Cells defend themselves against drugs with membrane pumps — ATP-binding cassette (ABC) transporters such as P-glycoprotein (ABCB1) — that actively expel chemotherapeutics before they reach their targets, producing multidrug resistance. RNA sequencing (RNA-seq) measures how strongly every gene is read into RNA, and the field's reflex is to look up pump-gene abundance to explain or predict resistance. But three observations in the literature say this reflex fails. The Bruton's tyrosine kinase inhibitor RN486 fully reverses paclitaxel resistance in ABCB1-overexpressing cells — the resistance fold plummets from 707.30 to 1.30 — while leaving ABCB1 transcript and protein levels untouched RN486. Conversely, CRISPR knockout of the kinase CDK6 disables ABCB1 without shutting off its transcription: the cell still makes ABCB1 pre-mRNA, but 14 exon-skipping events splice it into truncated proteins that never reach the membrane CDK6-PI3K. And at the prediction level, RNA-seq underperforms proteomics for drug response because RNA and protein levels correlate only weakly (median Pearson $r$ between 0.436 and 0.539) BDRN.
The proposal here is that efflux-mediated resistance is encoded in the transcriptome, but not where single-gene differential expression looks for it. It lives in two places current pipelines ignore: the splicing layer (which exons are included in pump and regulator transcripts) and the pathway layer (the coordinated spliceosome, hypoxia, and redox programs that sustain pump function). I propose to build and test a single model that makes both layers first-class inputs — an extension of the regularized enrichment model pareg pareg — and to evaluate it on the one existing dataset where the ground truth is known: the deep RNA-seq and splicing data from CDK6/PI3K-deficient multidrug-resistant cancer cells CDK6-PI3K.
Core Claim
Efflux-mediated drug resistance can be recovered and predicted from bulk RNA-seq, provided the analysis models exon-level splicing events and inter-dependent pathways jointly rather than reading off ABC-transporter mRNA abundance. Concretely, I claim three testable things: that a splicing-aware, dependency-aware enrichment model run on the CDK6/PI3K knockout RNA-seq recovers the known ABCB1 exon-skipping events and nominates the upstream splicing factors the original study could not identify; that pathway-level features predict resistance to efflux-substrate drugs in cell-line panels better than pump-gene abundance does; and that pathway-plus-splicing features close part of the documented gap between transcriptomic and proteomic drug-response prediction (Pearson $r$ of 0.5311 for RNA versus 0.5467 for protein models) BDRN.
Background: What Each Prior Result Does and Does Not Give Us
Four papers form the load-bearing prior work. The CDK6-PI3K study CDK6-PI3K is the motivating evidence: it connects deep RNA-seq, differential expression (8,974 genes downregulated on CDK6 knockout), MATS-based splicing detection (14 skipped-exon events in ABCB1 pre-mRNA, 10 conserved across CDK6-, PI3K p110α-, and p110β-deficient cells), and complete in vivo tumor elimination in 6 of 9 CDK6-deficient mice on doxorubicin. But it stops at the phenotype: it neither identified the splicing factors bridging the kinase axis to the spliceosome nor applied any functional enrichment to its own differential expression output — it explicitly leaves this integration open. The pareg paper pareg supplies the method I will extend: it treats pathway enrichment as a sparse regularized regression over gene-level differential-expression p-values, with a network-fusion penalty coupling overlapping pathways, and it outperformed Fisher's exact test, topGO, MGSA, and blitzGSEA in mean area under the precision-recall curve on synthetic benchmarks — though no numeric value is reported for that AUC, and it has never been applied to pharmacogenomic or efflux-resistance data, nor to anything but gene-level p-values. The hypoxia studies CoCl2-LOVO and UCP2-NSCLC constrain the biology: efflux pumps are embedded in a HIF-1α/PPARγ/ROS/Nrf2 cascade with metabolic and apoptotic co-regulation (Bcl-2/Bax shifts, glucose uptake), all measured only by targeted assays — never genome-wide. Finally, two transfer-learning systems show what bulk and single-cell RNA-seq prediction currently achieves without any of this: VAE_LD, a residual variational autoencoder with knowledge distillation that reaches an average AUC of 0.81 and F1 of 0.92 on the Genomics of Drug Sensitivity in Cancer (GDSC) benchmark across 72 drugs VAE_LD, and SCAD, which transfers drug-sensitivity labels from bulk to single-cell data and predicts pre-existing resistant subpopulations at AUC 0.831 SCAD. Neither isolates efflux-mediated resistance from other mechanisms, and neither uses functional enrichment at all — the exact gap this proposal fills.
