Internal Transcript Controls: Reading Drug Effects out of Single-Arm Trials via the miRNA–m$^6$A Axis

In a single-arm trial, every patient receives the experimental therapy, so the study observes what happened under treatment but never what would have happened without it. Medicine works around this with external comparisons, matched historical cohorts, or simply cautious language: single-arm trials in aggressive cancers are labeled hypothesis-generating because confounders like natural disease fluctuation and concurrent lifestyle changes can mimic a drug effect (Patritumab deruxtecan in HR+HER2− advanced breast cancer: a phase 2 trial, Effects of a Dipeptidyl Peptidase 4 Inhibitor Sitagliptin on Glycemic Control and Lipoprotein Metabolism in Patients with Type 2 Diabetes Mellitus (GLORIA Trial)). This proposal argues that for one class of therapies — drugs that act on the epitranscriptomic machinery governing epithelial–mesenchymal transition (EMT), the process by which stationary carcinoma cells dissolve their junctions and become invasive — the control group can be built inside the treated patient rather than found outside.

The mechanism is concrete. The enzyme METTL3 decorates only certain messenger RNAs with $m^6$A — a small chemical mark that changes how long a transcript survives and how much protein it yields. Su et al. showed that in pancreatic cancer, METTL3 places $m^6$A near the stop codon of the AREG transcript, stabilizing it and driving EMT and invasion; the microRNA miR-33a-3p normally silences METTL3, and its loss in tumors unleashes this whole cascade (miR-33a-3p regulates METTL3-mediated AREG stability and alters EMT to inhibit pancreatic cancer invasion and metastasis). The key observation for trial design: a METTL3 inhibitor should change the stability of marked transcripts and leave unmarked transcripts in the same cell untouched. Those unmarked transcripts, matched to marked ones on baseline properties, are a built-in control group — a within-patient placebo arm defined by molecular structure rather than by randomization.

Core Claim

A single-arm trial of a therapy targeting the microRNA–$m^6$A axis can support valid causal estimation of the therapy's effect on the EMT transcriptome without any external control cohort, by comparing transcripts that carry the drug's target mark to matched transcripts that do not, within the same pre- and post-treatment samples. The regulatory axis itself supplies the stratification that randomization would otherwise provide. We will build this estimator, validate it against wet-lab ground truth where rescue experiments establish the true causal direction, and test whether it survives the two threats a skeptic should raise: that marked and unmarked transcripts differ systematically for reasons unrelated to the drug, and that peak-calling in $m^6$A data is biased by transcript abundance.

This differs from all existing work at the intersection. Multi-Task Adversarial Learning for Treatment Effect Estimation in Basket Trials estimates treatment effects in single-arm settings, but it needs multiple cancer indications sharing a DNA mutation to "borrow strength" across arms, and it models only static baseline mutations. Estimating treatment effects from single-arm trials via latent-variable modeling constructs synthetic controls from external electronic health records, requiring sufficient biological overlap between the trial and real-world cohorts — overlap that often does not exist in rare, molecularly defined malignancies. Causal disentanglement for single-cell representations and controllable counterfactual generation generates counterfactual cellular states, but demands a fully prespecified acyclic causal graph, while the axis we study contains genuine feedback loops, such as the self-reinforcing circuit between the transcription factor SP1 and miR-92b in head and neck cancer (Specificity protein 1/microRNA-92b forms a feedback loop promoting the migration and invasion of head and neck squamous cell carcinoma). Our design needs neither external cohorts, cross-indication pooling, nor a complete causal graph: it needs only local, directional knowledge of which transcripts the writer enzyme marks — knowledge that is directly measurable.

