modes. We define peptides that bind to HLA as potential binders if both the predicted %Rank_BA and %Rank_EL scores for at least one HLA allele are <0.5 (or %Rank < 2 for exploratory analysis in the Shiny web server).

Transcriptome response similarity and drug similarity calculation

To quantify the similarity of transcriptome responses between samples, we extracted the the Wald statistic (stat) column from the DESeq2 differential expression results and calculated pairwise sample correlations using the kendall_cor function from the R package kendallknight [55]. Kendall's tau correlation is a non-parametric rank correlation coefficient that measures the strength and direction of the monotonic relationship between two variables. Given its robustness to outliers, it is particularly suitable for assessing the similarity of transcriptomic changes.

We employed two distinct approaches to calculate inter-drug similarities. The first approach involved computing the Tanimoto similarity based on the SMILES representations of the molecules. Briefly, the SMILES strings were parsed into molecule objects using the parse.smiles function in the rcdk package, and molecular fingerprints were subsequently generated via the get.fingerprint function. Finally, the Tanimoto similarity matrix among these fingerprints was computed using the fp.sim.matrix function from the fingerprint package. The second approach utilized MoLFormer [18] to derive 768-dimensional embedding vectors from the drug SMILES strings, and the cosine similarity between these vectors was then calculated.

Gene set enrichment analysis

Gene set enrichment analysis (GSEA) was performed using the R package fgsea [56]. To rank the genes, the Wald statistic (stat) column derived from the DESeq2 differential expression analysis was utilized. Immune escape and antigen presentation machinery, namely APM, was obtained from the MSigDB database. Down_Immune_Escape gene set was download from supplementary table in [19]. This study identify genes that enable cancer cells to evade killing mediated by cytotoxic T lymphocytes (CTLs) by performing genome-wide CRISPR screens across a panel of genetically diverse mouse cancer cell lines that were cultured in the presence of CTLs. UP_Immune_Escape gene set was downloaded from LIN_TUMOR_ESCAPE_FROM_IMMUNE_ATTACK in MSigDB. This gene set contains genes up-regulated in highly immune-resistant cancer cell line developed from a susceptible cancer using an in vivo selection strategy [20]. Regarding the spliceosome analysis, the KEGG_SPLICESOME gene set from the MSigDB database was employed. Additionally, over-representation analysis based on the hypergeometric test was conducted using the enrichKEGG function from the clusterProfiler package.

Cell culture and RNA-seq experiments

KB-0742-treated and DMSO control groups were established, with three independent biological replicates included in each group. For each biological replicate, approximately 1 × 10^6 K562 cells were seeded and cultured in RPMI 1640 complete medium. After the cells reached a stable growth condition, cells in the treatment group were exposed to KB-0742 at a final concentration of 1.2 μM, whereas cells in the control group were treated with an equal volume of DMSO. Cells were harvested after 8 h of treatment. After treatment, the cell suspension was transferred into centrifuge tubes and centrifuged at 500 g for 5 min to collect the cells. The supernatant was discarded, and the cells were gently resuspended in PBS, followed by another centrifugation at 500 g for 5 min. After removal of the PBS supernatant, an appropriate volume of TRIzol reagent was added to the cells for complete lysis. The lysed samples were used for RNA extraction, RNA-seq library preparation, and transcriptome sequencing analysis. Libraries were sequenced on the illumina platform. The previously mentioned analytical pipeline was used to process the data obtained from RNA-seq sequencing.

Implementation of database and web server

The Shiny application primarily consists of four modules: Transcriptome Response, Drug Enrichment, Connectivity Analysis, and AS-derived Neoantigen. The Connectivity Analysis module, inspired by the Connectivity Map (CMap) database, calculates the similarity between a user-provided query signature and multiple drug-induced transcriptome profiles. Adopting the strategy and source code from the WebCMap R package [57], we computed six similarity metrics: eXtreme Sum (XSum), Connection Strength Score (CSS), Weighted Connectivity Score (WTCS), Spearman correlation, Pearson correlation, and Cosine similarity. For CSS and XSum, empirical P-values are determined using a permutation test. Similar to the approach used in WebCMap, we pre-generated 10,000 random sets for each of the 931 signatures in our DRIVE dataset and calculated their corresponding CSS and XSum values in advance, thereby constructing the null distributions required for the empirical P-value calculations. The database and platform were developed using R version 4.4.2 and Shiny version 1.10.0. The Shiny application is hosted on a cloud server equipped with an Intel(R) Xeon(R) CPU E5-2697 v2 with 8 cores and 16 GB of memory.

Statistical analysis

Statistical analysis was performed by R software. Wilcoxon rank-sum test was performed to compare two groups of independent samples. A P value of <0.05 was considered statistically significant. Other statistical details of each analysis can be found in the Figure legend or the Method details section.

Availability of data and code

Data for building DRIVE database and web-server has been deposited at Zenodo (10.5281/zenodo.19816802). All data can be viewed online in https://componclab.com/DRIVE. Codes required to reproduce the analysis outlined in this manuscript are available in https://github.com/ComOncLab/DRIVE/, and analysis report are available online in https://comonclab.github.io/DRIVE/. The code for developing Shiny is also available in https://github.com/ComOncLab/DRIVE.

Funding

This work is supported by startup funding of postdoctoral project from Yangjiang People's Hospital.

Appendix. Supplementary materials.

Supplementary Figure S1-S7.

CRediT authorship contribution statement

Tao Wu: Writing – review & editing, Writing – original draft, Visualization, Validation, Software, Project administration, Methodology, Investigation, Funding acquisition, Formal analysis, Data curation, Conceptualization. Hong-Feng Tang: Writing – review & editing. Wei-liang Wang: Writing – review & editing, Supervision.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.