In silico RNA isoform screening to identify potential cancer driver exons with therapeutic applications

Anglada-Girotto, Miquel; Ciampi, Ludovica; Bonnal, Sophie; Head, Sarah A.; Miravet-Verde, Samuel; Serrano, Luis

doi:10.1038/s41467-024-51380-z

In silico RNA isoform screening to identify potential cancer driver exons with therapeutic applications

Article
Open access
Published: 15 August 2024

Volume 15, article number 7039, (2024)
Cite this article

Download PDF

You have full access to this open access article

From

View current issue

In silico RNA isoform screening to identify potential cancer driver exons with therapeutic applications

Download PDF

3003 Accesses
147 Altmetric
18 Mentions
Explore all metrics

Abstract

Alternative splicing is crucial for cancer progression and can be targeted pharmacologically, yet identifying driver exons genome-wide remains challenging. We propose identifying such exons by associating statistically gene-level cancer dependencies from knockdown viability screens with splicing profiles and gene expression. Our models predict the effects of splicing perturbations on cell proliferation from transcriptomic data, enabling in silico RNA screening and prioritizing targets for splicing-based therapies. We identified 1,073 exons impacting cell proliferation, many from genes not previously linked to cancer. Experimental validation confirms their influence on proliferation, especially in highly proliferative cancer cell lines. Integrating pharmacological screens with splicing dependencies highlights the potential driver exons affecting drug sensitivity. Our models also allow predicting treatment outcomes from tumor transcriptomes, suggesting applications in precision oncology. This study presents an approach to identifying cancer driver exon and their therapeutic potential, emphasizing alternative splicing as a cancer target.

Splicing isoform-specific functional genomic in cancer cells

Article Open access 14 December 2018

RNA splicing dysregulation and the hallmarks of cancer

Article 10 January 2023

Alternative splicing: a new breakthrough for understanding tumorigenesis and potential clinical applications

Article 19 January 2023

Introduction

Cancer originates from the concurrent appearance of cancer driver genomic alterations, such as mutations, copy number variations, and chromosomal rearrangements, that activate oncogenes and inactivate tumor suppressor genes^1,2,3. These genetic changes affect non-genetic molecular regulatory layers, including chromatin accessibility, gene expression, alternative splicing, and translation efficiency, to reprogram the cell toward abnormal proliferation^{4,5,6,7,8,9,10,11,12,13}. Hence, both genetic and non-genetic alterations contribute to cancer initiation and progression and thereby constitute important therapeutic opportunities. However, while identifying cancer driver genomic alterations is well established, the systematic identification of cancer driver non-genetic alterations remains unaddressed.

Alternative splicing is a key regulatory process that extensively contributes to the molecular and functional diversity of the cell. Through the activity of over 300 splicing factors, human cells select which combination of exons to use depending on the physiological conditions⁸. This regulation is crucial, as it can have cascading effects on downstream molecular levels, such as mRNA stability, translation, protein activation, and protein-protein interactions¹⁴. In cancer, splicing is often dysregulated to favor tumor progression^9,10. So far, the ASCancer Atlas¹⁵ compiled 814 exons from the literature with experimentally validated contributions to tumor progression. Splicing dysregulation in cancer can be targeted by different drugs including splice-switching oligonucleotides (SSOs), which can be designed to disrupt specific splicing events through base pairing with a gene’s pre-mRNA¹⁶, representing a promising strategy for targeting cancer driver splicing events. However, the identification of such events relied on laborious, case-by-case experimental validations, leaving their potential as cancer drivers largely underexplored.

Gonatopoulos-Pournatzis et al.¹⁷ and Thomas et al.¹⁸ used CRISPR-based screens to genomically delete a maximum of ~2,000 different exons of the 143,592 annotated exons¹⁹. They observed decreased cell proliferation upon deletion of 124 exons in epithelial-derived RPE1 cells¹⁷, 339 exons in cervix-derived HeLa cells, and 152 exons in lung-derived PC9 cells¹⁸. On the other hand, they observed increased cell proliferation upon deletion of 70 or 36 exons in the cervix-derived HeLa cells or lung-derived PC9 cells, respectively¹⁸. These exons constitute potential cancer driver exons, as their perturbation alters cell proliferation. Even though this approach is powerful, extending the list of tested exons remains experimentally challenging. This underscores the importance of developing new approaches to assess the functional impact of perturbing splicing isoforms.

While conducting genome-wide screens for cancer driver splicing exons poses challenges, gene-level screens are commonly used to investigate the contribution of individual genes to cancer phenotype hallmarks^20,21. For example, the effect of decreasing the expression of individual genes with short hairpin RNA (shRNA) on cancer cell proliferation has been largely investigated²². Measuring the fitness effects upon gene silencing (hereafter, gene dependencies) not only helps to identify important genes for cell proliferation but also can be repurposed to understand the mechanisms of action of drugs²³.

Cancer cell lines exhibit variability in splicing isoforms of the same gene and in their dependency on specific genes following knockdown (KD). When shRNA KD screens target the entire pool of isoforms of a gene, diverse responses in different cell lines may result from variations in the inclusion levels of specific exons. We propose that exons whose inclusion is associated with gene dependency are potential cancer driver exons. Thus, gene dependency screens can be repurposed to identify splicing dependencies, revealing critical splicing events and potential drug-splicing synergies.

Here, we identify 1073 potential cancer driver exons by statistically associating exon inclusion with their corresponding gene dependency. Many of these exons belong to low-mutation genes within cancer driver pathways, suggesting alternative splicing as a parallel oncogenic mechanism. Our statistical models predict outcomes of splicing perturbations, enabling in silico RNA isoform screening. Across 13 cancer types, we prioritize and validate eight exons as potential therapeutic targets. Additionally, inferred splicing dependencies align with drug sensitivity profiles, creating models for drug-exon interactions. Our computational framework, spotter, is available at https://github.com/MiqG/target_spotter, with data accessible at https://spotter.crg.eu.

Results

Genome-wide identification of high-confidence potential cancer driver exons

To identify exons that influence cell proliferation statistically, we hypothesized that potential cancer driver exons are those whose inclusion, measured as percent spliced in (PSI), correlates with their corresponding gene dependency (see Methods). For instance, the inclusion of two experimentally validated cancer driver exons in the genes KRAS (HsaEX0034998)²⁴ and SMNDC1 (HsaEX0060482)¹⁸ (VastDB nomenclature¹⁹) correlate with their gene dependency (p = 2.3 × 10^–9, Spearman correlation = –0.23, and p = 0.00019, Spearman correlation = 0.15, respectively) (Supplementary Fig. 1a, b). However, the expression of the genes, measured as transcript per million (TPM), also correlates strongly with gene dependencies (Supplementary Fig. 1a, b). This indicates that, although exon inclusion correlates significantly with gene dependencies, we should account for gene expression when identifying potential cancer driver exons. While exon-gene dependency may be influenced by various molecular regulatory layers beyond gene expression, we intentionally refrained from incorporating additional measurements. This decision was made to assess the effectiveness of simpler models, which could have broader applicability.

Gene dependencies quantify gene contributions to cell proliferation, pinpointing potential cancer driver genes as well as pan-essential ones with limited therapeutic potential due to their fundamental role in viability across all cell types²⁵. Therefore, we evaluated the intersection of gene sets comprising known cancer driver exons from a hand-curated list and the ASCancer Atlas¹⁵, along with pan-essential genes sourced from the Cancer Dependency Map and cancer driver genes identified through genetic alterations in COSMIC’s Cancer Gene Census (CGC)^26,27 (Supplementary Data 1; see Methods). Only 157 out of 634 (24.7%) genes with known cancer driver exons are listed as pan-essential genes (n = 73), mutational cancer driver genes (n = 70), or both (n = 14) (Fig. 1a). Comparing the overlaps of each gene set with biological pathways reveals a high pathway similarity between genes with known cancer driver exons and mutational cancer driver genes (p < 2.2 × 10⁻¹⁶, Pearson correlation = 0.88) compared with pan-essential genes with the former (p = 4.4 × 10⁻¹¹, Pearson correlation = 0.62) and the latter (p = 4.7 × 10⁻⁷, Pearson correlation = 0.5) (Supplementary Fig. 1c; Supplementary Data 2). Considering that pan-essential and mutational cancer driver gene lists are well established and that cancer driver exons are understudied, this suggests that (i) splicing modulation of some genes not known to be pan-essential or mutational cancer genes could play a role in cancer, (ii) mutational cancer driver genes can also be targeted at the splicing level, and (iii) a potentially large number of unexplored cancer driver genes may derive from splicing alterations rather than other genetic changes.

**Fig. 1: Modeling gene dependencies to identify potential cancer driver exons.**

To find potential cancer driver exons, we defined linear models that predict gene dependency (i.e., Demeter2 score, see Methods) as a function of exon inclusion and gene expression (i.e., PSI and TPM, respectively). We did this for all 143,592 annotated exons in the VastDB human transcriptome that could be mapped to 635 cell lines from Cancer Cell Line Encyclopedia (CCLE) with data types adequately accessible (Fig. 1b; Supplementary Data 3–7; see Methods). We used likelihood ratio (LR) tests comparing full models including both gene expression and exon inclusion to a null model with gene expression only to identify exons that improved model fitting (Fig. 1b; see Methods). We then evaluated the area under the curve (AUC) for a receiver operating characteristic (ROC) analysis to assess the ability of p-values obtained from these LR tests to retrieve the known cancer driver exons, mutational cancer driver genes, and pan-essential genes (Supplementary Data 8; see Methods). Note that for mutational cancer driver gene and pan-essential gene sets, we assigned each gene with the lowest LR p-value from all exons in that gene (as applicable). LR p-values for distinguishing mutational cancer driver genes showed the best predictive power (AUC = 0.6), followed by known cancer driver exons (AUC = 0.58) and pan-essential genes (AUC = 0.55) (Supplementary Fig. 1d, left). On the other hand, LR p-values show the best precision-recall predictive power to identify known cancer driver exons (AUC = 1) followed by mutational cancer driver genes (AUC = 0.96) and pan-essential genes (AUC = 0.9) (Supplementary Fig. 1d, right). Considering that these lists only have a small overlap and that the number of annotated cancer driver exons is likely underestimated (Fig. 1a), this suggests that our approach holds the potential to discover exons of functional relevance for cancer cell proliferation.

Gene dependency scores are calculated by modeling shRNA depletion considering gene knockdown and off-target effects²². The varying number of shRNAs targeting each gene in the libraries can lead to strong associations when an exon is targeted, making cell lines without that exon resistant to shRNA knockdown. To determine if this biased our analysis, we mapped all shRNA sequences to the human transcriptome (Supplementary Data 9; see Methods). Out of 12,129 genes, 130 (1.1%) were targeted with shRNAs designed against only one of their exons (Supplementary Fig. 1e). In 27 (20.7%) of these 130 genes (i.e., about 0.02% of the total genes), the targeted exon is annotated in VastDB and could, thus, bias our analysis (Supplementary Fig. 1f). Further, LR p-value distributions of targeted and non-targeted exons are similar regardless of the number of exons targeted per gene (Supplementary Fig. 1g). This indicates that gene dependencies are robust against the number of shRNAs targeting each gene, making it adequate for our approach.

Prior to shortlisting high-confidence potential cancer driver exons, we confirmed the minimum number of reads needed for an exon to consider the value reliable (See Supplementary Fig. 2; Supplementary Data 10, 11; see Methods). We found that a threshold of ≥ 10 reads was sufficient to avoid significant PSI estimation errors. Then, we selected the significance threshold for LR p-values that simultaneously maximized the recovery of a true-positive set of cancer driver exons while minimizing the recovery of true-negative sets (LR p < 0.025) (Supplementary Fig. 3a–c; see Methods). We further filtered the remaining 3,984 exons by their model’s test set Pearson correlation. A Pearson correlation > 0.2 resulted in 1,073 exon models –corresponding to a total of 813 genes– that accurately ranked gene dependencies at single-sample resolution (Fig. 1c; Supplementary Fig. 3d; Supplementary Data 12–14; see Methods). Although these models were trained to predict gene dependencies, our approach modeled the contribution of each exon’s alternative splicing to their corresponding gene-level dependencies. Hence, from now on, we will refer to our model predictions as splicing dependencies. Overall, genes containing at least one of the shortlisted exons are enriched for functions related to the splicing machinery with a false discovery rate (FDR) < 0.05 (Fig. 1d; Supplementary Data 15; see Methods). Further, splicing of these exons is often annotated to cause ORF disruption upon exclusion (37%) or alternative protein isoform generation (35%); proportions that align with the expected background distribution (Fig. 1e; Supplementary Fig. 4a).

Predicted splicing dependencies quantify the expected change in cell proliferation by depleting an isoform at a given gene expression and exon inclusion. Based on their splicing dependency, exons are predicted to have different effects, similar to how genes can act as potential oncogenes, tumor suppressor genes, or double agents depending on how they affect cancer cells when perturbed^28,29. Thus, conceptually, potential cancer driver exons can act: (i) as oncoexons when when inclusion of the exon leads to inhibition of cell proliferation alongside KD of the corresponding gene (negative splicing dependency), (ii) as tumor suppressor exons when inclusion of the exon leads to promotion of cell proliferation when the corresponding gene is knocked down (positive splicing dependency), and (iii) as double-agent exons, when their inclusion can either inhibit or promote cell proliferation after KD of the corresponding gene, depending on the circumstances (extremely positive and negative splicing dependencies across different cancer samples). Using the models of splicing dependency, we asked how each potential cancer driver exon is expected to behave by estimating splicing dependencies for each cancer cell line in the CCLE (n = 1015). Out of 1073 explored exons, 813 (75.8%) align with our oncoexon definition, showing negative splicing dependency distributions. (Supplementary Fig. 4b; Supplementary Data 16 and 17; see Methods).

