Isoform Switching on the Inflammation-to-Cancer Timeline

A single gene does not spell out a single protein. By stitching its RNA pieces together in different orders — a process called alternative splicing — one gene can yield several distinct proteins, and those proteins can do opposite things: one version of a receptor mops up a growth signal and restrains a tumor, while a sibling version delivers the same signal and feeds it. Almost everything we know about this comes from static snapshots: a tumor and its neighbor tissue, photographed once, years after the disease began. Meanwhile, a separate literature on inflammation-driven cancer has worked out how chronic cytokine exposure — IL-6, prostaglandin E2, sustained NF-κB signaling — pushes cells toward malignancy, but it almost entirely ignores splicing, treating each gene as one unit with one output. Nobody has watched, over time, how a chronically inflamed tissue rewrites its isoform choices, in which cells, under which cytokine, and whether that rewrite precedes the onset of cancer-like behavior.

The proposal here is to do exactly that, and to make the result predictive rather than descriptive. The core hypothesis is that in inflammation-driven oncogenesis, the earliest molecular rewiring happens at the level of which isoform a gene produces, not how much of the gene is expressed — and that this isoform choice is a deterministic-enough function of the inflammatory context (cytokine identity, dose, source, cell type, and time) to be learned by a model. Concretely, we will build a longitudinal, isoform-resolution, single-cell atlas of the colitis-to-colorectal-cancer transition in mice, spanning an in vitro cytokine-dose-by-time surface in human colonic organoids, and train a model that predicts isoform proportions from a continuous description of inflammatory context. Genes whose isoform switch demonstrably precedes the stemness shift along the trajectory become causal candidates, which we test by blocking the switch with splice-targeting drugs. Context-dependency is treated as a first-class axis throughout: the same cytokine is expected to induce different isoform programs in epithelial versus immune cells, mirroring known cases where a molecule's function flips with its spatial source.

Core Claim

Three claims, in increasing order of ambition. First, descriptive: chronic inflammatory signaling induces a coordinated, cell-type-specific reshaping of isoform choice — including coupled transcription-start-to-splicing decisions and 3′ UTR processing — that begins before gene-expression-level or stemness changes mark the cell as transformed. Second, predictive: isoform proportions can be predicted from a continuous context vector of cytokine concentrations, cell type, and time, and a model built for continuous conditioning will beat the strongest existing context-aware system, CDBridge, which conditions only on categorical tissue labels. Third, causal: for a small set of early-switching genes, blocking the inflammation-induced isoform change with antisense oligonucleotides or a splicing-factor degrader will delay or reduce malignant transformation markers, establishing splicing as a driver rather than a passenger of inflammation-driven oncogenesis.

The existing literature motivates each piece but stops short of connecting them. The long-read colorectal cancer atlas mapped 125,205 isoforms and 394 dysregulated transcript structures in patient tumors, but only in static resections with no inflammatory or temporal resolution An isoform-resolution transcriptomic atlas of colorectal cancer from long-read single-cell sequencing. The IL-6/FRA1 work showed a cytokine flipping a transcription factor's behavior through deacetylation, promoting stem-like properties in a 123-patient cohort — but it treats FRA1, NANOG, and CD44 as monolithic genes and never asks whether splicing variants mediate the switch The inflammatory cytokine IL-6 induces FRA1 deacetylation promoting colorectal cancer stem-like properties. The KRAS4A study is the closest precedent for the mechanism we posit: microenvironmental stress (hypoxia, ER stress) shifted isoform balance of a single oncogene, and depleting the minor isoform via the RBM39/DCAF15 pathway eradicated cancer stem cells Targeting KRAS4A splicing through the RBM39/DCAF15 pathway inhibits cancer stem cells. We generalize that finding from one gene under acute stress to the genome under chronic cytokine exposure over time. On the modeling side, CDBridge predicts tissue-specific isoform activity from sequence, achieving an average $R^2$ of 0.387 and Spearman correlation of 0.618 on the GTEx-Benchmark across 40+ tissues — but its context is a discrete tissue label, with no axis for time or inflammatory state CDBridge: A Cross-omics Post-training Bridge Strategy for Context-aware Biological Modeling. And the 189-gene Stem Cell Gene Set provides a validated temporal ruler — its first principal component ordered a 14-day embryonic stem cell differentiation course day by day and stratified tumor grade — but it is built entirely from aggregate gene-level signal in bulk tissue A gene expression profile of stem cell pluripotentiality and differentiation is conserved across diverse solid and hematopoietic cancers. We reuse that ruler, at isoform resolution, as the coordinate system against which isoform switches are timed.

