Alternative splicing — the process by which one gene's RNA is cut and re-joined into different protein isoforms — decides which version of a protein a cell makes, yet we have almost no way to watch that decision happen. The clearest symptom of this blindness is documented in a study of PARP-inhibitor resistance in ovarian cancer: the splicing kinase CLK1 controls which ERCC1 isoform a cell expresses, but no commercially available antibody can tell the functional ERCC1-202 isoform from the others, so isoform-specific validation stalls (Genome-wide CRISPR-Cas9 screening identifies CLK1 inhibition as a strategy to restore PARP inhibitor sensitivity via ERCC1 isoform switching, Genome-wide CRISPR-Cas9 screening identifies CLK1 inhibition as a strategy to restore PARP inhibitor sensitivity via ERCC1 isoform switching). Meanwhile, every method we have for finding the master regulators of splicing is static and purely computational: GLIMS ranks cancer driver genes by running PageRank over co-splicing networks built from bulk RNA-seq (GLIMS: A two-stage gradual-learning method for cancer genes prediction using multi-omics data and co-splicing network), and ESPACE relaxes a sparsity penalty around pre-defined "hub" genes to reconstruct regulatory networks (Enhanced construction of gene regulatory networks using hub gene information). Neither touches a living cell's dynamics.
The same blindness limits peptide manufacturing. Scalable production of bioactive peptides — antimicrobials in particular — is stuck behind an economic gap: the best low-cost pipeline, an E. coli fusion-carrier system with acid-cleavable tags, reached 29.3 mg/L at 96% purity for the peptide Kiadin but cannot handle peptides with internal Asp-Pro motifs and cannot fold disulfide-rich defensins, the class that includes plectasin, whose yields must rise by orders of magnitude to compete with cheap antibiotics in agriculture (Zheng et al., Porto et al., Zhang et al.). Disulfide-rich peptides need eukaryotic hosts, and eukaryotic hosts splice — but nobody can see, let alone engineer, the splicing inside a production strain.
Core Claim
I propose a closed-loop platform with three parts. First, a panel of ratiometric minigene reporters that turn isoform choice into a live optical signal: each reporter carries an alternative exon wired so that exon inclusion puts firefly luciferase in frame and exon skipping puts Renilla luciferase in frame, so the ratio of the two light outputs directly measures the percentage spliced in (PSI). This reuses the ratiometric noise-normalization logic of dual-luciferase transcription assays (Wei Sun) but points it at splicing — something no luciferase study in this literature does, including the homoharringtonine work where luciferase measured only promoter output (The small-molecule drug homoharringtonine targets HSF1 to suppress pancreatic cancer progression). Second, the reporter panel is read out over time across a systematic perturbation panel (siRNA, CRISPRi, and compounds targeting splicing kinases, phosphatases, SR proteins, and hnRNPs), and the resulting time-resolved perturbation-response network is ranked by betweenness centrality — the fraction of shortest paths passing through a node — rather than by degree or PageRank. Third, the top-ranked gatekeeper regulators are engineered in a eukaryotic production host expressing a defensin-class peptide, and purified yield is the endpoint.
The reason for betweenness rather than degree is concrete: Xiao Dong et al. showed that bridge hubs coordinate cross-talk between disease modules without high degree or large expression changes, and CLK1 itself is exactly this kind of moderate-connectivity bottleneck — invisible to fold-change filtering and plausibly under-ranked by PageRank. GLIMS's own stage-2 design already admits that splicing information reorders driver rankings, but it computes static partial correlations on bulk cohorts and validates nothing in living cells; we measure perturbation responses directly and feed real kinetics back into the ranking.
Research Program
The first deliverable is the reporter construct itself, and the most defensible first target is ERCC1, because the field already knows the answer we must reproduce: TG003-mediated CLK1 inhibition shifts splicing away from the ERCC1-202 isoform in OVCAR8 cells, quantified laboriously by bulk rMATS RNA-seq and confounded by absent isoform-specific antibodies (Genome-wide CRISPR-Cas9 screening identifies CLK1 inhibition as a strategy to restore PARP inhibitor sensitivity via ERCC1 isoform switching, Genome-wide CRISPR-Cas9 screening identifies CLK1 inhibition as a strategy to restore PARP inhibitor sensitivity via ERCC1 isoform switching). The reporter converts this into a plate-reader measurement. Reporter PSI is the calibrated ratio
$$\Psi_{\text{rep}} = \frac{F}{F + \gamma R},$$
where $F$ and $R$ are firefly and Renilla photon rates and $\gamma$ is a per-cell-line calibration constant fixed once against matched RNA-seq PSI values. Reporter exon choice will follow the design principles the splicing literature has established: splice-site thermodynamic strength gates how much a factor's effect depends on trans-acting recruitment (Choi et al.), and competing RNA-binding proteins set the baseline inclusion level (Feng et al.), so we will deliberately include exons with a range of splice-site strengths to keep the panel sensitive to both kinase-driven and competition-driven regulation. A second construct family targets Bcl-x and caspase 9, whose isoform switches under homoharringtonine treatment are PP1-dependent (Sun et al.).