Potential cancer driver exons enlarge the universe of potential cancer driver genes

Of the set of 813 genes with shortlisted exons (Fig. 1c), only 26 (3.2%) genes carry a known cancer driver exon while 74 (9.1%) overlap with the mutational cancer driver gene set (Fig. 1c). Despite pan-essential genes typically having stable gene dependency across cell types, 241 (29.6%) show strong co-varying gene dependency with exon inclusion. As anticipated by our ROC analyzes above, associating gene dependencies with their exon inclusion prioritizes exons in true cancer driver genes as well as exons in important genes for general cell viability. However, the majority, 508 (62.5%), of the genes that carry shortlisted exons remain uncategorized. To check if uncategorized genes are related to cancer-specific pathways, we explored whether they share pathways with cancer-related and pan-essential genes. Pathway overlap frequencies of uncategorized genes are highly correlated to those of genes with known cancer driver exons (p = 3.1 × 10⁻¹¹, Pearson correlation = 0.63) and to those of mutational cancer driver genes (p = 8.1 × 10⁻⁸, Pearson correlation = 0.53) but are not correlated to those of pan-essential genes (p = 0.4, Pearson correlation = 0.089) (Fig. 1f; Supplementary Data 2). This suggests uncategorized genes with shortlisted exons specifically participate in cancer pathways and not in those of general cell viability. Thus, this expands the catalog of genes with cancer driver potential by 508 uncharted additions. As the majority of 1,073 shortlisted exons (n = 701; 65.3%) are linked to known cancer driver genes or pathways, and the pan-essentiality of the remaining 372 exons needs further validation, we label all as potential cancer driver exons due to this study’s exploratory nature.

We then wondered whether the small overlap but high pathway similarity between mutational cancer driver genes and genes harboring known or potential cancer driver exons indicates that cancer cells might use alternative splicing in parallel to mutational processes to tune gene function (Fig. 1c, f). To test this hypothesis, we quantified how often uncategorized genes undergo mutagenesis compared to the rest (see Methods). For all non-intronic types of mutation effects (n = 11), we found that the rate of mutations in genes with potential cancer driver exons was significantly lower than in genes without, regardless of whether we included or not the genes from the mutational cancer driver and pan-essential lists (p < 0.05, two-sided Wilcoxon rank sum test) (Fig. 1g; Supplementary Fig. 4c; Supplementary Data 18). This suggests that cancer cells require genes with potential cancer driver exons to be functional and that their driver activity arises from splicing regulation rather than mutagenesis.

Splicing dependency models recapitulate single and combinatorial splicing perturbations

Certain splicing factors are recurrently mutated in cancer, and this generates new splicing vulnerabilities^10,30. To showcase the value of splicing dependency analysis, we simulated how reported cancer driver exons change their sensitivity to perturbation given missense mutations in recurrently mutated splicing factors (Supplementary Fig. 4d; Supplementary Data 19; see Methods). Specifically, we calculated maximum harm scores by multiplying predicted splicing dependencies in a sample by the necessary change in inclusion to fully exclude each exon (see Methods). Notably, the experimentally validated cancer driver exon HsaEX0005606_ARHGEF11^11,31 exhibited the strongest significant decrease in expected harm score, from –0.68 to –6.75 (p = 0.00934, two-sided Wilcoxon test) with mutant RBM10, indicating that perturbing this exon in the RBM10-mutated context may lead to a stronger inhibitory effect.

To assess the predictive power of our splicing dependency models, we leveraged the results of the previous CRISPR-Cas9-based studies evaluating the impact of deleting single exons on cell proliferation^17,18. The removal of these exons caused a wide range of effects on cell proliferation, from neutral (i.e., close to 0) to strong effects (i.e., far from 0) (Supplementary Fig. 5a, b). We found a positive correlation between the experimental changes in cell proliferation and our harm scores, that is the cell proliferation effects upon splicing modulation predicted from our splicing dependency models (see Methods) for PC9 cells, as reported in Thomas et al.¹⁸ (p = 0.023, Pearson correlation = 0.56); and for RPE1 cells, as reported in Gonatopoulos-Pournatzis et al.¹⁷ (p = 0.013, Pearson correlation = 0.51) (Fig. 2b). The correlation was weaker in HeLa cells (p = 0.57, Pearson correlation = 0.22), possibly due to the smaller dynamic range of observed changes in cell proliferation and harm scores as well as to an outlier case (Supplementary Fig. 5c; Supplementary Data 20–23). Note the sole purpose of harm scores is to rank potential cancer driver exons by their expected harmfulness when perturbed; hence, their magnitude is not biologically meaningful. Overall, this analysis validates our ability to predict the harmfulness of individual splicing changes in potential cancer driver exons on cell proliferation.

**Fig. 2: Inferred splicing dependencies recapitulate the effects of single and multiple splicing perturbations on cancer cell proliferation.**

Knocking down RNA-binding proteins (RBPs) leads to numerous changes in splicing profiles that may contribute to inhibiting cell proliferation. As part of the ENCORE project, Van Nostrand et al.³² measured the transcriptomic profiles of liver carcinoma-derived HepG2 and leukemia-derived K562 cancer cell lines before and after knocking down a total of 263 different RBPs (Supplementary Fig. 5d, e). We thus tested whether modeling the splicing dependency using the exon inclusion and gene expression changes would recapitulate the experimental gene dependencies of each corresponding KD (Supplementary Data 24–28). We calculated harm scores for each potential cancer driver exon and assumed additive effects by summing the individual harm scores for the combined perturbations (Fig. 2a; Supplementary Fig. 5f) (see Methods). We measured the Pearson correlation between the combined harm scores and gene dependency for different numbers of the most harmful exons to evaluate the predictive power of combined harm scores (see Methods). We found that the correlation steadily increased in both cell lines and reached a plateau when considering only the top 10 most harmful (i.e., negative) potential cancer driver exons, with a Pearson correlation of 0.24 (p = 0.0036) in HepG2 cells and of 0.32 (p = 1.4 × 10^–5) in K562 cells (Fig. 2c; Supplementary Fig. 5g). As a control, we repeated the procedure with reversed and subsetted harm scores to confirm that only the most harmful exons had predictive power (Fig. 2c). This suggests that, regardless of the various biological processes that may be altered upon knocking down an RBP, our models can disentangle the contributions of each exon to inhibiting cell proliferation and predict their combined effect using a simple linear combination.

To determine how frequently each harmful exon contributed to impairing cancer cell proliferation, we counted how often each exon appeared among the top 10 most harmful exons for each RBP KD. Changing the splicing of only 14 different exons (six of which are shared between the two cancer cell lines) led to cell harm. For instance, the HsaEX0051262_PUF60 exon is the most frequently used to harm the cell when knocking down RBPs, in both cell lines. PUF60, or Poly(U) binding splicing factor 60, is involved in various nuclear processes, including alternative splicing, apoptosis, and transcriptional regulation^33,34. On the other hand, the HsaEX0016775_COPS3 exon, which belongs to COP9 signalosome subunit 3, and the HsaEX0050345_PRPF18 exon, a splicing factor, are the most harmful targets in HepG2 and K562 cells, respectively (Fig. 2d). Of note, these three frequently-harmful exons belong to pan-essential genes, underlying the ability of our approach to prioritize exons with functional relevance.

In conclusion, solely based on transcriptomic information, our models of splicing dependency can predict how perturbing single or multiple potential cancer driver exons in a sample affects cell proliferation.

Streamlined discovery of splicing-based anticancer therapies

To identify exons with the potential to act as both cancer drivers and therapeutic targets, we analyzed 13 cancer types from The Cancer Genome Atlas (TCGA) and their seven clinically relevant subtypes, as certain cancers can behave differently at the subtype level³⁵ (Supplementary Fig. 6a, b). We considered only those potential cancer driver exons whose inclusion generates an annotated new proteoform and whose modulation would preferentially target tumor cells rather than healthy tissue, and we ranked them according to their predicted maximum harm scores in tumor samples (see Methods). With these criteria, we identified 40 targetable exons with therapeutic potential across cancer types (Supplementary Fig. 6c; Supplementary Data 29), and 21 by cancer subtype (Supplementary Fig. 6d; Supplementary Data 30). Some of these had been previously described as cancer driver exons (Supplementary Fig. 6c, d; Supplementary Data 1). Most prioritized exons are predicted to act as oncoexons (i.e., negative splicing dependency), meaning that excluding them may be a useful treatment strategy for tumors.

For experimental validation, we selected eight potential oncoexons from the above list whose exclusion is predicted to inhibit cell proliferation in lung-derived A549 and breast-derived MDA-MB-231 cancer cell lines (Fig. 3b; Supplementary Data 31). As selected exons generate protein isoforms, their proteoforms could be targeted through small molecules, provided shortlisted exons are not found in disordered protein domains³⁶. However, according to ESMFold structural predictions^37,38 (Supplementary Data 32; see Methods), six of the eight selected exons are indeed in a disordered domain of their proteoform, and none of the proteoforms is targetable using drugs listed in the DrugBank database³⁹ (Table 1; Supplementary Fig. 7). To overcome this, we used SSOs to induce the skipping of our prioritized potential cancer driver exons (Supplementary Data 33 and 34). Of these exons, two were expected to have strong effects (HsaEX0049558_ PPP1R12A, HsaEX0044398_NUP85); four, to have mild effects (HsaEX0034998_KRAS, HsaEX0070392_VLDLR, HsaEX0071941_YAP1, and HsaEX0052877_RCC1); one, to be a strong oncoexon in A549 cells but a mild oncoexon in MDA-MB-231 cells (HsaEX0050345_PRPF18); and one, to be a mild oncoexon in tumor samples but a mild tumor-suppressor exon in cancer cell lines (HsaEX0026116_FNBP1) (Table 1). Of note, we did not consider the strongest oncoexon, HsaEX1036699_SFPQ, because of the lower chances of designing a functional SSO for such a long exon (3,459 base pairs).

**Fig. 3: Applying models of splicing dependency to samples from cancer patients to prioritize splicing vulnerabilities with therapeutic potential.**

Table 1 Selected potential cancer driver exons for experimental validation

Full size table

The skipping of the two exons predicted to have strong effects and the one predicted to be a strong oncoexon in A549 cells significantly decreased cell proliferation compared to the scrambled SSO control (p < 0.05; two-sided t-test). Of the exons expected to have mild effects, HsaEX0052877_RCC1 and HsaEX0071941_YAP1 skipping caused strong proliferative effects, while HsaEX0034998_KRAS, HsaEX0070392_VLDLR, and HsaEX0050345_PRPF18 worked as predicted (Fig. 3c; Supplementary Figs. 8 and 9; Supplementary Data 35; see Methods). As expected from our previous validations (Fig. 2b; Supplementary Fig. 5c), our statistical models show a higher level of precision for predicted values that are further from 0 (Supplementary Fig. 8c), ensuring a reliable performance over a wider range of predictions.

Many approved chemotherapeutics work by targeting ubiquitous proteins that are more essential for highly proliferative cells than for cells that proliferate at lower rates^40,41. In our prioritization criteria, we required the inclusion of potential cancer driver exons to be higher in tumors than the corresponding healthy tissue (i.e., tumor-specific), which greatly reduced the targets with therapeutic potential to be considered. However, these exon inclusion differences were not large in most cases (Supplementary Fig. 6c, d), potentially compromising the specificity of our SSO treatments. To address this, we tested whether the proliferative effect diminished after excluding the eight selected exons in cells with a slower proliferation rate, such as the intestine-derived HT29 cancer cell line, which doubles ~3 times less than A549 or MDA-MB-231 cells in 96 h (Supplementary Fig. 10a; Supplementary Data 36). Even though the relative splicing dependency predictions from exon exclusion with SSOs suggested similar effects as those observed in A549 and MDA-MB-231 cells (Supplementary Fig. 10b; Supplementary Figs. 10–13), none of the SSOs decreased cell proliferation significantly as compared to the scrambled negative control (Supplementary Fig. 10c). Hence, we conclude that targeting potential cancer driver exons has a stronger effect on highly proliferative cells; thus, we may disregard our tumor-specific criterion and consider all identified potential cancer driver exons as putative targets in future prioritizations.

Our models of splicing dependency could have been influenced by a sampling bias regarding the impact on cell proliferation, as we relied on multiple cancer cell lines for their generation and assumed that the inclusion of identified potential cancer driver exons is correlated with cell proliferation in slower-proliferating cell lines. To rule this out, we corroborated that the same predictions generated from models using either only high- or low-proliferating cancer cell lines agreed with the full model (Supplementary Fig. 10d; Supplementary Data 35; see Methods). In agreement with this, Webb et al.⁴² observed that perturbing alternative splicing especially impedes highly proliferative cells.