Research Program

The first deliverable is a matched-difference estimator for transcript-level drug effects, which we call an internal-transcript-control (ITC) design. For each patient in a single-arm METTL3-inhibitor or miR-33a-3p-mimic trial, measured transcript stability or abundance changes from baseline are split into a treated stratum (transcripts with $m^6$A peaks near the regions where METTL3 deposits marks) and a control stratum (transcripts without such peaks). Each marked transcript is matched to one or more unmarked transcripts on pre-treatment expression, transcript length, GC content, and baseline abundance trajectory, so that the comparison is not confounded by intrinsic differences between the two strata. The average difference in post-treatment change between matched pairs is the estimated causal effect of removing the mark. This is a within-sample analogue of difference-in-differences, and its identifying assumption is explicit and testable: absent treatment, matched marked and unmarked transcripts would have followed parallel trajectories. Pre-treatment longitudinal samples, where available, let us check that parallelism directly.

The second deliverable extends this to patient-level EMT outcomes. Transcript-level effect estimates feed a latent-variable synthetic-control module in the spirit of the identifiable deep latent-variable model, which separates source-dependent measurement noise from shared disease biology and handles informative missingness in real-world data (Estimating treatment effects from single-arm trials via latent-variable modeling). Unlike that work, which requires external cohorts with overlapping biology, our module can run with molecular profiles alone: patients whose marked-transcript strata fail to show the expected drug effect serve as internal evidence of non-response, even with no control arm.

The third deliverable is a semi-synthetic benchmark generator. Using the tissue-specific $m^6$A site annotations, somatic mutations, and deep site predictor from m6A-TSHub: Unveiling the Context-Specific m6A Methylation and m6A-Affecting Mutations in 23 Human Tissues, we will simulate single-arm trials in which the true effect on each transcript is known, with ground-truth EMT outcomes induced through the documented axis (METTL3 → $m^6$A-marked AREG → EGFR hyperactivation → EMT markers). This gives us something the field lacks: a benchmark where causal truth for a miRNA–m$^6$A intervention is known at transcript and patient resolution.

Proposed Experiments

Three experiments, in order of decisiveness. First, ground-truth validation on the published axis. Su et al. performed the crucial causal experiment for us: reintroducing METTL3 completely reversed the anti-metastatic phenotype induced by miR-33a-3p mimics in PANC-1, MIA PaCa-2, and BxPC-3 cells (miR-33a-3p regulates METTL3-mediated AREG stability and alters EMT to inhibit pancreatic cancer invasion and metastasis). We will apply the ITC estimator to the mimic-treated condition, holding out the METTL3-rescue condition, and ask whether it correctly predicts which transcripts' stability reverses under rescue. Predicting the rescue outcome from mimic-arm data alone is a falsifiable test: if the internal controls are contaminated by intrinsic marked-versus-unmarked differences, the reversal predictions will fail. Second, benchmark evaluation on the semi-synthetic trials, reporting PEHE and ATE error, and on the News-16 multi-treatment benchmark used by the MTAL authors, so the estimator machinery can be compared to published causal machinery under identical conditions. Third, robustness sweeps: how effect estimates degrade as $m^6$A annotation quality drops from measured peaks (100–200 nucleotide MeRIP resolution) to predicted sites — the m6A-TSFinder predictor reports an average AUROC of 0.8593 across 23 tissues, which quantifies exactly this degradation (m6A-TSHub) — and as expression-based peak-calling bias is induced artificially, since low-abundance transcripts escape peak detection for technical rather than biological reasons. This last sweep matters because m6A-TSHub documents that 93.25% of cancer mutations affecting $m^6$A sites are unique to a single tissue: marked/unmarked status is context-dependent, so mis-specified tissue context is a realistic failure mode, not a corner case.

