MicroRNA-Regulon Anchors for Drift-Stable Drug Response Prediction in Cancer

A machine-learning model that matches specialists on a static test set can quietly fail years later: patient populations change, sequencing platforms are upgraded, treatment guidelines are revised, and the model's inputs drift away from its training distribution. This "performance rot" is well documented in clinical tabular data — in intensive-care models, a sepsis predictor lost up to 0.090 AUROC over calendar time, translating to one additional missed septic patient for every eleven (Guo et al.) — and rolling backtests show that random train/test splits systematically overstate deployment performance (Zhou et al.). Yet almost nothing is known about whether molecular models — genomics-based diagnostic classifiers and drug-sensitivity predictors — rot the same way, because every major molecular benchmark uses time-agnostic splits (The TCGA Meta-Dataset Clinical Benchmark).

This proposal makes a specific claim about how to slow that rot in cancer drug response prediction. MicroRNAs — short non-coding RNAs, each of which represses dozens of downstream messenger RNAs — act as multi-gene rheostats, and the biology literature shows they are the switches through which chemoresistance emerges: loss of miR-136 restores hyper-efficient DNA crosslink repair in ovarian cancer (Zhao et al.), depletion of miR-125a uncaps the repair enzyme Rad51 in cervical cancer (Huang et al.), and loss of miR-17/20 or miR-1268b stabilizes survival kinases Akt1 and ERBB2 in breast cancer (Yu et al., Zhu et al.). The only feature-level intervention that has ever preserved discriminative power under temporal shift in clinical machine learning was expert knowledge-based grouping — clinicians manually mapping drifting codes into stable biological concepts — while automated statistical reductions like PCA failed (Guo et al.). My claim is that curated microRNA-target regulons are exactly such a grouping, available off the shelf from the resistance literature, and that anchoring drug-response features to them preserves both calibration and discrimination as cohorts age — where pooled Empirical Risk Minimization, invariance penalties, and periodic recalibration all decay.

Core Claim

When a gene-expression drug-response model represents each sample not by 20,530 raw gene counts but by activity scores over biologically curated microRNA-target regulons (plus mRNA pathway scores), its performance degrades substantially more slowly under rolling temporal backtesting than the same architecture trained on raw or statistically reduced features. The mechanism is not statistical invariance — which Guo et al. showed actively harms predictive power by discarding valid time-varying features — but biological anchoring: the miRNA-gene interaction graph imposes a stable, literature-validated structure on the feature space that is invariant to sequencing-platform and protocol drift even as individual gene measurements move. Additionally, the regulon scores themselves change interpretive role: a regulon whose activity drifts across collection years becomes a candidate therapeutic target, because the biology literature already tells us that miRNA dysregulation is a cause of resistance, not merely a marker of it.

Research Program

Three components. First, build the anchor: a curated microRNA-target regulon map assembled from the mechanistic resistance literature catalogued above — miR-136/DNA-repair, miR-125a/Rad51, miR-17/20/p53-Akt1, miR-1268b/ERBB2-PI3K-Akt, miR-195-5p/REGγ-Wnt-β-catenin in renal carcinoma (Chen et al.), miR-302e/MBNL3-AKT-VEGFA in gastric adenocarcinoma (Journal of Microbiology and Biotechnology, 2024), and the oncomiR axes such as miR-21 in lung cancer (MicroRNA-21: A potential therapeutic target in lung cancer (Review)) — extended with long non-coding RNA sponges that sequester tumor-suppressive microRNAs, since these competing-endogenous-RNA loops are part of the same post-transcriptional layer. Each regulon $k$ yields one activity score per sample: the median of the z-scored expression of its target genes, with sign flipped for repressed targets,

$$R_k = \mathrm{median}\big(\{-z_g : g \in T_k\}\big),$$

where $T_k$ is the target-gene set of microRNA $k$ and $z_g$ is the standardized expression of gene $g$. The median makes the score robust to any single platform-sensitive gene drifting, which is precisely the failure mode raw features suffer. This mirrors the move that let Precily transfer from bulk to single-cell data (Chawla et al.), but at a regulatory rather than pathway layer.

Second, build the model: a drug-sensitivity predictor in the Precily mold — pathway activity scores plus drug descriptors — but with microRNA regulon scores added as input features. Precily is the direct predecessor here: it integrates 1,329 mRNA pathway scores with drug descriptors in a deep network and reaches Pearson correlation $\rho = 0.88$ on the Genomics of Drug Sensitivity in Cancer (GDSC) test set, $\rho = 0.84$ on the independent CTRPv2 panel, and $\rho = 0.85$ at single-cell resolution, while completely omitting non-coding RNA (Chawla et al.). Our model keeps everything Precily does and adds the regulon layer, so any gain is attributable to the anchor.

Third, build the evaluation: the first molecular temporal-backtest benchmark. We re-partition the TCGA meta-dataset's 174 clinical prediction tasks across 33 cancer types (The TCGA Meta-Dataset Clinical Benchmark) by sample acquisition date, applying the EMDOT rolling-origin protocol from Zhou et al.: train only on samples with date $\le t$, test on samples after $t$, roll $t$ forward across the collection horizon, and report the full decay curve rather than a single number. The same protocol is applied to GDSC/CTRPv2 drug-response prediction by cell-line creation and screening batch date, and to TCGA patient survival stratification, where Precily reported 72% five-year survival for predicted responders versus 28% for non-responders — a gap we will re-measure per time window to see whether it holds or collapses as training data ages. This is the piece nobody has done: Zhou et al. explicitly note EMDOT has never been applied to high-dimensional genomic data, and the miRNA diagnostic classifiers that do exist — NSCLC subtyping from tissue miRNA-Seq (miRNA-Seq Tissue Diagnostic Signature) and the miRNA-183/HE4/CA-125 ovarian panel (Combined measurement of miRNA-183, HE4, and CA-125) — are validated on single, temporally unstratified cohorts.