Cancer cells must orchestrate a multitude of pathways to thrive, which, in turn, generate vulnerabilities. While most anti-cancer drugs are directed against well-known proliferation-related pathways^43,44, vulnerabilities may occur in other pathways as well. Thus, methods such as spotter, that unbiasedly spot such vulnerabilities, may uncover new targetable pathways. In this experimental validation, we hindered cell proliferation by targeting exons in genes involved not only in cell cycle-related pathways (MAPK, Hippo, Ran, Rho) but also in endocytosis (HsaEX0070392_VLDLR), splicing (HsaEX0050345_PRPF18) and the nuclear pore complex (HsaEX0044398_NUP85) (Fig. 3d; Table 1).

Altogether, these results not only confirm the ability of our approach to efficiently pinpoint exons that drive cancer cell proliferation but also suggest that targeting prioritized exons, validated here with SSOs, might be a promising treatment strategy.

Potential cancer driver exons mechanistically mediate drug sensitivity

Alternative splicing can impact drug effectiveness in two ways: directly, by influencing the function of the drug’s target protein, or indirectly, by affecting the activity of genes that either regulate the target or are regulated by it. Drug mechanisms of action can be elucidated systematically at the gene level by associating cell proliferation signatures of drug sensitivity with gene dependencies²³. With our models of splicing dependency, we can now adapt this approach to the splicing level by associating cell proliferation signatures of drug sensitivity to splicing dependencies, thereby uncovering exons mediating drug sensitivity. These mediators have the potential to improve the effectiveness of a drug by increasing its potency or reverting drug resistance in cancer cells while reducing side effects.

To explore the relationship between alternative splicing and drug sensitivity, we analyzed data from the Cancer Genome Project (GDSC1 and GDSC2)⁴⁵. Specifically, we regressed each drug’s cell proliferation signature of drug sensitivity measured as log-scaled half-maximal inhibitory concentration (IC₅₀) against the predicted splicing dependencies of potential cancer driver exons (Fig. 4a; Supplementary Fig. 14a; see Methods). Out of a total of 531,135 associations between the 1,073 potential cancer driver exons and the 495 drug sensitivity profiles, we identified 14,325 significant associations between drug sensitivity and exons (with FDR < 0.1 and n > 20). Of these, 17 were found to occur between exons belonging to drug targets (p = 3.63 × 10^–10, Fisher’s exact test) (Fig. 4b; Supplementary Fig. 14b–f; Supplementary Data 37–39).

**Fig. 4: Splicing dependencies uncover exons influencing drug mechanism of action.**

In the case of indirect relationships, we would expect that exons must somehow be connected to the drug target, namely in the same pathway. Direct drug–exon associations have lower p-values than indirect associations (p = 1.4 × 10^–6, two-sided Wilcoxon rank-sum test) (Fig. 4c), while indirect drug–exon associations are more likely to be found in the same pathways (p = 1.3 × 10^–8, two-sided Wilcoxon rank-sum test) (Fig. 4d; Supplementary Data 40). By mapping these associations onto STRINGDB’s protein–protein interaction (PPI) networks⁴⁶ we measured the distance from each gene to each drug target(s). In PPI space, drug–exon associations are located closer to the drug target(s) (p = 2.4 × 10^–84, Chi-squared) (Fig. 4e; Supplementary Data 41; see Methods). These results suggest that our approach of integrating drug sensitivity data with splicing dependencies can identify both direct and indirect splicing events that may influence drug effectiveness.

To determine whether these statistical associations could distill mechanistic interactions relevant for further characterization, we inspected the coefficients of splicing dependency and drug–exon models to uncover putative drug–exon mechanistic synergies (Supplementary Fig. 15a). To validate our approach, we further analyzed drugs with at least one splicing-mediated target (Supplementary Fig. 15b, c) and focused on the drug nutlin-3A, which has a well-described mechanism of action, of unleashing apoptotic signals by binding to MDM2 and blocking the ubiquitination of TP53 (Fig. 4f). In physiological conditions, TP53 is ubiquitinated when interacting with MDM2 (E3 ubiquitin ligase activity), and this reaction is favored by MDM4 (also known as MDMX), which stabilizes the interaction between the other two⁴⁷. Among the top 10 nutlin-3A drug–exon associations, we found exons (such as HsaEX0038400_MDM2 and HsaEX0038414_MDM4) that shorten the corresponding proteins when excluded (Supplementary Fig. 16).

Changing the isoforms of these proteins modulates the activity of TP53^48,49,50 and the effect of nutlin-3A⁵¹. Nutlin3-A is less effective upon exclusion of HsaEX0038400_MDM2, which disrupts its interaction with TP53 and reduces the affinity for this isoform⁵¹. In agreement, our analysis shows a positive coefficient of the splicing term (splicing dependency model) and the splicing dependency term (drug–exon interaction model) for HsaEX0038400_MDM2 (Supplementary Fig. 16a), recapitulating the higher drug sensitivity upon exclusion (Supplementary Fig. 15a). Further, structural analysis of MDM2 with FoldX⁵² corroborated the relevance of HsaEX0038400_MDM2 in stabilizing the binding site for TP53 (Stability Gain = −15.28 kJ/mol), suggesting that cells including the exon are more resistant to nutlin-3A because there are more MDM2-TP53 functional complexes to block (Fig. 4g). On the other hand, the inclusion of HsaEX0038414_MDM4 introduces an early STOP codon in the transcript that leads to NMD⁵³ and to a truncated version of the MDM4 protein reported to strongly bind and inactivate TP53⁵⁰. In this case, we found that there was a positive splicing term coefficient (splicing dependency model) and a negative splicing dependency term coefficient (drug–exon interaction model), recovering the higher sensitivity (Supplementary Fig. 15a and Supplementary Fig. 16a).

Apart from this well-known drug–exon interaction, our approach uncovered other yet unreported interactions between drugs and exons in their target gene that could be explored in future studies. By inspecting the coefficients of splicing dependency and drug-exon models, we found interactions involving three exons in MDM2 and nutlin-3A, two exons in FGFR2 and the drugs PD173074 and AZD4547, two exons in ERBB2 and the drugs afatinib and CUDC-101, and one exon in EZH2 and the drug GSK343 (Supplementary Fig. 15). These exons likely influence drug sensitivity by (i) generating new proteoforms (n = 2) of the drug target, (ii) disrupting the gene’s ORF through inclusion (n = 1) or exclusion (n = 3) into their transcript, or (ii) by including or not regulatory sequences encoded in the transcript 5’ UTR (n = 2).

In conclusion, similarly to gene dependencies, inferred splicing dependencies reveal how splicing mediates drug responses mechanistically, in both a direct and an indirect manner. These associations open the possibility of generating testable hypotheses to discover new synergies between splicing modulation and drug treatments with translational potential.

Splicing-based treatment recommendation

Our analysis identified drug-exon associations that encode mechanistic information about how drugs may be effective. Similarly to splicing dependency models, the linear models we generated to perform drug-exon statistical associations can be used to predict drug sensitivity from splicing dependency. In this case, because we have a model for each exon and drug, we unified drug sensitivity predictions across exons into a single prediction of sensitivity per drug (see Methods). This enables us to concatenate both splicing dependency and drug-exon interaction models to predict the sensitivity to the 453 GDSC drugs from the transcriptome of a given sample (Fig. 5a). Confirming the predictive power of this approach, our models effectively ranked drug sensitivities with good accuracy (median Pearson correlation = 0.7) in both GDSC screens, even having used a smaller training set than the testing set (Fig. 5b; Supplementary Data 42).

**Fig. 5: Using drug-exon interaction models to predict drug sensitivity of individual cancer samples.**

To further investigate whether drug recommendations from drug-exon interaction models are applicable in a clinical setting, we tested how well our models trained on cancer cell lines could predict the sensitivity of cancer patients to chemotherapy before treatment. We collected RNA-seq data from 49 patients diagnosed with high-grade serous ovarian cancer (HGSOC)⁵⁴ and, using drug-exon interaction models, we predicted their expected sensitivity to their corresponding chemotherapy regimens (Fig. 5a). We then labeled the patients based on their progression-free interval (PFI) as resistant (PFI ≤ 6 months) or sensitive (PFI > 6 months) (Fig. 5c; see Methods). The predicted IC₅₀ values could reliably distinguish between sensitive and resistant samples from patient tumor samples before treatment (p-value_GDSC1 = 0.026; p-value_GDSC2 = 0.0058) (Fig. 5d; Supplementary Fig. 17; Supplementary Data 43–45).

Thus, this analysis shows how our method focused on alternative splicing could help oncologists in determining the best risk-benefit trade-off for cancer treatment.

Discussion

Traditional identification of cancer driver genes has primarily focused on genetic alterations^26,27. While genetic mutations can initiate cancer, the disease progresses by rewiring multiple molecular levels, including pre-mRNA alternative splicing. This process creates new vulnerabilities that must be considered to effectively combat cancer. However, despite transcriptomic data being widely available, there is a lack of studies systematically exploring the interplay between alternative splicing and cancer. As a result, reported cancer driver splicing events in databases like ASCancer Atlas¹⁵ are likely biased by researchers’ interests and available gene function information. These biases may lead to an underestimation of the true number of cancer driver splicing events. Current exon-level CRISPR screens are unable to assess the contribution of exons on the same scale as gene-level perturbation screens. Therefore, new approaches are necessary to identify potential cancer driver splicing events across the genome, as well as highlight the most important ones in a given sample. In this study, we validate simple statistical models based on publicly available data that can help tackle both problems.

Inspired by the association between MDM4’s exon 6 (HsaEX0038414_MDM4) inclusion and gene dependency reported by Gandhi et al.⁵⁵, we hypothesized that extending this concept to all exons could reveal potential cancer driver exons. Using linear models to evaluate the contribution of exon inclusion on top of gene expression when predicting gene dependency, we shortlisted 1,073 potential cancer driver exons. Notably, 56.5% of these exons (n = 607) belong to genes with a low mutation rate and are not known to drive cancer or are considered pan-essential. However, they are found in pathways with known cancer driver genes in similar proportions. This suggests that our approach uncovered previously uncharacterized potential cancer driver genes whose driver potential arises from a putative alternative splicing cancer program, acting alongside well-studied mutational cancer driver genes. This finding supports previous observations that revealed alternative splicing remodels protein-protein interacting domains in known cancer driver genes or pathways⁵⁶, indicating that mutational and splicing alterations can converge to drive cancer progression.

Our approach not only pinpointed relevant exons but also generated models for splicing dependency analysis using solely transcriptomic data. Despite their simplicity, these models can mimic the cell proliferation effects observed when potential cancer driver exons are perturbed individually or in combination with reliable accuracy.

To assess the potential of our approach to accelerate drug development, we prioritized exons with therapeutic potential as targets for SSO treatments across 13 different cancer types. Out of these, we selected eight exons for experimental validation. Six of them belonged to genes not previously identified as mutational cancer driver genes²⁶, of which four belonged to pan-essential genes, which were expected to lack cancer specificity upon perturbation. Additionally, six of the validated exons are located in disordered protein domains. Nevertheless, all validated exons decreased cell proliferation as predicted when specifically targeted with SSOs in the highly proliferative cancer cell lines tested, but not in low-proliferative cells. This highlights the potential of splicing modulation as a strategy to expand the therapeutic window of genes that are inherently difficult to target, leveraging the simplicity of designing SSOs to target pre-mRNAs compared to proteins.

To explore the potential clinical usefulness of splicing dependency models, we associated estimated splicing dependencies with the sensitivity profiles of anticancer drugs. This enabled us to explore the direct and indirect mechanisms of drug–exon interactions, such as those involving nutlin-3A, HsaEX0038400_MDM2, and HsaEX0038414_MDM4. In addition, by integrating drug–exon interaction models as a drug recommendation system, we demonstrate the ability to predict patient response before treatment in an ovarian cancer patient cohort. This illustrates how our statistical models may impact a wide range of research fields: from bench prioritization of exons that may boost drug response, to bedside solutions for precision oncology.

While our splicing dependency and drug-exon interactions models hold broad applicability, they constitute an effort to establish the basis for splicing dependency analysis. We systematically identify potential cancer driver exons using straightforward statistical models that can then be used to generate testable hypotheses. However, our approach likely underestimates the number of potential cancer driver exons identified since we are only considering gene dependency measured for cell proliferation, which represents only one of the twelve cancer hallmarks. Further, we anticipate considering interactions in multi-exon genes, other confounding factors from regulatory molecular layers, or cell lineage specificity in future model iterations will be crucial for a more comprehensive understanding of cancer driver splicing events.

Our work represents an important step towards the efficient systematic characterization of cancer splicing vulnerabilities. To our knowledge, this study represents an unprecedented attempt to overcome experimental limitations to identify and model potential cancer driver splicing events at the genomic scale at single-sample resolution. We envision our approach to harnessing cancer heterogeneity in our favor can illuminate the interplay between alternative splicing and cancer, from its underlying biology to its application in cancer treatments.

To facilitate predicting splicing dependencies on any given cancer sample, we included the 1,073 exon-level linear models in an open-source computational framework called spotter. Given transcriptomic data, spotter predicts the effect of perturbing each potential cancer driver exon. These predictions not only provide insights into the cancer-driving behavior of potential cancer driver exons but also facilitate the generation of data-driven testable hypotheses.