Baselines

The do-nothing reference point is the naive pre-post comparison: average stability change of marked transcripts from baseline to post-treatment, with no matching and no unmarked-transcript referent. This is what a standard single-arm pharmacodynamic readout effectively reports, and it is the estimand that natural EMT drift, batch effects, and intrinsic transcript-turnover differences all contaminate; in our semi-synthetic benchmark it will carry a known injected bias, and beating it is necessary but not interesting. The strongest existing competitor is the identifiable deep latent-variable model for single-arm trials, which reports a median survival error of 0.135 versus 0.173 for CEVAE on its synthetic survival benchmark and the lowest absolute error for the treated-group effect on its 833-patient chronic-kidney-disease RCT paired with Helsinki EHR data (Estimating treatment effects from single-arm trials via latent-variable modeling). We will run it on our semi-synthetic trials with the molecular covariates it was never designed for, giving it the same samples, features, and tuning budget as our estimator, both with and without an external cohort supplied (when an external cohort is supplied, it is at its best; when we withhold one — the realistic rare-cancer setting — we expect its common-support assumption to fail, and we will report that honestly rather than engineer the failure). It is the strongest competitor because it is the only published method that both targets the single-arm setting and handles informative missingness, which our benchmark also injects. The second published competitor is MTAL, which achieved an ATE error of 5.12 on News-16 versus 5.76 for the next best baseline (Multi-Task Adversarial Learning for Treatment Effect Estimation in Basket Trials); we will run it on the News-16 benchmark for comparability and on our trials configured as epitranscriptomic "baskets" (stratified by miR-33a/METTL3 status across cancer indications, using the 499,369 cancer-condition peaks and 587,983 somatic mutations catalogued in m6A-TSHub) with identical data and compute. Beating both means lower PEHE and ATE error on the semi-synthetic benchmark while remaining valid when the external cohort is removed — a condition under which neither baseline was designed to work. Finally, the key-ingredient ablation replaces matched unmarked transcripts with unmatched ones: same estimator, same data, no matching. If matching contributes nothing, the internal-control idea reduces to a two-sample comparison and the proposal's core claim is wrong; this ablation is therefore decisive, and we expect effect-estimate bias to grow substantially without it.

Why not just…?

Why not just use untreated cell lines as controls? For the mechanistic questions, this is exactly right and it is what Su et al. did, comparing PANC-1, MIA PaCa-2, and BxPC-3 against HPNE normal ductal epithelium (miR-33a-3p regulates METTL3-mediated AREG stability and alters EMT to inhibit pancreatic cancer invasion and metastasis). But the same paper states that in vivo validation is missing, and the translational endpoint is a human single-arm trial in a disease with a 5-year survival below 7%, where withholding an investigational therapy to create a control arm is ethically and practically out of reach. Cell-line controls do not travel to that setting; the internal-transcript design does, because it only requires each patient's own pre- and post-treatment biopsies.

Why not just add $m^6$A features to MTAL? This is a real threat and we test it rather than argue around it: the "MTAL with epitranscriptomic covariates" run in the baselines above is precisely this alternative. Two structural reasons suggest it will fall short, though. MTAL pools across multiple indications sharing a mutation to obtain balance through an adversarial discriminator; a single-arm epitranscriptomic trial in one indication gives it nothing to pool over. And MTAL treats baseline features as static, whereas the drug's target — the $m^6$A mark — changes under treatment, which is exactly the signal our design exploits. If the MTAL variant matches us on the semi-synthetic benchmark anyway, the simpler method wins and we will say so in print.

Why not just hand-draw the causal graph and use CausCell? Because the graph is not acyclic where it matters. CausCell requires a fully prespecified causal directed acyclic graph, and the axis under study contains reciprocal circuits, most clearly the SP1/miR-92b self-reinforcing loop documented in head and neck cancer (Specificity protein 1/microRNA-92b forms a feedback loop promoting the migration and invasion of head and neck squamous cell carcinoma); the microRNA and $m^6$A arms of the axis also regulate each other's machinery. Our design deliberately uses only one directed local fact — writer-enzyme site occupancy — rather than a full graph, so feedback elsewhere in the network does not invalidate the estimate. We will still run CausCell with the best currently documented subgraph as an ablation; if its diffusion-based counterfactuals recover the rescue outcomes better than ITC does, that would point to combining the two rather than to either alone.

Decisive Checks

The proposal lives or dies on three outcomes: whether the ITC estimator predicts the METTL3-rescue reversal from mimic-arm data alone (it should, and a failure would falsify the identifying assumption); whether matching unmarked transcripts is load-bearing (the ablation must show bias growth without it); and whether the estimate survives the annotation-degradation sweep from measured peaks to predicted sites. If all three hold, the practical payoff is a template for reading causal efficacy signals out of the single-arm trials that epitranscriptomic therapies in lethal cancers will, realistically, always have.