Research Program
The first component is an efflux-focused pathway catalog. Standard Gene Ontology and KEGG categories treat "ABC transport" as one term among thousands; I will build a compact ontology that adds the regulator programs documented in the literature — spliceosome and RNA-binding-protein categories (motivated by the ABCB1 splicing result), the HIF-1α/PPARγ/Nrf2 hypoxia axis, PI3K and p38 MAPK signaling, and apoptosis-threshold terms — so that the enrichment model has somewhere to put the coordinated programs that single-pump assays miss. This is curation plus database work, not new data collection.
The second component is the model itself. pareg fits pathway activity as a sparse linear function of gene-pathway memberships on gene-level p-values. I extend it with exon-event predictors, so that a pathway's activity can be driven both by which member genes change and by which exons are skipped in them:
$$ y_p \;=\; \sum_{g} X_{g p}\, \beta_g \;+\; \sum_{e} Z_{e p}\, \gamma_e \;+\; \lambda_1 \|\beta\|_1 \;+\; \lambda_2\, \beta^\top S\, \beta, $$
where $y_p$ is the activity score of pathway $p$ derived from differential-expression p-values, $X_{gp}$ indicates whether gene $g$ belongs to pathway $p$, $Z_{ep}$ indicates whether exon-skipping event $e$ falls in a member gene of pathway $p$, $\beta$ and $\gamma$ are the fitted weights, and $S$ is a pathway-similarity matrix (Jaccard or semantic overlap) so that overlapping pathways are estimated coherently rather than redundantly, exactly as in pareg pareg. The key new ingredient is the $\gamma$ block: it lets the model attribute pathway dysregulation to splicing events, and the splicing events with the largest weights point directly at candidate regulatory splicing factors. Because pareg's runtime is already substantial (up to ~30 minutes in parallelized mode), I will pre-filter exon events to those in catalog genes and fit with warm-started regularization paths to keep the extension tractable.
The third component is analysis and prediction. I will re-analyze the CDK6/PI3K knockout RNA-seq (KB-C2 and H460/MX80 multidrug-resistant lines) gene-level with a negative-binomial differential-expression model of the DESeq family — the count model that, unlike fold changes or t-tests, accounts for variance growing with the mean — following the standard pipeline exemplified by the Osmanthus transcriptome study Osmanthus, and event-level with a MATS-style skipped-exon detector. Then, on GDSC (72 drugs, 1,280 cell lines VAE_LD), I will define an efflux-substrate prediction task: for drugs that are known ABCB1/ABCG2 substrates (paclitaxel, doxorubicin, and analogues), predict resistance from three feature sets — pump-gene abundance alone, pareg-style pathway scores alone, and the full splicing-plus-pathway model — under identical folds and tuning budgets. TCGA provides the translational check: whether the fitted efflux-pathway scores stratify patient survival on efflux-substrate regimens, mirroring how VAE_LD's explainable-AI gene sets stratified cohorts across 1,836 patients in five cancer types VAE_LD.
The fourth component is a small validation layer. The model's top-ranked splicing factors will be knocked down in KB-C2 cells to test whether resistance reverses and whether the 10 conserved ABCB1 skipped-exon events appear — the direct test of whether the model found the bridge that the CDK6 study left unidentified. Separately, a chronic-exposure time course under RN486 will test whether inhibitor pressure drives compensatory pathway shifts (upregulation of ABCG2 or ABCC1 programs) that the RN486 study explicitly left unaddressed RN486.