Methods

Quantification of splicing and gene expression

Unless stated otherwise, gene expression and exon inclusion were quantified using vast-tools¹⁹. Specifically, we aligned each raw RNA-seq sample to the hg38 (Hs2 in VastDB) genome annotation with vast-tools align --sp Hs2 --EEJ_counts --expr. Then, we combined all samples with vast-tools combine --sp Hs2 --keep_raw_reads --keep_raw_incl --TPM. Finally, by running vast-tools tidy -min_N 1 -min_SD 0 --min_ALT_use 25 --noVLOW, we masked as NAs those exons and samples without enough coverage according to vast-tools benchmarks (at least 10 reads in total). This workflow resulted in a table of event inclusion as percentage spliced in (PSI) units ranging from 0 to 100 and gene expression as transcripts per million (TPM) units. The details of each command can be found at https://github.com/vastgroup/vast-tools.

Identification of potential cancer driver exons

Data preprocessing

For each cancer cell line in the CCLE, we quantified its gene expression (TPM) and exon inclusion (PSI) as explained above. Raw transcriptomic matrices contained data for 41,614 genes and 206,293 exons across 1019 cell lines. We obtained the metadata for each cell line from the DepMap website (see Data Availability). We only kept those genes and exons that could be quantified in at least 50 cell lines and those cell lines whose identifiers were also found in the metadata table. After this step, preprocessed transcriptomic matrices contained data for 39,982 genes and 164,653 exons across 1015 cell lines.

In parallel, we downloaded the Demeter2 gene dependency scores matrix from the DepMap website (see Data Availability). The raw Demeter2 matrix contained data for 17,309 genes across 712 cell lines. Because the Demeter2 dataset results from an integration of 3 datasets, not all gene KDs were measured for each cell line. We kept those genes for which at least 50 scores were measured and kept those cell lines whose identifiers were also found in the CCLE metadata table downloaded previously. After this step, the preprocessed Demeter2 matrix contained data for 17,280 genes across 712 cell lines.

Compilation of lists of known cancer driver exons, mutational cancer driver genes, and pan-essential genes

We mapped the cancer driver exons listed in the ASCancer Atlas¹⁵ to VastDB identifiers using hg19 coordinates. Out of the 814 exons listed, 598 were mapped to VastDB identifiers with exact coordinates. Hence, in our analyzes, we only used the 598 VastDB exons that could be mapped as known cancer driver exons.

We obtained known mutational cancer driver genes from the COSMIC website (see Data Availability).

We obtained a list of pan-essential genes identified from CRISPR knockout screens from the DepMap website (see Data Availability). And, we obtained a list of pan-essential genes identified from shRNA KD screens by applying Krill-Burger et al.⁵⁷’s method. Briefly, using the Demeter2 gene dependency scores preprocessed as explained above, for each cell line in the matrix, we ranked its genes and computed the ranking percentage by dividing by the total number of genes detected in the sample. This results in a distribution of ranking percentages for each gene from which we saved the 90th percentile for each gene. These 90th percentiles form a bimodal distribution in which genes with low 90th percentile ranking percentages will tend to be essential in most samples. To label them as pan-essential, we perform a density estimation of the distribution to find the local minimum splitting the bimodal distribution into low and high 90th percentile ranking percentages. Pan-essential genes from KD screens correspond to the genes whose 90th percentile ranking percentages belong to the lower end of the distribution.

All these lists were integrated in Supplementary Data 1.

Training models of splicing dependency

To identify those exons whose inclusion strongly co-varies with their gene dependencies across cancer cell lines, we fitted a linear model for each exon using exon inclusion and gene expression as variables to predict Demeter2 gene dependency scores. To associate these datasets we could only consider the 635 cell lines in common between the preprocessed transcriptomic (n = 1015) and Demeter2 (n = 712) datasets. Note that 1811 cell lines are listed in the CCLE metadata table used in the preprocessing steps above.

For each exon, we built a linear model with two predictor (dependent) variables and one outcome (independent) variable, measured across the 635 cell lines (observations). Predictor variables are exon inclusion from the preprocessed exon inclusion matrix and the corresponding gene expression from the preprocessed gene expression matrix. The outcome variable is the Demeter2 gene dependency score vector for each cell line from the preprocessed Demeter2 matrix. Before fitting, we log-scaled gene expression values ($\log 2({TPM}+1)$) and standardized both log-scaled gene expression and exon inclusion variables to ensure they have the same mean and standard deviation. We dropped those observations containing a missing value in any of the variables and, to measure the generalizability of the model in a later step, we randomly split the observations into training (85%) and test (15%) sets with the sklearn.model_selection.train_test_split function. At this point, we proceeded to fit the full linear model on the training set with the statsmodels.api.OLS function using standardized exon inclusion and standardized log-scaled gene expression as variables to predict Demeter2 gene dependency scores:

$${{GeneDependency}}_{{gene}}= {{\beta }_{{Intercept}}}_{{exon}} \cdot {Intercep}{t}_{{exon}} \\ +{\beta }_{{gene}} \cdot \log 2({{TPM}}_{{gene}}+1)+{\beta }_{{exon}} \cdot {{PSI}}_{{exon}}$$

(1)

We wanted to evaluate two aspects of the fitted model: the contribution of the exon inclusion variable compared with the gene expression variable alone and the generalizability of the model. To measure the former, we used the training set to fit a null model consisting of the standardized gene expression as the only variable to predict Demeter2 gene dependency scores. Then, we compared the full and the null models by performing a two-sided Likelihood-Ratio (LR) statistical test that returns a statistic indicating whether the full model, containing both the exon inclusion and gene expression variables, is significantly better than the null model, containing only the gene expression variable. On the other hand, to measure the generalizability of the full model, we used the 15% test set to make predictions and compared them with real Demeter2 gene dependency scores computing a Pearson correlation (test-set Pearson correlation). For each fitted model, we obtained two evaluation metrics, an LR statistic, and a test-set Pearson correlation, quantifying the contribution of the exon inclusion variable to the full model and its gene generalizability, respectively.

The variables used for modeling may contain missing values that cannot be considered for fitting. This may lead to uncertain evaluation metrics. To account for this uncertainty, we repeated the fitting step 1000 times for each exon, using different random seeds for the train-test splits. For each exon, we saved the parameters for the 1000 fitted full models and corresponding LR statistics and test-set Pearson correlations. Finally, for each exon, we obtained the LR test p-value using the mean of the LR statistics across iterations with the scipy.stats.chi2.sf function, and we summarized test-set Pearson correlations computing their mean.

By fitting the model for each exon multiple times using different data splits we not only obtain robust evaluation metrics but also enable performing robust predictions across all the fitted models of each exon rather than a single-point estimate. Because these models predict gene-level dependency from splicing variation, we termed their predictions “splicing dependency” instead of “gene dependency”.

Evaluation of predictive power of LR p-values

As we hypothesized our statistical approach is capable of identifying potential cancer driver exons, after fitting the models we evaluated the ability of LR p-values to re-discover either known cancer driver exons, known mutational cancer driver genes, or pan-essential genes identified from CRISPR- and shRNA-based screens (Supplementary Data 1). We performed TPR-FPR and Precision-Recall curve analyzes with the pROC::roc function using LR p-values as the predictor variable for each prior knowledge label. For those genes with multiple exons, there were multiple p-values when performing this analysis with known mutational cancer driver genes and pan-essential genes as ground truth. In those cases, we assigned the lowest LR p-value to the gene.

Evaluation of biases in Demeter2 gene dependency scores due to shRNA targeting

Demeter2 scores were statistically derived upon modeling the proliferation effects of targeting gene transcripts across cancer cell lines²². To maximize KD efficiency, most genes are targeted by a variable number of shRNAs. Hence, we wondered whether differences in shRNA targeting coverage differences could impact our analyzes. We obtained shRNA sequences from the DepMap website and mapped them to VastDB exons and to GENCODE v44 transcripts (see Data Availability). Note that VastDB only annotates alternatively spliced exons, hence, to map shRNAs to the full transcriptome we had to use GENCODE v44 transcripts.

To map shRNA sequences to VastDB exons, we loaded the VastDB splicing event information and Demeter2 shRNA sequences tables directly downloaded from their corresponding websites (see Data Availability). The VastDB splicing event table contains the sequences of each exon and those of its neighboring exons. As some shRNAs could map on the splice junctions of a given exon with its neighbors, we combined the 7 last nucleotides of the upstream neighbor exon and the 7 first nucleotides of the downstream neighbor exon with the corresponding exon sequence. Then, we used pandas.Series.str.contains to map –allowing for 0 mismatches– the first 19 nucleotides of each shRNA to the expanded exon sequences of the corresponding gene annotated in the Demeter2 shRNA sequences table. This mapping relates each shRNA to VastDB exons.

To align shRNA sequences to the GENCODE v44 transcripts, we saved each shRNA sequence (see Data Availability) into a fasta file using only the first 19 nucleotides as indicated in McFarland et al.²². Then, we built the bowtie index for GENCODE v44 transcripts with bowtie-build with default parameters, we aligned the shRNA fasta-formatted sequences to the indexed transcripts allowing for 0 mismatches running bowtie -v 0. We ran samtools view -b to convert the SAM alignment file into BAM and bedtools bamtobed to convert the BAM file into BED format. In this step, we used bowtie version 1.3.0, samtools version 1.6, and bedtools version 2.32.1. This alignment returns where each shRNA maps on human transcripts but not on the corresponding genomic coordinates. To obtain the genomic coordinates, we downloaded the GENCODE v44 genome annotation (see Data Availability) and used the function ensembldb::ensDbFromGtf to convert the downloaded gtf-formatted file into an SQLite database that can be read with the function ensembldb::EnsDb. Then, using the ensembldb::transcriptToGenome function we converted transcript coordinates to genomic coordinates for the BED-formatted alignment. This mapping generates a table relating each shRNA to all annotated transcripts.

Evaluation of uncertainty of predicted splicing dependencies

PSI values represent ratios of the number of reads supporting the inclusion of a given splicing event divided by the total number of reads supporting both the inclusion and the exclusion of that splicing event in mature mRNA transcripts. Depending on the total number of reads, these proportions are more or less noisy such that the more reads we detect for a given splicing event the more stable the proportion is to changes in few inclusion/exclusion read counts. Binomial distributions model this problem. Hence, we performed random samplings from the binomial distribution to simulate and evaluate the noise introduced in PSI inference with different total read counts of an exon.

We simulated how in theory the error in PSI estimation increases as we decrease the number of total read counts to compute a PSI. We used the base::rbinom(n = 1000) function with 1000 random samples for all PSIs between 0 and 100 with steps of 5 units, with either 2, 10, 100, or 1000 total read counts. We saved the average from each random sample.

We also simulated how this error may propagate in the prediction of splicing dependencies with our splicing dependency models. For the CCLE cell lines with the fewest (ACH − 000934, 32,750,708 reads) and the most (ACH − 000143, 331,178,064 reads) total raw read counts, we randomly sampled their PSI values 100 times with the function numpy.random.binomial(n_random_samples = 100) setting the n parameter to the corresponding total read counts observed for each specific exon in each sample. To perform this step, we obtained total read counts for each exon and sample from the table resulting from running the vast-tools combine command. This table contains not only PSI values for each exon but also the number of reads supporting its inclusion and exclusion with the format InclusionJunction1=InclusionJunction2=ExclusionJunction. We wrote a custom Python script to extract the counts and to sum them obtaining a table with the total read counts mapped to each exon and sample. In both cell lines, we observe the same trend: PSI and splicing dependency errors are larger when fewer reads are used to calculate PSIs (Supplementary Fig. 2c; Supplementary Data 11). However, the high correlations between real and simulated values for PSI or splicing dependency (Pearson correlation = 1, p < 2.2 × 10^–16 in all cases) indicated that our threshold of ≥ 10 reads was sufficient to avoid remarkable PSI estimation errors.

Model selection

We used a knowledge-driven approach to find the significance threshold for LR-tests. We combined our hand-curated list of exons that have been reported to modulate cell growth at least in one cell line and condition with the cancer driver exons from the ASCancer Atlas¹⁵ (Supplementary Data 1). Since not all ground truth exons will likely have the same effect across all cancer cell lines, we defined the true-positive set by picking the exons belonging to the 100 genes with the highest standard deviation in gene dependency across cancer cell lines. Accordingly, we defined true-negative sets by picking the exons belonging to the 100, 1000, and 10,000 genes with the lowest standard deviation in gene dependency that did not appear in our ground truth list. We considered an increasing size of the true-negative set but not the true-positive set because we expected cancer-driver exons to be a minority. Considering only those exon models with test Pearson correlation > 0, we measured true- and false-positive rates for different LR p-value thresholds and selected the LR p-value threshold as the value with the maximum cumulative sum of differences between true-positive and false-positive rates (Supplementary Fig. 3c). Significant splicing dependency models identify exons whose inclusion contributes to the predictive power of the model. However, this selection step does not guarantee the individual predictions from these models are well calibrated such that ranking predictions of individual models recapitulate experimental observations.