Research Program

The program has three parts: generate the missing data, learn the context-to-isoform mapping, and perturb the earliest switches.

The data part closes the two gaps that every cross-cutting study in this space leaves open. In vivo, we use the dextran sulfate sodium colitis model (the standard chronic-colitis paradigm used in the PGE2 work, which found a nearly 5-fold increase in colonic PGE2 during active disease and showed that autocrine versus paracrine PGE2 has opposite effects on T cells) combined with a carcinogen to produce the colitis-to-cancer trajectory, sampling colon tissue at roughly five time points from healthy through inflamed through dysplastic to carcinoma mPGES-1-Mediated Production of PGE2 and EP4 Receptor Sensing Regulate T Cell Colonic Inflammation. Each point is profiled with paired short-read and PacBio long-read single-cell sequencing, using the SQANTI3 classification scheme (full splice matches, novel in catalog, novel not in catalog) exactly as deployed in the colorectal atlas, so that coupled splicing decisions — which accounted for 299 of the 394 dysregulated structures in that study — are actually resolvable rather than fragmented An isoform-resolution transcriptomic atlas of colorectal cancer from long-read single-cell sequencing. In vitro, we expose human colonic organoids to a factorial grid of IL-6, TNF, and PGE2 at several doses across several time points, giving a dense, controlled context surface that the mouse model cannot provide. Cell types are kept separate throughout — epithelium, myeloid, lymphoid — because the myeloid NSCLC work showed a splicing switch that converts a soluble decoy receptor into a membrane-bound signaling receptor in tumor-associated myeloid cells, the kind of context inversion that bulk averaging destroys Identification of Reprogrammed Myeloid Cell Transcriptomes in NSCLC.

The modeling part learns a function from context to isoform choice. For gene $g$ with isoforms $i$, let $\pi_{g,i}$ be the proportion of that gene's transcripts that are isoform $i$. The model predicts

$$\hat{\pi}_{g,i} \;=\; \mathrm{softmax}_{i}\!\left(f_\theta\!\left(d_{g,i},\; \mathbf{c},\; t\right)\right),$$

where $d_{g,i}$ encodes what makes isoform $i$ different from its siblings (exon structure, junction motifs, UTR processing, taken from the SQANTI3 output), $\mathbf{c}$ is the continuous context vector — cytokine concentrations, cell type, and source attribution — and $t$ is time. Two design choices respond directly to documented limitations of prior models. First, conditioning is continuous and injected by feature-wise modulation, the mechanism ProtQueSt uses to make protein-function retrieval depend on the question asked; here the "question" is the inflammatory state ProtQueSt: Query-Conditioned Retrieval-Augmented Generation for Protein Function Annotation. Second, we deliberately avoid the same-gene contrastive pretraining of Orthrus, whose objective maximizes similarity between isoforms of the same gene and therefore suppresses exactly the differences we need to represent Orthrus: Towards Evolutionary and Functional RNA Foundation Models. Pretraining may still help, so we initialize from a frozen Orthrus-style encoder but train the difference-encoding head on top; whether that pretraining helps or hurts is itself an ablation.