The second deliverable is the perturbation network. Roughly 100 regulators — the SR and hnRNP families plus the kinases and phosphatases that set their phosphorylation state (CLK1, PP1, and their relatives) — will be perturbed one at a time, with reporter readouts at multiple time points, in both a cancer line (OVCAR8) and a production-relevant line (CHO). A directed edge is drawn from each regulator to each reporter exon whose PSI it changes beyond a pre-registered effect threshold. We then compare three rankings of the same network: degree, PageRank (GLIMS's choice), and betweenness. The pre-registered ground truth is the ovarian-cancer CRISPR screen: CLK1 was its top synthetic-lethal hit, so a ranking method is judged by whether it places CLK1 and the PP1 catalytic subunit near the top using far fewer experiments than a genome-wide screen.
The third deliverable is the manufacturing application. The top betweenness-ranked splicing regulator in the CHO network becomes the engineering target: overexpression or knockdown of the chosen regulator (or a small-molecule modulator of it) in a CHO strain expressing a plectasin-like defensin fused to a protective carrier, purified by a non-chromatographic heat-and-isoelectric-precipitation workflow adapted from Zheng et al.. The success criterion is quantitative: purified yield in mg/L at stated purity, MIC against E. coli K88 and MRSA, with the Kiadin benchmark defining the bar.
Baselines
The "do nothing" reference point is the unengineered CHO strain expressing the same defensin construct under the same fermentation conditions, plus a random-gene perturbation arm (non-targeting and unrelated-gene knockdowns) to establish the noise floor of the reporter panel. Any claimed yield improvement must exceed this floor, and any claimed hub effect must exceed what random perturbations produce.
The strongest competitor for isoform measurement is endpoint bulk RNA-seq with rMATS quantification, exactly the method the CLK1 study used to discover the ERCC1-202 switch; for targeted validation, isoform-specific qRT-PCR. For the comparison to be fair, the same cells, same perturbations, and same time points will be assayed by reporter and by RNA-seq, and the reporter is only credited with an advantage where a full time course is required. The acceptance bar for reporter fidelity is agreement with RNA-seq PSI at the reliability level the field already accepts for splicing models: CellSpliceNet reports a Spearman correlation of ρ = 0.88 on held-out data, so a reporter-vs-RNA-seq Spearman of at least 0.8 across the perturbation time course is the pre-registered threshold.
The strongest competitor for hub ranking is GLIMS's stage-2 approach: PageRank centrality over a co-splicing network, run here on our own perturbation data rather than on static cohorts, so both methods see identical inputs and compute budgets (, GLIMS: A two-stage gradual-learning method for cancer genes prediction using multi-omics data and co-splicing network). GLIMS's stage-1 multi-omics GCN achieved accuracy 0.858, AUROC 0.904, and AUPRC 0.768 on pan-cancer data, with the co-splicing refinement consistently improving on it; no reported result found on Lacuna for the cohort-specific values, and no reported result found on Lacuna for ESPACE's exact FDR and MCC numbers against RegulonDB, so those methods are compared here by recovery of CRISPR-validated regulators, not by transferable scores. Beating the baseline means betweenness ranks CLK1 and PP1 in the top few percent with fewer perturbations than PageRank needs, and that betweenness-ranked targets deliver yield gains that degree-ranked targets do not. The key ablation is reranking the identical perturbation-response network with degree, PageRank, and betweenness — this isolates the centrality definition as the only variable. A second ablation builds the network from endpoint RNA-seq instead of reporter time courses, isolating the value of dynamics.
The manufacturing reference point is the Kiadin pipeline: 29.3 mg/L purified at 96% purity, MIC 32 μg/mL against E. coli K88, zero macrophage cytotoxicity up to 128 μg/mL, survival raised from 0% to 50% in lethal murine sepsis, and over 98% reduction of organ bacterial burden (Zheng et al.). Beating this bar means at least comparable mg/L and purity in the eukaryotic host for a disulfide-rich peptide that the E. coli system cannot fold, not merely a yield above the unengineered CHO floor.
Why not just…?
Why not just measure isoforms by qRT-PCR or rMATS RNA-seq? This is the most serious objection, and for endpoint measurements it is partly right — the CLK1 study needed no reporter to find CLK1. The honest answer is that the reporter earns its keep only where time resolution and throughput matter: a perturbation panel of ~100 regulators across a multi-time-point course is prohibitive at RNA-seq depth per sample, while a ratiometric plate readout covers it in a fraction of the cost and self-normalizes transfection noise, the specific confounder dual-luciferase assays were invented to remove (Wei Sun). The episomal caveat is real: the PIK3CA 3' UTR reporter contradicted patient PBMC expression data (A Functional Variant at miR-520a Binding Site in PIK3CA Alters Susceptibility to Colorectal Cancer in a Chinese Han Population), so every reporter finding will be checked against endogenous RNA-seq, and the platform claims validity only where reporter and endogenous measurements agree.