To ensure that the predictions of the different models maintained experimental rankings at the single sample level, we applied a second filter based on the test-set Pearson correlation of each exon model. As this metric quantifies how well each individual model recapitulates test-set Demeter2 gene dependencies during fitting, we reasoned selecting models with high test-set Pearson correlation should lead to calibrated predictions. To evaluate this, for different test-set Pearson correlation thresholds, we selected sets of fitted splicing dependency models with test-set Pearson correlation higher than each threshold and predicted splicing dependencies across cell lines. If the selected set of models is well calibrated, ranking their predicted splicing dependencies across exons for each cell line should match the corresponding ranking of the real Demeter2 gene dependencies. We quantified this by computing a Spearman correlation for each cell line between predicted splicing dependencies and real Demeter2 gene dependencies. In the case of selecting more than one exon for the same gene, to avoid duplicate predictions, we considered only the model with the worst individual test-set Pearson correlation. This produced the distribution of Spearman correlations for each test-set Pearson correlation threshold plotted in Supplementary Fig. 3d. We selected 0.2 as the test-set Pearson correlation threshold because Spearman correlations plateaued at this point (Supplementary Fig. 3d).

The exons corresponding to selected models (LR p-value < 0.025 and test-set Pearson correlation > 0.2) are potential cancer driver exons.

Gene set overlap analysis

Unless stated otherwise, all statistical tests for gene set overlap analysis were performed using the function clusterProfiler::enricher.

Gene-level mutation frequency

We computed gene-level mutation frequencies by counting the number of mutations per gene ($i$) and variant classification ($j$) across cancer cell lines from the CCLE and dividing them by the corresponding gene lengths computed from the GeneBreak::ens.gene.ann.hg19 object:

$${Mutation\; Frequency}=\frac{{{{\rm{\#}}}}{{Mutations}}_{{ij}}}{{{GeneLength}}_{i}}$$

(2)

We only considered those mutation effects present in more than 10 different genes either bearing at least a potential cancer driver exon or not.

Simulating single-exon perturbations with harm scores

Given gene expression and exon inclusion, fitted models of splicing dependency predict the splicing dependency for each of the selected potential cancer driver exons. To make predictions, both variables are standardized internally to correspond to the variables used during the model-fitting step. We interpret predicted splicing dependencies as the expected change in cell proliferation of perturbing a gene isoform given its expression and exon inclusion. Note that even if an exon is not included predicted splicing dependencies may not be 0 necessarily. Also, splicing dependency models do not directly account for different exon perturbation strengths.

To simulate how much “harm” we may infringe when excluding or including an exon in a sample, we multiply predicted splicing dependencies using the exon inclusion in the basal state (${{SplicingDependency}}_{{basal},{exon}}$) by the desired change in exon inclusion obtained by subtracting the basal exon inclusion to the perturbed exon inclusion (${\varDelta {PSI}}_{{exon}}$). This results in a harm score (${{HarmScore}}_{{exon}}$) that reflects the expected effect of including or excluding an exon in a sample.

$${\varDelta {PSI}}_{{exon}}={{PSI}}_{{perturbed}}-{{PSI}}_{{basal}}$$

(3)

$${{HarmScore}}_{{exon}}=(-1)\cdot {\varDelta {PSI}}_{{exon}}\cdot {{SplicingDependency}}_{{basal},{exon}}$$

(4)

Validation of the predictive power of splicing dependency models for single exon perturbations

Data preprocessing

We obtained data from two exon-level CRISPR screen datasets^17,18 to assess how well splicing dependency models can predict how much single-exon deletions affect cancer cell proliferation. To that end, we required both the measurements quantifying how every single exon deletion affected the proliferation of each cancer cell line and the basal RNA-seq profile of the cell lines used in each screen.

Thomas et al.¹⁸ provided both types of data for HeLa and PC9 cancer cell lines. We then mapped the coordinates of their screened exons to the VastDB exon identifiers and quantified the splicing and gene expression from their raw RNA-seq files with vast-tools as explained above.

On the other hand, Gonatopoluos-Pournatzis et al.¹⁷ provided the cell proliferation measurements upon exon deletion already mapped to VastDB exon identifiers for the RPE1 cell line. However, instead of profiling the cell line’s transcriptome by RNA-seq themselves, they obtained these data from Hart et al.⁵⁸. Following the same approach, we used this last dataset to quantify the basal splicing and gene expression of the cell line using vast-tools, as explained above.

Harm score computation

To predict the expected change in proliferation upon deletion of every mapped potential cancer driver exon, we computed the harm scores as explained above setting ${{PSI}}_{{final},i}=0$ to simulate deleting each exon.

Evaluation

For those exons targeted in the CRISPR screens and those selected as potential cancer driver exons, we computed the Pearson correlation between harm scores derived from estimated splicing dependencies and measured changes in cell proliferation using stats::cor(method = ”pearson”).

Prediction of cancer cell proliferation from combinatorial splicing changes

Data preprocessing

To assess whether our models of splicing dependencies can predict overall changes in cell proliferation upon simultaneous changes in the inclusion of multiple cancer driver exons, we require both the transcriptomic changes and the measured change in cell proliferation upon perturbation. Van Nostrand et al.³² measured the transcriptomic profiles before and after knocking down different RNA binding proteins (RBPs) with short-hairpin RNAs (shRNAs) in two cancer cell lines, HepG2 and K562. Specifically, they knocked down 223 RBPs in HepG2 and 217 RBPs in K562. We then quantified splicing and gene expression from Van Nostrand et al.³²’s raw RNA-seq files using vast-tools as explained above. Unfortunately, Van Nostrand et al.³² did not measure the changes in cell proliferation caused by their RBP KDs. Therefore, although they result from another study, we used the overlapping Demeter2²² gene dependencies (149 for HepG2 and 172 for K562) as the observed effects in cell proliferation upon perturbation because it is the only available dataset measuring the effects of knocking down the same RBPs in the same cancer cell lines with the same screening approach.

Combined harm score computation

To infer how combined changes in the inclusion of multiple cancer driver exons affect cell proliferation, we computed harm scores as explained above comparing the basal versus the knocked-down transcriptomic profiles of detected cancer driver exons. Specifically, we computed harm scores for each perturbed sample as explained above averaging the basal splicing dependency and cancer driver exon inclusions (PSI). In case there were replicates available in the perturbed condition, we averaged the harm scores predicted for every exon. To combine harm scores from $N$ individual exons and predict their combined effects, we assumed additive effects by summing their harm scores:

$${{CombinedHarmScore}}_{\{i,...,N\}}={\sum }_{i=1}^{N}\,{{HarmScore}}_{{exo}{n}_{i}}$$

(5)

With this approach, for a given set of exons $\{i,\,...,N\}$, we obtained a combined harm score for each RBP KD.

Evaluation

To quantify the predictive power of our combined harm scores for a given set of exons, we correlated the combined harm scores across all RBP KDs with observed effects in cell proliferation using stats::cor(method = ”pearson”) function. For both cancer cell lines, we computed this correlation with an increasing number of exons ranked according to different criteria: considering first exons whose inclusion changes are predicted to be the most harmful (“most”); considering exons whose inclusion changes are predicted to be the least harmful (“less”); and, dropping exons whose inclusion changes are predicted to be the most harmful (“higher”).

Systematic prioritization of targetable cancer driver exons

Data preprocessing

For 13 different types of cancer from the TCGA project with more than 10 “primary tumor” and “solid tissue normal” samples (BLCA, BRCA, COAD, HNSC, KICH, KIRC, KIRP, LIHC, LUAD, LUSC, PRAD, THCA, UCEC), we used TCGAbiolinks::GDCquery and TCGAbiolinks::GDCdownload to obtain gene expression as gene read counts. To quantify exon inclusion, after obtaining dbGaP’s permission to access the raw RNA-seq data in TCGA, we used vast-tools as explained above without running the module to quantify gene expression. In this case, we had to change the alignment command to “vast-tools --ns --stepSize 24 --trimLen 48 --sp Hs2 --EJ_counts” to accommodate the pipeline to sequencing reads shorter than 50 bp. Subsequently, we dropped formalin-fixed, paraffin-embedded (FFPE) samples and, for those patients with more than one sample, we kept the sample with fewer missing values in their PSI profile.

Differential splicing analysis

For each cancer type, we performed a two-sided Wilcoxon Rank Sum test to compare the inclusion of cancer driver exons between primary tumors and solid tissue normal (STN) samples using the function scipy.stats.mannwhitneyu. We repeated this procedure at the cancer subtype level using the classification from Mina et al.³⁵. Specifically, BRCA primary tumors were stratified into NotBasal, Basal, and Her2; UCEC primary tumors were stratified into CN_low, CN_high, POLE, and MSI. These primary tumor subtypes were then compared to all available STN samples for the corresponding tissue. Study abbreviations for TCGA datasets can be found online at https://gdc.cancer.gov/resources-tcga-users/tcga-code-tables/tcga-study-abbreviations.

Splicing dependency analysis

To pinpoint the splicing vulnerabilities in patients’ cancer samples from TCGA, we inferred the splicing dependencies of each sample and a maximal harm score setting ${{PSI}}_{{perturbed}}$ according to whether the potential cancer driver exon was expected to act as an oncoexon (exclusion is harmful; ${{PSI}}_{{perturbed}}=0$) or a tumor suppressor exon (inclusion is harmful ${{PSI}}_{{perturbed}}=100$). In this case, to predict splicing dependencies, we converted gene mRNA read counts into TPMs with the function target_spotter.utils.count_to_tpm using the same effective gene lengths used by vast-tools: the file “Hs2_mRNA-50-SS.eff” in the Hs2 (hg38) human annotation from VastDB.

Prioritization criteria

We considered as targetable potential cancer driver exons with therapeutic potential those exons generating protein isoforms that were significantly differentially spliced (|ΔPSI | > 5 and FDR < 0.05) in the direction that, in case of treatment with splice-switching oligonucleotides (SSOs), would preferentially harm the tumoral tissue. Precisely, we kept exons with negative splicing dependency more significantly included in primary tumors, and those exons with positive splicing dependency more significantly excluded in primary tumors using the TCGA dataset. Finally, we selected exons expected to result in different responses upon exclusion in our cancer cell lines.

Experimental validation of selected targetable cancer driver exons with therapeutic potential

In-house RNA-seq of the cell lines used in the manuscript

RNA was extracted from A549, MDA-MB-231, and HT29 cells with the RNeasy Mini kit (QIAGEN). Standard polyA-selected Illumina libraries were generated at the CRG Genomics Unit and were sequenced in an Illumina NextSeq2000, producing an average of ~50 million paired-end 151 bp reads.

Selection of exons for experimental validation

From the targetable cancer driver exons with therapeutic potential prioritized above, we selected eight potential oncoexons that generated new proteoforms according to VastDB’s annotation and whose exclusion, according to the predictions from splicing dependency models, should also decrease cell proliferation in 3 cancer cell lines (A549, MDA-MB-231, HT29).

Analysis of the impact of selected exons on their main proteoform

For each exon, we obtained the Ensembl identifier of its corresponding main inclusion proteoform listed in VastDB. Then we used Ensembl’s REST API to retrieve the corresponding amino acid sequences and the coordinates of selected exons. In addition, we retrieved all protein domain annotations available. We used the full amino acid sequence in ColabFold to predict the corresponding structure of each proteoform with ESMFold obtaining a PDB file. Proteoform sequences and their corresponding coordinates and identifiers are available in Supplementary Data 32. Predicted PDB files can be found in the publication’s code repository (see Data Availability).

Design and validation of SSOs

For each target, we designed SSOs to target either the 3’ss (3SS), the 5’ss (5SS) and, when possible, the exonic element (exon). Each SSO is 21 nt long and has been purposefully designed for each exon considering the sequences of the cis-regulatory elements trying to avoid the activation of possible cryptic 3’ss and 5’ss in close proximity to the annotated splice sites and with a GCC as close as possible to 50%. We verify that the SSO target sequence is unique in the genome with BLAST⁵⁹ (https://blast.ncbi.nlm.nih.gov/Blast.cgi) and that SSOs are not prone to form secondary structures. Additionally, ESEfinder⁶⁰ (https://esefinder.ahc.umn.edu/cgi-bin/tools/ESE3/esefinder.cgi) and cisBP-RNA⁶¹ (http://cisbp-rna.ccbr.utoronto.ca/TFTools.php) tools were utilized to investigate the presence of putative exonic regulatory motifs. In addition, we used a scrambled SSO negative control and an SSO targeting HsaEX0044216_NUMB (Patent ID: PCT/EP2019/056889 [https://patentscope.wipo.int/search/en/detail.jsf?docId=WO2019180046]). 2’-O-methyl-modified phosphorothioate SSOs were purchased from IDT and sequences are available in Supplementary Data 33.

Cell culture and SSO transfection

MDA-MB-231, HT29, and A549 cells were cultured in Dulbecco’s modified Eagle’s medium (DMEM) supplemented with 10% fetal bovine serum (FBS) and penicillin/streptomycin. Transfection with 30 nM SSOs was performed using Lipofectamine RNAi max (Thermofisher #13778100) and Opti-MEM according to the manufacturer protocols. At the time of transfection, the cells were kept in a non-supplemented medium for 5 hours before changing to a fully supplemented medium.