Baselines
The naive reference point is pump-gene mRNA abundance: predict efflux resistance from ABCB1, ABCG2, and ABCC1 expression alone. This is what most resistance studies implicitly do, and it is cheap and standardized. It is the right "do nothing" baseline because the literature already documents its two failure modes: resistance can be fully reversible with zero abundance change RN486, and abundant pre-mRNA can coexist with a nonfunctional pump due to splicing CDK6-PI3K. Beating it means strictly higher prediction accuracy on the held-out GDSC efflux-substrate folds, with confidence intervals, under the same features-per-sample accounting.
The strongest methodological competitor is pareg itself, unmodified pareg: same regularized regression core, same network-fusion penalty, but gene-level p-values only, no exon-event terms, and the standard GO/KEGG catalog rather than my efflux-focused one. This baseline isolates exactly my two additions. It will be run on the same KB-C2/H460 data and the same GDSC folds, with the same cross-validation budget for its regularization hyperparameters, and beaten only if the extension recovers more of the 10 conserved ABCB1 skipped-exon events at the same false-discovery level and predicts resistance more accurately. The file reports that pareg outperformed Fisher's exact test, topGO, MGSA, and blitzGSEA on synthetic benchmarks in mean precision-recall AUC but gives no numeric value; no reported result found on Lacuna for pareg's numeric mean PR-AUC on that benchmark, so I will re-run those comparisons myself on the KB-C2 data.
The strongest predictive competitor is VAE_LD VAE_LD, which reports an average AUC of 0.81 and F1 of 0.92 on the full GDSC benchmark across 72 drugs. Those numbers are on the full 72-drug task, not the efflux-substrate subset, so for a fair comparison I will retrain and tune VAE_LD on the identical efflux-substrate folds and compute its AUC there; its published 0.81 serves as the reference point for what deep transfer learning achieves in the easier, mechanism-mixed setting. For the regression framing, the numbers to beat are from the Bi-modal Drug Response Network comparison BDRN: a transcriptomic model reaching median Pearson $r = 0.5311$ on log-$IC_{50}$ prediction, against a proteomic ceiling of $r = 0.5467$. The decisive result for my second claim is a transcriptomic model whose pathway-plus-splicing features push past 0.5311 and toward the proteomic ceiling on efflux-substrate drugs.
Finally, two ablations isolate the key ingredients: the full model with the network-fusion penalty removed (a plain LASSO regression, testing whether pathway-dependency modeling matters, as pareg's authors claim), and the full model with the exon-event block zeroed out (testing whether splicing contributes beyond pathway aggregation). If either ablation matches the full model, the corresponding ingredient is cut from the claim.
Why not just…?
Why not just co-administer a pump inhibitor like RN486 or lapatinib? Because the clinical record of direct pump blockers is poor and the mechanism paper itself documents why: they suffer dose-limiting toxicity, disrupt physiological transporters in the blood-brain barrier, liver, and kidney, and are often cleared as pump substrates themselves CDK6-PI3K. The comparison with chemical CDK4/6 inhibitors is instructive — ribociclib was far less effective at reversing resistance than genetic CDK6 knockout, because knockout disables the pump through splicing rather than transient occupancy. My proposal is complementary, not competing: the RN486 study explicitly leaves open whether chronic exposure triggers compensatory adaptation of other ABC transporters RN486, and my time-course experiment is designed to answer exactly that question, which no inhibitor study can answer without transcriptomics.
Why not just run qPCR and Western blots for the three pump genes? Because targeted assays measure abundance, and the central finding of this literature is that abundance is the wrong observable. The CDK6 study showed cells transcribing plentiful ABCB1 pre-mRNA while producing truncated, membrane-incompetent protein through 10 conserved exon-skipping events — invisible to any assay that counts transcript copies CDK6-PI3K. The hypoxia studies similarly show that pump upregulation is one node of a coordinated program spanning metabolism, redox signaling, and apoptosis thresholds CoCl2-LOVO, UCP2-NSCLC. Targeted assays are the right validation tool — I use them for the splicing-factor knockdown — but they cannot discover what to target, which is the point of the genome-wide model.