The causal part operationalizes timing. For each gene we estimate a switch time $t^*_g$: the point on the trajectory at which its dominant isoform changes. The Stemness Index, recomputed at isoform resolution on our single-cell data, gives a second timeline. Genes whose isoform switch precedes the local stemness rise are driver candidates; those that lag are passengers — a direct attack on the correlation-versus-causality problem that the Stemness Index authors themselves flag as unsolved A gene expression profile of stem cell pluripotentiality and differentiation is conserved across diverse solid and hematopoietic cancers. Top candidates are then tested with the two perturbation tools validated in this literature: antisense oligonucleotides against splice sites, which suppressed oncogene expression in vivo in the 3′ UTR splicing study Pan-cancer pervasive upregulation of 3′ UTR splicing drives tumourigenesis, and Indisulam-induced degradation of the splicing factor RBM39, which depleted KRAS4A and eliminated cancer stem cells in xenografts Targeting KRAS4A splicing through the RBM39/DCAF15 pathway inhibits cancer stem cells. Success is measured by delayed transformation in the mouse model and reduced CD44+/CD133+ stem-like populations, the same markers used in the IL-6/FRA1 cohort The inflammatory cytokine IL-6 induces FRA1 deacetylation promoting colorectal cancer stem-like properties. CD44 is the flagship test case: its standard and variant isoforms have documented opposing roles in stemness and metastasis, and the review compiling that evidence explicitly names the cytokine triggers of the CD44s-to-CD44v switch as unresolved CD44 and its implication in neoplastic diseases.

Baselines

The naive reference point is a context-free isoform model: predict, for every gene, the isoform mixture observed in healthy tissue, regardless of cytokine, cell type, or time, with total abundance taken from a gene-level model. It anchors how much of isoform behavior is simply the baseline mixture and how much the context genuinely explains. No reported result found on Lacuna for gene-to-isoform transfer under inflammatory time courses, so we establish this number ourselves on our held-out conditions.

The strongest existing competitor is CDBridge, the cross-omics model that bridges DNA and protein foundation models with a tissue-conditioned decoder and holds the best reported numbers on the GTEx-Benchmark: an average $R^2$ of 0.387 and Spearman correlation of 0.618, against Enformer's $R^2$ of 0.127, plus a coding-region segmentation AUC of 0.993 CDBridge: A Cross-omics Post-training Bridge Strategy for Context-aware Biological Modeling. We run it two ways under the same compute and tuning budget as our model: first on the GTEx-Benchmark with its published splits, to confirm we can reproduce its reported numbers in our hands; second on our atlas, by mapping every (cytokine, dose, time, cell type) condition to the best-fitting categorical label its decoder accepts, which gives it a fair but discretized view of continuous context. Beating it means exceeding $R^2$ 0.387 / Spearman 0.618 on the GTEx-Benchmark splits, and, more decisively, outperforming it on held-out continuous conditions — unseen cytokine doses and later time points — where categorical conditioning cannot interpolate. A second strong competitor is a frozen Orthrus encoder with a regression head, motivated by its reported linear-probing Pearson correlation of 0.69 on human RNA half-life (versus 0.45 for HyenaDNA), which certifies the representation quality but, by the authors' own account of the contrastive objective, may under-represent isoform-distinguishing features Orthrus: Towards Evolutionary and Functional RNA Foundation Models. The key ablation is our own architecture with continuous context replaced by categorical labels — time bucketed into discrete stages, cytokine dose collapsed to on/off. If the categorical version matches the continuous one, the central design claim fails; this ablation is the proposal's load-bearing test.

Why not just…?

Why not just run short-read splice-event quantification on a time course? This is the cheapest serious alternative — tools like ASprofile already track percent-spliced-in events across developmental series, as done across four porcine testis stages where transcription-start and termination changes dominated (41.0% and 32.9% of 92,738 events, versus only 7.6% exon skipping) Identification and Analysis of Alternative Splicing Events in Sus Scrofa Using RNA-Seq Data. The problem is resolution: in the colorectal atlas, 299 of 394 dysregulated transcript structures arose from coupled multi-event splicing that short reads cannot connect, and start/end-site choices — precisely the dominant event classes — are what short reads fragment most badly An isoform-resolution transcriptomic atlas of colorectal cancer from long-read single-cell sequencing. We are honest that this is partly a resource question, so we run the short-read event pipeline on our own time course as an explicit ablation and count how many of the long-read-identified early switchers it recovers. If it recovers most of them, the cheap path wins for discovery and we will say so.