Validation of oncoexons by RT-PCR

All cell lines used were plated in 24 well plates (0.25 ×10⁶ cells per well) with complete medium. 24 hours after transfection, cells were harvested, RNA was extracted using an RNeasy mini kit (Qiagen, 74104), and 1 µg of RNA was reverse-transcribed into cDNA using SuperScriptIII Reverse Transcriptase (Invitrogen #18080044). Splicing detection was achieved by RT-PCR using GoTaq® Flexi DNA Polymerase kit (Promega). Primers were obtained directly from VastDB¹⁹, which provides designed primers annealing to the flanking upstream and downstream constitutive exons. PCR products were resolved on 2% agarose gel in SB buffer for 25 minutes at 100 V. Gel images were acquired with the GelDoc XR+ system (Bio-Rad) using non-saturating settings. Band intensity was quantified using ImageJ and PSI was quantified as: inclusion / (inclusion + exclusion) x 100. All primer sequences are provided in Supplementary Data 34.

Clonogenic assay

Clonogenic assays were performed in technical and biological triplicates⁶². Cells were plated at low density (5000 cells per well) in 6 well plates with complete medium and transfected as described above. After 7 days, cells were washed with PBS, fixed, and stained with a solution of 25% methanol and 0.5% crystal violet in PBS (2 ml per well) for 1 h at room temperature. After washing and drying, the adsorbed dye was resolubilized with 70% ethanol, which was transferred into 96-well plates and measured photometrically (570 nm) in a microplate reader. The clonogenic potential was calculated by normalizing the absorbance to that of untreated controls. Finally, we averaged 3 technical replicates for each of the 3 biological replicates and performed two-sided T-tests using stat::t.test to determine which changes in cell proliferation upon SSO treatment were statistically significant (p < 0.05) with respect to the negative control, a scrambled SSO.

Resazurin cell proliferation assay

Following the protocol from Head et al.⁶³, resazurin sodium salt (Sigma-Aldrich Quimica S.A. R7017) stock was prepared at 0.2 mg/mL in PBS and sterilized by filtration through a 0.22 um PVDF filter (Merck Chemicals, and Life Science S.A SLGP033RS). Cells were seeded at a density of 1000 cells/well in 200 μL of complete media with each condition in duplicate. Resazurin was added at the time of seeding and every 24 h to 1 plate (10 μL/well), and fluorescence intensity was measured after 5 h using an Infinite M-Plex plate reader from Tecan (Ibérica Instrumentación S.L, Barcelona, Spain) (ex = 560 nm, em = 590 nm). For each cell line, we normalized the mean fluorescence of cells with respect to the first time point.

Splicing dependency analysis of HT29 cells with high- and low-proliferation models

To confirm the effect of proliferation rate in our main model predictions, we fitted splicing dependency models for each of our selected exons for experimental validation using either highly or lowly proliferative cancer cell lines. We combined cell doubling times from Tsherniak et al.⁶⁴ and Ghandi et al.⁵⁵ by taking the minimum doubling time in those cases where it had been measured for the same cell line in both studies. Then, we split cancer cell lines into two groups, highly or lowly proliferative, considering the top and bottom 25% respectively. For each group, we generated a linear model for each exon selected for experimental validation using spotter as explained above to infer their corresponding splicing dependencies and harm scores.

Drug-exon statistical interactions

Data preprocessing

We started this analysis using the preprocessed transcriptomic data matrices to predict splicing dependencies. We fed fitted and selected splicing dependency models with two data matrices consisting of 39,982 genes and 164,653 exons across 1,015 cell lines. This resulted in a data matrix of splicing dependencies for the 1,073 selected exons across 1,015 cell lines.

On the other hand, we downloaded the GDSC drug screens⁴⁵ (see Data Availability). The raw table contained the drug sensitivities measured as IC₅₀ for a total of 974 different cell lines (973 were used in GDSC1 and 794 were used in GDSC2) and 453 drugs (320 were used in GDSC1 and 175 were used in GDSC2).

In this analysis, we wanted to associate predicted splicing dependencies with drug sensitivity. Hence, to ensure the independence of our findings we performed these associations using only those cell lines that were not present in the Demeter2 dataset for model fitting. For each GDSC screen, we kept the measurements for those cancer cell lines in the transcriptomic data matrices and we placed the drug sensitivities measured in non-Demeter2 cell lines in the training sets (194 for GDSC1 and 159 for GDSC2), and the rest in the validation sets (442 for GDSC1 and 393 for GDSC2).

In addition, each cell line’s growth rate was inferred following Gonçalves et al.²³’s approach; we PCA-decomposed the preprocessed drug sensitivity screen training matrices, either GDSC1 or GDSC2, using sklearn.decomposition.PCA(n_components = 1), and saved the first component as the vector for cancer cell growth rate. The GDSC1 training set matrix contained a minimum of 13 and a maximum of 124 missing values per drug (n = 194) and the GDSC2 training set matrix contained a minimum of 0 and a maximum of 158 missing values (n = 159). To perform this PCA, missing values were imputed as the median drug sensitivity for each drug across cancer cell lines following Gonçalves et al.²³’s method. Imputed matrices were exclusively used to generate the vector of cancer cell growth rates described in this step.

Training models of drug-exon interaction

We associated predicted splicing dependencies from cancer cell transcriptomes with experimentally measured drug sensitivities from the GDSC drug screens adapting the method from Gonçalves et al.²³. For each drug and exon, we used the training set to predict preprocessed drug sensitivities (${\log ({{IC}}_{50})}_{{cell},{drug}}$) from predicted splicing dependencies (${{SplicingDependency}}_{{exon},{cell}}$) and cell growth rates (${{GrowthRate}}_{{cell}}$). After standardizing the splicing dependency and growth rate variables, we performed an 85%-15% train-test random split of the input training data. Then, we fitted a full and a null model without the splicing dependency variable using the glimix_core.lmm.LMM function:

$$ {H}_{o} \! :\,\log({IC}_{50})_{cell,\, drug}=\beta_{Intercept_{exon}} \cdot {Intercept_{exon}}+{\beta_{cell}} \cdot GrowthRate_{cell}+\varSigma \\ {H}_{1} \! :\,\log ({IC}_{50})_{{cell},{drug}}={\beta }_{{Intercep}{t}_{{exon}}} \cdot {{Intercept}}_{{exon}}+{\beta }_{{exon}} \cdot \\ {{SplicingDependency}}_{{exon},{cell}}+{\beta }_{{cell}} \cdot {{GrowthRate}}_{{cell}}+\varSigma$$

(6)

The covariance term, $\varSigma$, is a constant term calculated using the pandas.DataFrame.cov function on the splicing dependency matrix of the input training data. We included this term following Gonçalves et al.²³ ‘s method.

After model fitting, to find significant drug-exon interactions, we performed an LR-test comparing the full model, ${H}_{1}$, to the null model, ${H}_{o}$, resulting in a p-value of the contribution of the splicing dependency term in the full model. Drug-exon statistical interactions with FDR < 0.1 and supported by more than 20 observations were considered significant. On the other hand, to estimate model generalizability, we computed a test Pearson correlation for the 15% test split of the input training data.

Evaluation of mechanistic drug-exon interactions

Data preprocessing

We used HGNC’s standard nomenclature to correct non-standard gene symbols in the drug targets table from GDSC⁴⁵ (see Data Availability).

Pathway-level assessment

For each drug, we counted the number of significant drug-exon interactions that occur between a gene with at least one exon within the ReactomeDB⁶⁵ pathway(s) targeted by the drug. To test whether these pathway-specific indirect interactions could be explained by random chance, for each drug, we randomized the labels for significant drug-exon interactions and counted how many drug-exon interactions occur in the same targeted pathways as explained before. Finally, we performed a two-sided Wilcoxon rank-sum test using stat::wilcox.test to evaluate whether the real and random distributions were significantly different.

Protein-protein interaction assessment

We hypothesized that if statistical drug-exon interactions carried mechanistic information we should observe more significant interactions as we search closer to a drug’s target(s). Following Gonçalves et al.²³’s approach, we mapped drug-exon interactions on STRINGDB⁴⁶ protein-protein interaction network and computed the relative frequency of significant versus non-significant statistical associations occurring at different shortest path lengths from the gene with a drug-exon interaction to the drug target(s).

Exploring drug-exon mechanistic interactions mediating Nutlin-3A’s effectiveness

To assess whether drug-exon models recapitulated known mechanistic interactions, we further explored the case of Nutlin-3A.

We sorted significant Nutlin-3A-exon interactions according to the fitted coefficients of the splicing dependency term and considered the top 10 interactions. Among these, we focused on direct and indirect drug-exon interactions with HsaEX0038400_MDM2 and HsaEX0038414_MDM4, respectively, as they impacted their corresponding genes at the protein level according to VastDB’s protein impact annotations and their interaction with TP53 and Nutlin-3A had already been studied in the literature.

HsaEX0038400_MDM2 protein structure stability analysis

We downloaded the crystal structure of MDM2 bound to TP53 with identifier 4hfz [https://www.rcsb.org/structure/4HFZ]⁶⁶ from PDB⁶⁷. We then removed TP53’s atoms and processed MDM2 into three different structures: full-length MDM2 (full), keeping only residues Val28 to Glu52 corresponding to exon HsaEX0038400_MDM2 (exon), and removing residues Thr26 to Glu52 corresponding to exon HsaEX0038400_MDM2 and 2 N-terminal residues (no exon). We then ran “foldx --command=RepairPDB” to repair preprocessed structures and “foldx --command=Stability” to measure their stability using FoldX⁵². Finally, to measure the gain in the stability of including HsaEX0038400_MDM2, we subtracted the sum of the stabilities of considering only the exon and the structure without the exon to the stability of the full structure (Fig. 4g).

Estimation of drug sensitivity with drug-exon interaction models

To estimate the expected drug sensitivity of a given sample, for each drug, we combined drug-exon interaction models as follows:

$${{if\; PearsonCorr} \, > \, 0:w}_{{exo}{n}_{i}}=\frac{{{PearsonCorr}}_{{exo}{n}_{i}}}{{\sum }_{i=1}^{n}\,{{PearsonCorr}}_{{exo}{n}_{i}}}{{{\rm{;}}}} \, {else}:{w}_{{exo}{n}_{i}}=0$$

(7)

$${\log ({{IC}}_{50}){\prime} }_{{drug},{cell}}={\sum }_{i=1}^{N}\,\left[{w}_{{exo}{n}_{i}} \cdot ({{\beta }_{{Intercep}{t}_{{exo}{n}_{i}}} \cdot {Intercept}}_{{exo}{n}_{i}}+{\beta }_{{exo}{n}_{i},{cell}} \cdot {SplicingDependency}_{{exo}{n}_{i},{cell}}+{\beta }_{{cell}} \cdot {{GrowthRate}}_{{cell}})\right]$$

(8)

For each drug $j$, we inferred its log-scaled IC₅₀, $\log ({{IC}}_{50})^{\prime}$. We summed individual predictions of each drug-exon model for that drug weighted by the scaled test-set Pearson correlation, ${PearsonCorr}$. Those with negative test-set Pearson correlation were not considered, ${w}_{{exo}{n}_{i}}=0$.

Validation of drug-exon interaction models for personalized drug recommendation

In vitro predictive power

We used drug-exon interaction models to predict the drug sensitivities for all cancer cell lines from the CCLE with publicly available transcriptomes that had also been screened in the GDSC dataset. We then measured the ability of these models to correctly infer the drug sensitivity for each sample by computing the Pearson correlation between the real and predicted drug IC₅₀s for each cancer cell line.

Clinical predictive power

We obtained the pre-treatment tumor transcriptomic data and corresponding clinical metadata from HGSOC patients from Pietilä et al.⁵⁴. We quantified their exon inclusion and gene expression with vast-tools as explained above. For each patient, we estimated the sensitivity to each corresponding treatment regimen with splicing dependency and drug-exon interaction models. For those patients treated with multiple drugs, we summarized their predicted drug sensitivities by computing the median resulting in an overall drug sensitivity prediction for each patient. Patients with PFI shorter or longer than 6 months were labeled as “Sensitive” or “Resistant”, respectively, following Pietilä et al.⁵⁴’s approach. We then assessed whether predicted drug sensitivities were lower in “Sensitive” samples compared to “Resistant” samples by performing a two-sided Wilcoxon Rank Sum test using stat::wilcox.test. Finally, we confirmed the predictive power of the models given different PFI thresholds to classify patients through ROC analysis using pROC::roc.