Why not just run standard overrepresentation analysis (Fisher's exact test, topGO)? This is the most serious objection, because ORA is fast, universal, and would take seconds where my model takes tens of minutes. The pareg benchmarking gives principled reason to expect it to fail: ORA treats pathways as independent, so overlapping categories (ABC transport, xenobiotic detoxification, lipid translocation) produce redundant term lists and inflated false discoveries, and pareg consistently outperformed Fisher's exact test and topGO in mean precision-recall AUC across noise levels pareg — though again, no numeric AUC is reported in the file. More fundamentally, ORA cannot represent splicing events at all, since it operates on thresholded gene lists, discarding the continuous evidence and the exon-level signal that my core claim rests on. That said, I will run Fisher's exact test on the KB-C2 differential expression output as a baseline rather than argue around it: if plain ORA recovers the ABCB1 splicing story as cleanly as the extended model, the pathway-modeling machinery is not earning its cost and the claim must be weakened. I consider this unlikely precisely because ORA has no representation for the phenomenon, but the experiment settles it.
Proposed Experiments and Decisive Checks
Experiment one is the recovery test on the KB-C2/H460/MX80 knockout data CDK6-PI3K. The model is run without being told the answer and must recover the 10 conserved ABCB1 skipped-exon events among its top-ranked splicing attributions at a pre-registered false-discovery threshold, and its ranked splicing-factor nominations must include factors whose knockdown reproduces the splicing phenotype. This is the cleanest ground-truth check in the corpus because the answer is already known from the original study.
Experiment two is the prediction test on GDSC. On the efflux-substrate drug subset, the full model must beat the pump-abundance baseline and the retrained VAE_LD, and in the regression framing push the transcriptomic Pearson $r$ above the 0.5311 transcriptomic figure toward the 0.5467 proteomic figure reported for BDRN BDRN. A secondary check uses SCAD-style single-cell transfer SCAD: whether the pathway signature detects pre-existing resistant subpopulations before drug exposure, distinguishing rare pump-driven clones from inducible network-wide adaptation — the open question SCAD left for efflux specifically.
Experiment three is the translational test on TCGA: whether fitted efflux-pathway scores stratify patient survival on efflux-substrate regimens, benchmarked against the stratification that VAE_LD achieved with explainable-AI gene signatures VAE_LD.
Experiment four is the mechanistic closure: knockdown of the model's top-ranked splicing factors in KB-C2 cells, with resistance-fold and exon-skipping assays as endpoints, plus the chronic RN486 exposure time course to map compensatory adaptation. The proposal succeeds if experiments one and two confirm the core claim, and the mechanistic experiment then turns the model's nominations into the splicing-factor identification that the CDK6-PI3K study explicitly left open — the missing link between kinase signaling and pump integrity.
References
- Bruton’s Tyrosine Kinase (BTK) Inhibitor RN486 Overcomes ABCB1-Mediated Multidrug Resistance in Cancer Cells
- CDK6-PI3K signaling axis is an efficient target for attenuating ABCB1/P-gp mediated multi-drug resistance (MDR) in cancer cells
- Drug Response Modeling across Cancers: Proteomics vs. Transcriptomics
- Coherent pathway enrichment estimation by modeling inter-pathway dependencies using regularized regression
- The hypoxia-mimetic agent CoCl2 induces chemotherapy resistance in LOVO colorectal cancer cells
- Uncoupling protein 2 downregulation by hypoxia through repression of peroxisome proliferator-activated receptor γ promotes chemoresistance of non-small cell lung cancer
- Interpretable Transfer Learning for Cancer Drug Resistance: Candidate Target Identification
- Enabling Single‐Cell Drug Response Annotations from Bulk RNA‐Seq Using SCAD
- Transcriptomic Analysis of the Candidate Genes Related to Aroma Formation in Osmanthus fragrans