Why not just extend the Stemness Index to isoform level? The Stemness Index already orders time and stratifies tumor grade with a 189-gene set, so a natural cheap proposal is to recompute it on isoform counts and rank genes by how early their isoform-level score shifts A gene expression profile of stem cell pluripotentiality and differentiation is conserved across diverse solid and hematopoietic cancers. We actually adopt this as the trajectory coordinate — it is a real strength, not a straw man — but it is a correlation ruler: it cannot say which isoform within a gene moved, cannot separate epithelial from infiltrating immune cells in bulk, and cannot predict anything about an unseen context. The decisive comparison is driver-candidate ranking: we will measure the overlap between the isoform-level-SCGS ranking and our model's predicted early switchers, and check both rankings against the perturbation outcomes. If the simple index ranks causal hits as well as the model, the model's contribution collapses to interpretation, and the experiments will reveal that.

Why not just fine-tune Orthrus or CDBridge on the new data? This is the most genuine threat, because it requires no new architecture. The documented obstacles are specific: Orthrus's contrastive objective pulls isoforms of the same gene together, suppressing the signal we need, and CDBridge's conditioning is categorical, with no slot for a continuous time or dose axis Orthrus: Towards Evolutionary and Functional RNA Foundation Models; CDBridge: A Cross-omics Post-training Bridge Strategy for Context-aware Biological Modeling. But fine-tuning can partially patch both, so rather than argue, we run both fine-tuned baselines head-to-head with our model under identical data and tuning budgets, and the extrapolation conditions (unseen dose, unseen late time point) are where the architectures should separate. If fine-tuned CDBridge matches us there, the simple path wins and the paper becomes an honest negative result about architecture — which is itself worth publishing given how much current investment flows into these foundation models.

Proposed Experiments and Decisive Checks

Experiment one builds the atlas: five time points of the colitis-to-cancer mouse model plus the organoid cytokine-dose-by-time grid, each with paired short-read and long-read single-cell profiling and SQANTI3-classified isoform catalogs per cell type. The primary descriptive readout is the fraction of inflammation-associated isoform changes that are (a) coupled multi-event structures, (b) cell-type-specific, and (c) present before any gene-level or stemness-index change — the descriptive claim stands or falls on whether a substantial early isoform layer exists at all.

Experiment two benchmarks the model on three tiers of difficulty: reproduction of the GTEx-Benchmark against CDBridge's reported 0.387 $R^2$ / 0.618 Spearman CDBridge; interpolation to held-out cytokine doses and time points on the organoid grid; and extrapolation to the in vivo mouse conditions, the hardest transfer test of whether the learned context-to-isoform mapping captures biology rather than batch. We additionally probe whether predicted isoforms change predicted protein function, using RNA-GPS-style localization prediction — reported at 0.77 multi-compartment AUROC — to check that inflammation-induced 3′ UTR processing shifts transcript localization, as the pan-cancer 3USP work showed for CTNNB1 nuclear-export traps RNA-GPS predicts high-resolution RNA subcellular localization and highlights the role of splicing; Pan-cancer pervasive upregulation of 3′ UTR splicing drives tumourigenesis.

Experiment three runs the causal tests on the top-ranked early switchers, including CD44 as the positive-control candidate whose variant isoforms are already implicated in stemness and whose cytokine triggers are an open question CD44 and its implication in neoplastic diseases. Antisense oligonucleotides block candidate splice sites; Indisulam tests whether RBM39-dependent splicing, already shown to control KRAS4A under hypoxia, also mediates cytokine-driven switches Targeting KRAS4A splicing through the RBM39/DCAF15 pathway inhibits cancer stem cells. Endpoints are transformation latency and the CD44+/CD133+ population. A useful external comparison for prognostic value is the left-versus-right colon cancer splicing study, whose 10-event splicing signature reached an AUC of 0.84 for overall survival and whose central splicing factor RBM25 reached 0.672 — our early-switcher panel should be benchmarked against that bar on human validation cohorts Systematic profiling of alternative splicing events and splicing factors in left- and right-sided colon cancer.

The decisive checks, stated plainly: the descriptive claim fails if early isoform switches are rare or merely track aggregate expression; the predictive claim fails if the categorical-context ablation or fine-tuned CDBridge matches the continuous model on extrapolation; the causal claim fails if blocking early switches leaves transformation and stemness markers intact. Each check has a named experiment above, and the proposal is falsifiable at every tier.