Statistics & reproducibility

No statistical method was used to predetermine sample size. No data were excluded from the analyzes. The experiments were not randomized. The Investigators were not blinded to allocation during experiments and outcome assessment.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Data availability

The following datasets were generated: The RNA sequencing data for cell lines MDA-MB-231, HT29, and A549 generated in this study have been deposited in the Sequence Read Archive (SRA) under accession code PRJNA958127. Supplementary Data were made available at https://doi.org/10.6084/m9.figshare.23502006. Source data are provided with this paper. Previously published datasets used: The normalized gene dependencies (Demeter2) from the RNAi viability screen conducted by McFarland et al.²² are available from DepMap’s figshare repository at https://ndownloader.figshare.com/articles/6025238/versions/6. Molecular data from the Cancer Cell Line Encyclopedia (CCLE), including raw RNA-seq.fastq files, can be accessed from the European Nucleotide Archive (ENA) under project number PRJNA523380. Cell line metadata (“sample_info.csv”) and processed mutation mapping across CCLE cell lines (“CCLE_mutations.csv”) are available from DepMap’s figshare repository at https://ndownloader.figshare.com/articles/13681534/versions/1. The list of CRISPR-based pan-essential genes used in this study can be found on DepMap’s website at https://depmap.org/portal/download/all/?releasename=DepMap+Public+23Q2&filename=CRISPRInferredCommonEssentials.csv. The list of cancer driver genes (COSMIC-CGC) is available from the COSMIC Cancer Gene Census website at https://cancer.sanger.ac.uk/cosmic. The Demeter2 shRNA sequences used in the analysis were obtained from DepMap’s website at https://depmap.org/portal/download/all/?releasename=DEMETER2+Data+v6&filename=shRNA-mapping.csv. GENCODE’s version 44 of the human transcriptome (gencode.v44.transcripts.fa.gz) can be downloaded from their website at https://ftp.ebi.ac.uk/pub/databases/gencode/Gencode_human/release_44/gencode.v44.transcripts.fa.gz. GENCODE’s version 44 of the human genome annotations (gencode.v44.annotation.gtf.gz) can be accessed from their website at https://ftp.ebi.ac.uk/pub/databases/gencode/Gencode_human/release_44/gencode.v44.annotation.gtf.gz. Gene length information from the R object GeneBreak::ens.gene.ann.hg19 [https://www.rdocumentation.org/packages/GeneBreak/versions/1.2.0/topics/ens.gene.ann.hg19] is utilized in this study. For the exon deletion CRISPR/Cas9 screens, data from Thomas et al.¹⁸ are available in Supplementary Tables from the publication at https://static-content.springer.com/esm/art%3A10.1038%2Fs41588-019-0555-z/MediaObjects/41588_2019_555_MOESM3_ESM.xlsx. On the other hand, data from Gonatopoulos-Pournatzis et al.¹⁷ are available in Supplementary Tables from the publication at https://static-content.springer.com/esm/art%3A10.1038%2Fs41587-020-0437-z/MediaObjects/41587_2020_437_MOESM14_ESM.xlsx and https://static-content.springer.com/esm/art%3A10.1038%2Fs41587-020-0437-z/MediaObjects/41587_2020_437_MOESM15_ESM.xlsx. Raw RNA-seq.fastq files for RPE1 cell lines used in the Gonatopoulos-Pournatzis et al.¹⁷ study are available from the European Nucleotide Archive (ENA) under accession number PRJNA302748. For the KD screen of RNA-binding proteins (ENCORE), raw RNA-seq.fastq files from Van Nostrand et al.³² can be accessed from the ENCODE website at https://www.encodeproject.org/encore-matrix/?type=Experiment&status=released&internal_tags=ENCORE. Splicing event information and predicted protein impact from VastDB are available at https://vastdb.crg.eu/wiki/Downloads#Homo_sapiens_.28hg38.29. Molecular and clinical data from The Cancer Genome Atlas (TCGA) are accessible via the GDC portal [https://gdc.cancer.gov/access-data/gdc-data-portal]. Cell line doubling time data from Tsherniak et al.⁶⁴ and Ghandi et al.⁵⁵ are provided in their respective supplementary tables number 1. For drug sensitivity screens (GDSC), cell proliferation data for CCLE cell lines treated with various drugs can be downloaded from the Genomics of Drug Sensitivity (GDSC) project website at https://www.cancerrxgene.org/downloads/bulk_download. The human protein interaction network from STRING DB can be obtained from https://stringdb-static.org/download/protein.links.full.v11.5/9606.protein.links.full.v11.5.txt.gz. For structural analyzes of the interactions between MDM2, TP53 and Nutlin-3A, we downloaded crystal structures from PDB with the following accession numbers 4hfz [https://www.rcsb.org/structure/4HFZ] and 4hg7 [https://www.rcsb.org/structure/4hg7]. High-Grade Serous Ovarian Cancer (HGSOC) patient metadata from Pietilä et al.⁵⁴ is available in Supplementary Tables from their publication at https://static-content.springer.com/esm/art%3A10.1038%2Fs41467-021-24009-8/MediaObjects/41467_2021_24009_MOESM4_ESM.xlsx. Raw RNA-seq.fastq files for HGSOC are accessible from the European Genome Archive (EGA) under project number EGAD00001006456. All remaining data can be found in the Article, Supplementary and Source Data files. Source data are provided with this paper.

Code availability

To improve the reproducibility of our results, the full data analysis pipeline of this manuscript will be made publicly available at https://github.com/MiqG/publication_spotter (https://doi.org/10.5281/zenodo.12793641)⁶⁸ and the corresponding code of the spotter computational framework at https://github.com/MiqG/target_spotter (https://doi.org/10.5281/zenodo.12793649)⁶⁹.

References

Vogelstein, B. et al. Cancer genome landscapes. Science 339, 1546–1558 (2013).
Article ADS CAS PubMed PubMed Central Google Scholar
Acha-Sagredo, A., Ganguli, P. & Ciccarelli, F. D. Somatic variation in normal tissues: friend or foe of cancer early detection? Ann. Oncol. 33, 1239–1249 (2022).
Article CAS PubMed Google Scholar
Schneider, G., Schmidt-Supprian, M., Rad, R. & Saur, D. Tissue-specific tumorigenesis: context matters. Nat. Rev. Cancer 17, 239–253 (2017).
Article CAS PubMed PubMed Central Google Scholar
Hanahan, D. & Weinberg, R. A. Hallmarks of cancer: the next generation. Cell 144, 646–674 (2011).
Article CAS PubMed Google Scholar
Hanahan, D. Hallmarks of cancer: new dimensions. Cancer Discov. 12, 31–46 (2022).
Article CAS PubMed Google Scholar
Macaluso, M., Paggi, M. G. & Giordano, A. Genetic and epigenetic alterations as hallmarks of the intricate road to cancer. Oncogene 22, 6472–6478 (2003).
Article CAS PubMed Google Scholar
Terekhanova, N. V. et al. Epigenetic regulation during cancer transitions across 11 tumour types. Nature 623, 432–441 (2023).
Article ADS CAS PubMed PubMed Central Google Scholar
Rogalska, M. E., Vivori, C. & Valcárcel, J. Regulation of pre-mRNA splicing: roles in physiology and disease, and therapeutic prospects. Nat. Rev. Genet. 24, 251–269 (2023).
Article CAS PubMed Google Scholar
Bonnal, S. C., López-Oreja, I. & Valcárcel, J. Roles and mechanisms of alternative splicing in cancer — implications for care. Nat. Rev. Clin. Oncol. 17, 457–474 (2020).
Article PubMed Google Scholar
Bradley, R. K. & Anczuków, O. RNA splicing dysregulation and the hallmarks of cancer. Nat. Rev. Cancer 23, 135–155 (2023).
Article CAS PubMed PubMed Central Google Scholar
Robichaud, N., Sonenberg, N., Ruggero, D. & Schneider, R. J. Translational control in cancer. Cold Spring Harb. Perspect. Biol. 11, a032896 (2019).
Article CAS PubMed PubMed Central Google Scholar
Park, S. et al. Differential functions of splicing factors in mammary transformation and breast cancer metastasis. Cell Rep. 29, 2672–2688.e7 (2019).
Article CAS PubMed PubMed Central Google Scholar
Castelo-Branco, G. & Bannister, A. J. The epigenetics of cancer: from non-coding RNAs to chromatin and beyond. Brief. Funct. Genomics 12, 161–163 (2013).
Article PubMed Google Scholar
Bludau, I. & Aebersold, R. Proteomic and interactomic insights into the molecular basis of cell functional diversity. Nat. Rev. Mol. Cell Biol. 21, 327–340 (2020).
Article CAS PubMed Google Scholar
Wu, S. et al. ASCancer Atlas: a comprehensive knowledgebase of alternative splicing in human cancers. Nucleic Acids Res 51, D1196–D1204 (2023).
Article CAS PubMed Google Scholar
Celaj, A. et al. An RNA foundation model enables discovery of disease mechanisms and candidate therapeutics. https://doi.org/10.1101/2023.09.20.558508 (2023).
Gonatopoulos-Pournatzis, T. et al. Genetic interaction mapping and exon-resolution functional genomics with a hybrid Cas9–Cas12a platform. Nat. Biotechnol. 38, 638–648 (2020).
Article CAS PubMed Google Scholar
Thomas, J. D. et al. RNA isoform screens uncover the essentiality and tumor-suppressor activity of ultraconserved poison exons. Nat. Genet. 52, 84–94 (2020).
Article CAS PubMed PubMed Central Google Scholar
Tapial, J. et al. An atlas of alternative splicing profiles and functional associations reveals new regulatory programs and genes that simultaneously express multiple major isoforms. Genome Res. 27, 1759–1768 (2017).
Article CAS PubMed PubMed Central Google Scholar
Kiessling, M. K. et al. Identification of oncogenic driver mutations by genome-wide CRISPR-Cas9 dropout screening. BMC Genomics 17, 723 (2016).
Article PubMed PubMed Central Google Scholar
Lenoir, W. F. et al. Discovery of putative tumor suppressors from CRISPR screens reveals rewired lipid metabolism in acute myeloid leukemia cells. Nat. Commun. 12, 6506 (2021).
Article ADS CAS PubMed PubMed Central Google Scholar
McFarland, J. M. et al. Improved estimation of cancer dependencies from large-scale RNAi screens using model-based normalization and data integration. Nat. Commun. 9, 4610 (2018).
Article ADS PubMed PubMed Central Google Scholar
Gonçalves, E. et al. Drug mechanism-of-action discovery through the integration of pharmacological and CRISPR screens. Mol. Syst. Biol. 16, e9405 (2020).
Article PubMed PubMed Central Google Scholar
Chen, W.-C. et al. Targeting KRAS4A splicing through the RBM39/DCAF15 pathway inhibits cancer stem cells. Nat. Commun. 12, 4288 (2021).
Article ADS CAS PubMed PubMed Central Google Scholar
Chang, L., Ruiz, P., Ito, T. & Sellers, W. R. Targeting pan-essential genes in cancer: challenges and opportunities. Cancer Cell 39, 466–479 (2021).
Article CAS PubMed PubMed Central Google Scholar
Sondka, Z. et al. The COSMIC cancer gene census: describing genetic dysfunction across all human cancers. Nat. Rev. Cancer 18, 696–705 (2018).
Article CAS PubMed PubMed Central Google Scholar
Tate, J. G. et al. COSMIC: the catalogue of somatic mutations in cancer. Nucleic Acids Res 47, D941–D947 (2019).
Article CAS PubMed Google Scholar
Lee, E. Y. H. P. & Muller, W. J. Oncogenes and tumor suppressor genes. Cold Spring Harb. Perspect. Biol. 2, a003236 (2010).
Article CAS PubMed PubMed Central Google Scholar
Shen, L., Shi, Q. & Wang, W. Double agents: genes with both oncogenic and tumor-suppressor functions. Oncogenesis 7, 1–14 (2018).
Article Google Scholar
Zhang, Y., Qian, J., Gu, C. & Yang, Y. Alternative splicing and cancer: a systematic review. Signal Transduct. Target. Ther. 6, 1–14 (2021).
Google Scholar
Jun, Y. et al. Comprehensive analysis of alternative splicing in gastric cancer identifies epithelial–mesenchymal transition subtypes associated with survival. Cancer Res. 82, 543–555 (2022).
Article CAS PubMed Google Scholar
Van Nostrand, E. L. et al. A large-scale binding and functional map of human RNA-binding proteins. Nature 583, 711–719 (2020).
Article ADS PubMed PubMed Central Google Scholar
Page-McCaw, P. S., Amonlirdviman, K. & Sharp, P. A. PUF60: a novel U2AF65-related splicing activity. RNA 5, 1548–1560 (1999).
Article CAS PubMed PubMed Central Google Scholar
Sun, D., Lei, W., Hou, X., Li, H. & Ni, W. PUF60 accelerates the progression of breast cancer through down-regulation of PTEN expression. CMAR 11, 821–830 (2019).
Article CAS Google Scholar
Mina, M. et al. Conditional selection of genomic alterations dictates cancer evolution and oncogenic dependencies. Cancer Cell 32, 155–168.e6 (2017).
Article CAS PubMed Google Scholar
Hu, G., Wu, Z., Wang, K., Uversky, V. N. & Kurgan, L. Untapped potential of disordered proteins in current druggable human proteome. Curr. Drug Targets 17, 1198–1205 (2016).
Article CAS PubMed Google Scholar
Mirdita, M. et al. ColabFold: making protein folding accessible to all. Nat. Methods 19, 679–682 (2022).
Article CAS PubMed PubMed Central Google Scholar
Lin, Z. et al. Evolutionary-scale prediction of atomic-level protein structure with a language model. Science 379, 1123–1130 (2023).
Article ADS MathSciNet CAS PubMed Google Scholar
Wishart, D. S. et al. DrugBank: a comprehensive resource for in silico drug discovery and exploration. Nucleic Acids Res 34, D668–D672 (2006).
Article CAS PubMed Google Scholar
Otto, T. & Sicinski, P. Cell cycle proteins as promising targets in cancer therapy. Nat. Rev. Cancer 17, 93–115 (2017).
Article CAS PubMed PubMed Central Google Scholar
Rastogi, N. & Mishra, D. P. Therapeutic targeting of cancer cell cycle using proteasome inhibitors. Cell Div. 7, 26 (2012).
Article CAS PubMed PubMed Central Google Scholar
Webb, T. R., Joyner, A. S. & Potter, P. M. The development and application of small molecule modulators of SF3b as therapeutic agents for cancer. Drug Discov. Today 18, 43–49 (2013-1).
Zhong, L. et al. Small molecules in targeted cancer therapy: advances, challenges, and future perspectives. Sig Transduct. Target Ther. 6, 1–48 (2021).
Article Google Scholar
Kumar, B., Singh, S., Skvortsova, I. & Kumar, V. Promising targets in anti-cancer drug development: recent updates. Curr. Med. Chem. 24, 4729–4752 (2017).
CAS PubMed Google Scholar
Yang, W. et al. Genomics of Drug Sensitivity in Cancer (GDSC): a resource for therapeutic biomarker discovery in cancer cells. Nucleic Acids Res 41, D955–D961 (2013).
Article CAS PubMed Google Scholar
Szklarczyk, D. et al. STRING v11: protein–protein association networks with increased coverage, supporting functional discovery in genome-wide experimental datasets. Nucleic Acids Res 47, D607–D613 (2019).
Article CAS PubMed Google Scholar
Wade, M., Li, Y.-C. & Wahl, G. M. MDM2, MDMX and p53 in oncogenesis and cancer therapy. Nat. Rev. Cancer 13, 83–96 (2013).
Article CAS PubMed PubMed Central Google Scholar
Zheng, T. et al. Spliced MDM2 isoforms promote mutant p53 accumulation and gain-of-function in tumorigenesis. Nat. Commun. 4, 2996 (2013).
Article ADS PubMed Google Scholar
Jacob, A. G. et al. Stress-induced alternative splice forms of MDM2 and MDMX modulate the p53-pathway in distinct ways. PLoS One 9, e104444 (2014).
Article ADS PubMed PubMed Central Google Scholar
Rallapalli, R., Strachan, G., Tuan, R. S. & Hall, D. J. Identification of a domain within MDMX-S that is responsible for its high affinity interaction with p53 and high-level expression in mammalian cells. J. Cell. Biochem. 89, 563–575 (2003).
Article CAS PubMed Google Scholar
Bozzi, F. et al. In vitro and in silico studies of MDM2/MDMX isoforms predict Nutlin-3A sensitivity in well/de-differentiated liposarcomas. Lab. Invest. 93, 1232–1240 (2013).
Article CAS PubMed Google Scholar
Delgado, J., Radusky, L. G., Cianferoni, D. & Serrano, L. FoldX 5.0: working with RNA, small molecules and a new graphical interface. Bioinformatics 35, 4168–4169 (2019).
Article CAS PubMed PubMed Central Google Scholar
Bezzi, M. et al. Regulation of constitutive and alternative splicing by PRMT5 reveals a role for Mdm4 pre-mRNA in sensing defects in the spliceosomal machinery. Genes Dev. 27, 1903–1916 (2013).
Article CAS PubMed PubMed Central Google Scholar
Pietilä, E. A. et al. Co-evolution of matrisome and adaptive adhesion dynamics drives ovarian cancer chemoresistance. Nat. Commun. 12, 3904 (2021).
Article ADS PubMed PubMed Central Google Scholar
Ghandi, M. et al. Next-generation characterization of the cancer cell line encyclopedia. Nature 569, 503–508 (2019).
Article ADS CAS PubMed PubMed Central Google Scholar
Climente-González, H., Porta-Pardo, E., Godzik, A. & Eyras, E. The functional impact of alternative splicing in cancer. Cell Rep. 20, 2215–2226 (2017).
Article PubMed Google Scholar
Krill-Burger, J. M. et al. Partial gene suppression improves identification of cancer vulnerabilities when CRISPR-Cas9 knockout is pan-lethal. Genome Biol. 24, 192 (2023).
Article CAS PubMed PubMed Central Google Scholar
Hart, T. et al. High-resolution CRISPR screens reveal fitness genes and genotype-specific cancer liabilities. Cell 163, 1515–1526 (2015).
Article CAS PubMed Google Scholar
Boratyn, G. M. et al. Domain enhanced lookup time accelerated BLAST. Biol. Direct 7, 12 (2012).
Article CAS PubMed PubMed Central Google Scholar
Smith, P. J. et al. An increased specificity score matrix for the prediction of SF2/ASF-specific exonic splicing enhancers. Hum. Mol. Genet. 15, 2490–2508 (2006).
Article CAS PubMed Google Scholar
Ray, D. et al. A compendium of RNA-binding motifs for decoding gene regulation. Nature 499, 172–177 (2013).
Article ADS CAS PubMed PubMed Central Google Scholar
Bechara, E. G., Sebestyén, E., Bernardis, I., Eyras, E. & Valcárcel, J. RBM5, 6, and 10 differentially regulate NUMB alternative splicing to control cancer cell proliferation. Mol. Cell 52, 720–733 (2013).
Article CAS PubMed Google Scholar
Head, S. A. et al. Silencing of SRRM4 suppresses microexon inclusion and promotes tumor growth across cancers. PLoS Biol. 19, e3001138 (2021).
Article CAS PubMed PubMed Central Google Scholar
Tsherniak, A. et al. Defining a cancer dependency map. Cell 170, 564–576.e16 (2017).
Article CAS PubMed PubMed Central Google Scholar
Gillespie, M. et al. The reactome pathway knowledgebase 2022. Nucleic Acids Res 50, D687–D692 (2022).
Article CAS PubMed Google Scholar
Anil, B., Riedinger, C., Endicott, J. A. & Noble, M. E. M. The structure of an MDM2–Nutlin-3a complex solved by the use of a validated MDM2 surface-entropy reduction mutant. Acta Cryst. D. 69, 1358–1366 (2013).
Article CAS Google Scholar
Berman, H. M. et al. The protein data bank. Nucleic Acids Res 28, 235–242 (2000).
Article ADS CAS PubMed PubMed Central Google Scholar
Anglada-Girotto, M. et al. In silico RNA isoform screening to identify potential cancer driver exons with therapeutic applications - publication_spotter. Zenodo https://doi.org/10.5281/zenodo.12793642 (2024).
Anglada-Girotto, M. et al. In silico RNA isoform screening to identify potential cancer driver exons with therapeutic applications - target_spotter. Zenodo https://doi.org/10.5281/zenodo.12793650 (2024).