Why not just predict the perturbation effects computationally, as CellSpliceNet predicts PSI? CellSpliceNet integrates sequence, RNA secondary structure, junction regions, and splicing-factor expression to reach ρ = 0.88 in C. elegans neurons, and its ablations show every modality matters (dropping any one costs at least 0.06 in Spearman correlation). But it predicts splicing outcomes for a fixed expression state; it does not rank regulators by causal importance, and it has never been applied to human or production-host cells. It is a genuine threat as a perturbation-prioritization heuristic, so it will be run as a baseline: if its predicted sensitivity patterns identify CLK1-class regulators as well as our measured network, the wet-lab loop loses part of its justification. We use its insights — splice-site strength dependence and junction-focused encoding — to design reporter exons, which is a role it can already play well.
Why not just apply GLIMS-style PageRank to existing RNA-seq and skip the experiments? GLIMS already showed splicing data reorders driver rankings, so the objection has force. But its co-splicing network is built from static partial correlations with no experimental validation, and the betweenness evidence says the regulators we most need — bridge nodes like CLK1 — are precisely the ones raw connectivity measures miss (Xiao Dong et al.). The question is settled empirically, not rhetorically: the same perturbation data reranked by all three centrality measures. If PageRank ranks CLK1 as well as betweenness does, the betweenness claim should be abandoned and the loop reduced to a measurement contribution only.
Why not just stay in E. coli with the DAMP4-DP carrier system? For short, linear peptides, yes — 29.3 mg/L at 96% purity is the number to beat and the method to keep (Zheng et al.). But the system's own authors note the acid-cleavage protocol fails for peptides with internal Asp-Pro motifs and needs further work for longer and proline-rich sequences, and disulfide-rich defensins require oxidative folding the E. coli cytosol does not provide (Porto et al.); lasso peptides face the same host ceiling (Li et al.). The eukaryotic move is justified only for that peptide class, and if engineered CHO yields remain far below the E. coli benchmark for a peptide E. coli can handle, the application claim fails on its own terms.
Proposed Experiments and Decisive Checks
Experiment one validates the ERCC1 splicing reporter in OVCAR8 cells under TG003 and CLK1 knockdown, comparing reporter PSI to rMATS and isoform-specific qRT-PCR; acceptance is Spearman ≥ 0.8 against RNA-seq PSI across the time course. Experiment two builds the perturbation-response network and performs the three-way ranking comparison, with recovery of CLK1 and PP1 as the pre-registered success criterion. Experiment three uses the platform to resolve the standing contradiction in the homoharringtonine literature: Li et al. attribute HHT's action to direct binding of the transcription factor HSF1 (IC50 ~18–72 nM), while Sun et al. attribute it to PP1 activation and splicing shifts of Bcl-x and caspase 9. Running an HSF1-driven firefly promoter reporter and the Bcl-x splicing reporter in the same cells, with and without a PP1 inhibitor, tells us which event precedes the other on a single time axis — something neither endpoint study could see. Experiment four is the CHO engineering outcome: purified mg/L, purity, MIC against E. coli K88 and MRSA Mu50, with the Kiadin numbers as the reference bar and cell-culture-scale yield ratios reported against the unengineered strain.
The whole program stands or falls on two checks: the reporter must agree with endogenous isoform measurements wherever compared, and betweenness must demonstrably outrank degree and PageRank on the same data. Either failure has a graceful downgrade — to a validated measurement tool alone, or to a kinetic-network method in need of a better centrality — but both checks are cheap enough to run first.
References
- Genome-wide CRISPR-Cas9 screening identifies CLK1 inhibition as a strategy to restore PARP inhibitor sensitivity via ERCC1 isoform switching
- GLIMS: A two-stage gradual-learning method for cancer genes prediction using multi-omics data and co-splicing network
- Enhanced construction of gene regulatory networks using hub gene information
- High-level biosynthesis and purification of the antimicrobial peptide Kiadin based on non-chromatographic purification and acid cleavage methods
- U4 at the 3′ UTR of PB1 Segment of H5N1 Influenza Virus Promotes RNA Polymerase Activity and Contributes to Viral Pathogenicity
- The small-molecule drug homoharringtonine targets HSF1 to suppress pancreatic cancer progression
- Relative strength of 5’ splice-site strength defines functions of SRSF2 and SRSF6 in alternative splicing of Bcl-x pre-mRNA
- hnRNP A1 in RNA metabolism regulation and as a potential therapeutic target
- Homoharringtonine regulates the alternative splicing of Bcl-x and caspase 9 through a protein phosphatase 1-dependent mechanism
- CellSpliceNet: Explainable Multimodal Transformers for Splicing in C. elegans Neurons
- A Functional Variant at miR-520a Binding Site in PIK3CA Alters Susceptibility to Colorectal Cancer in a Chinese Han Population
- Effect of oral administration of microcin Y on growth performance, intestinal barrier function and gut microbiota of chicks challenged with Salmonella Pullorum