Download references

Acknowledgements

We thank Carolina Segura-Morales for the support in the experimental validations. We thank Javier Delgado and Damiano Cianferoni for their support in the structural analyzes. Also, we thank Juan Valcárcel, Manuel Irimia, Donate Weghorn, Xavier Hernandez-Alias, Marc Weber, and Reza Sodaei for the insightful discussions and feedback provided throughout the development of this project. We acknowledge the support of the Spanish Ministry of Science and Innovation through the Centro de Excelencia Severo Ochoa (CEX2020-001049-S, MCIN/AEI /10.13039/501100011033), and the Generalitat de Catalunya through the CERCA programme and to the EMBL partnership. We are grateful to the CRG Core Technologies Programme for their support and assistance in this work. The results shown here are in part based upon data generated by the TCGA Research Network (https://www.cancer.gov/tcga) and by Helsinki University (EGAD00001006456). Finally, we would like to thank the developer team behind VastDB as they created a resource that unifies and facilitates the splicing event-centric analysis of transcriptomes from the computational to the experimental sides. This project was funded by grants from the Plan Estatal de Investigación Científica y Técnica y de Innovación: PGC2018-101271-B-I00 and PID2021-122341NB-I00, (AEI/FEDER, UE) to L.S., and PID2021-122341NB-I00 project funded MICIU /AEI /10.13039/501100011033 / (FEDER, UE) to L.S.

Author information

Authors and Affiliations

Centre for Genomic Regulation (CRG), The Barcelona Institute of Science and Technology, Dr. Aiguader 88, Barcelona, 08003, Spain
Miquel Anglada-Girotto, Ludovica Ciampi, Sophie Bonnal, Sarah A. Head, Samuel Miravet-Verde & Luis Serrano
Department of Biology, Institute of Microbiology and Swiss Institute of Bioinformatics, ETH Zurich, Zurich, Switzerland
Samuel Miravet-Verde
Universitat Pompeu Fabra (UPF), Barcelona, Spain
Luis Serrano
ICREA, Pg. Lluís Companys 23, Barcelona, 08010, Spain
Luis Serrano

Authors

Miquel Anglada-Girotto
View author publications
You can also search for this author in PubMed Google Scholar
Ludovica Ciampi
View author publications
You can also search for this author in PubMed Google Scholar
Sophie Bonnal
View author publications
You can also search for this author in PubMed Google Scholar
Sarah A. Head
View author publications
You can also search for this author in PubMed Google Scholar
Samuel Miravet-Verde
View author publications
You can also search for this author in PubMed Google Scholar
Luis Serrano
View author publications
You can also search for this author in PubMed Google Scholar

Contributions

Conceptualization: M.A.-G., L.S., S.H., S.M.-V.; Methodology: M.A.-G., L.S., S.H., S.M.-V.; Software: M.A.-G.; Validation: M.A.-G., L.C., S.B., L.S., S.H., S.M.-V.; Formal analysis: M.A.-G., L.S., S.H., S.M.-V.; Investigation: M.A.-G., L.C.; Resources: L.S.; Data Curation: M.A.-G.; Writing-original draft: M.A.-G.; Writing-review and editing: M.A.-G., L.C., S.B., L.S., S.H., S.M.-V.; Visualization: M.A.-G., L.C.; Supervision: L.S., S.M.-V., S.H.; Funding acquisition: L.S. All authors read and approved the final manuscript.

Corresponding authors

Correspondence to Miquel Anglada-Girotto, Samuel Miravet-Verde or Luis Serrano.

Ethics declarations

Competing interests

All authors including M.A.-G., L.C., S.B., S.H., S.M.-V., and L.S. are listed as inventors in the European priority patent application NO. 24382535.3.

Peer review

Peer review information

Nature Communications thanks IHAB YOUNIS, and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary information

Supplementary Information

Reporting Summary

Peer Review File

Source data

Source Data

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.

Reprints and permissions

About this article

Cite this article

Anglada-Girotto, M., Ciampi, L., Bonnal, S. et al. In silico RNA isoform screening to identify potential cancer driver exons with therapeutic applications. Nat Commun 15, 7039 (2024). https://doi.org/10.1038/s41467-024-51380-z

Download citation

Received: 12 July 2023
Accepted: 06 August 2024
Published: 15 August 2024
DOI: https://doi.org/10.1038/s41467-024-51380-z
Springer Nature Limited

In silico RNA isoform screening to identify potential cancer driver exons with therapeutic applications

Abstract

Similar content being viewed by others

Introduction

Results

Genome-wide identification of high-confidence potential cancer driver exons

Potential cancer driver exons enlarge the universe of potential cancer driver genes

Splicing dependency models recapitulate single and combinatorial splicing perturbations

Streamlined discovery of splicing-based anticancer therapies

Potential cancer driver exons mechanistically mediate drug sensitivity

Splicing-based treatment recommendation

Discussion

Methods

Quantification of splicing and gene expression

Identification of potential cancer driver exons

Data preprocessing

Compilation of lists of known cancer driver exons, mutational cancer driver genes, and pan-essential genes

Training models of splicing dependency

Evaluation of predictive power of LR p-values

Evaluation of biases in Demeter2 gene dependency scores due to shRNA targeting

Evaluation of uncertainty of predicted splicing dependencies

Model selection

Gene set overlap analysis

Gene-level mutation frequency

Simulating single-exon perturbations with harm scores

Validation of the predictive power of splicing dependency models for single exon perturbations

Data preprocessing

Harm score computation

Evaluation

Prediction of cancer cell proliferation from combinatorial splicing changes

Data preprocessing

Combined harm score computation

Evaluation

Systematic prioritization of targetable cancer driver exons

Data preprocessing

Differential splicing analysis

Splicing dependency analysis

Prioritization criteria

Experimental validation of selected targetable cancer driver exons with therapeutic potential

In-house RNA-seq of the cell lines used in the manuscript

Selection of exons for experimental validation

Analysis of the impact of selected exons on their main proteoform

Design and validation of SSOs

Cell culture and SSO transfection

Validation of oncoexons by RT-PCR

Clonogenic assay

Resazurin cell proliferation assay

Splicing dependency analysis of HT29 cells with high- and low-proliferation models

Drug-exon statistical interactions

Data preprocessing

Training models of drug-exon interaction

Evaluation of mechanistic drug-exon interactions

Data preprocessing

Pathway-level assessment

Protein-protein interaction assessment

Exploring drug-exon mechanistic interactions mediating Nutlin-3A’s effectiveness

HsaEX0038400_MDM2 protein structure stability analysis

Estimation of drug sensitivity with drug-exon interaction models

Validation of drug-exon interaction models for personalized drug recommendation

In vitro predictive power

Clinical predictive power

Statistics & reproducibility

Reporting summary

Data availability

Code availability

References

Acknowledgements

Author information

Authors and Affiliations

Contributions

Corresponding authors

Ethics declarations

Competing interests

Peer review

Peer review information

Additional information

Supplementary information

Source data

Rights and permissions

About this article