Baselines

The strongest existing competitor is the Precily architecture itself, mRNA-pathway-scores-only, exactly as reported in Chawla et al.: it achieves $\rho = 0.88$ ($R^2 = 0.77$) on the GDSC test set and $\rho = 0.84$ on CTRPv2, and it already demonstrates the pathway-smoothing idea that motivates ours. It is the right competitor because it is the best-performing published drug-sensitivity model in the file and the most similar representation. To make the comparison fair, we reimplement it with the same architecture, the same 1,329 GSVA pathway scores and SMILESVec drug descriptors, the same train/test splits, and the same tuning budget as our regulon-augmented variant; the only difference is the added microRNA-regulon input block. Beating it means a higher Pearson $\rho$ on GDSC and, more importantly, a flatter decay curve under rolling temporal backtesting — specifically, a smaller per-year AUROC or $\rho$ loss with non-overlapping confidence intervals across rolling windows — since matching it statically while decaying slower is the real claim.

The second baseline is pooled Empirical Risk Minimization on raw gene counts, which Guo et al. established as the method proactive robustness algorithms fail to beat on clinical temporal shift; it is the "do nothing sophisticated" reference point and will be run with the same compute and tuning budget, trained on all data up to each rolling cutoff. The third is simple Logistic Regression, which deserves its slot because Zhou et al. showed complex models lose their advantage to it as data ages — on the TCGA meta-dataset it already reaches 68.78% average accuracy versus 71.19% for a multilayer perceptron (The TCGA Meta-Dataset Clinical Benchmark). A naive majority classifier (66.18% average accuracy on the same benchmark) sets the floor. The decisive ablation isolates the proposal's key ingredient: identical architecture and pipeline, but with regulons replaced by random gene groupings of matched size, plus a PCA-reduced feature set — since Guo et al. reported that PCA failed to preserve discrimination under shift. If random groupings or PCA decay as slowly as curated regulons, the biological anchoring story is dead; if curated regulons alone hold, the ingredient is the biology, not the dimensionality reduction.

Why not just…?

Why not just recalibrate the model periodically? Because recalibration fixes confidence, not ranking. Guo et al. found calibration broke down in 11 of 15 clinical studies and post-hoc calibration restored it, but discrimination, once feature-outcome relationships drift, cannot be recovered by any monotone rescaling of outputs — that is exactly the asymmetry this proposal targets at the feature level. Our experiments will make this concrete: at each rolling window we apply post-hoc recalibration to every baseline and report how much of the AUROC or $\rho$ loss it recovers. If it recovers most of the loss for the regulon-anchored model, the anchor was unnecessary; we predict it recovers little for raw-feature models.

Why not just use Logistic Regression and retrain often? This is the most serious alternative, and it partly works: Zhou et al. show simple linear models decay more slowly than deep ones, and Guo et al. show pooled ERM beats every proactive robustness method. But retraining needs fresh labels, and in oncology the ground-truth labels are drug response and survival, which take years to mature — the five-year survival endpoints in Chawla et al. are the point. A model that rots silently during the labeling lag directs patients to ineffective therapy, as the sepsis result — one extra false negative per eleven patients — makes vivid. We therefore include the retrain-often schedule as an explicit baseline at every rolling window and report its performance as a function of label lag, simulated by artificially withholding the most recent $k$ years of outcome labels. Our claim survives only if the regulon anchor beats retrained raw-feature models under realistic lags; if not, we report that honestly, because that would itself be the first quantification of label-lag-induced rot in molecular models.

Why not just model the 20,530 mRNAs and ignore microRNAs? Because the mRNA layer is downstream of the switches that actually move. Each microRNA simultaneously tunes an entire survival or repair network — miR-125a repressing Rad51, miR-17/20 gating the p53-Akt1 axis, miR-1268b gating ERBB2 — so a single regulon score summarizes a coordinated multi-gene state that a thousand individual gene measurements dilute across. And the file's own evidence cuts against purely statistical reductions: automated PCA failed to preserve discrimination under shift, while expert knowledge-based grouping was the only feature-level method that succeeded (Guo et al.). Precily shows the mRNA-only ceiling ($\rho = 0.88$ on GDSC) and simultaneously shows the gap — it tracks resistance evolution in xenografts but cannot see the non-coding layer that drives it (Chawla et al.). The random-grouping ablation is the decisive test: if matched-size random groupings do as well as curated miRNA regulons, the reviewer wins.

Proposed Experiments and Decisive Checks

Experiment one applies the EMDOT rolling-origin protocol to the TCGA meta-dataset's 174 tasks, re-partitioned by acquisition date. Metrics are average accuracy and AUROC per rolling window, plus calibration error, reported as decay curves for the multilayer perceptron (71.19% static average accuracy, the number to beat or match while decaying slower), Logistic Regression (68.78%), and the majority floor (66.18%) (The TCGA Meta-Dataset Clinical Benchmark). Experiment two is the drug-response backtest on GDSC and CTRPv2 with Precily as the baseline ($\rho = 0.88$ GDSC, $\rho = 0.84$ CTRPv2, $\rho = 0.85$ single-cell transfer (Chawla et al.)); the headline number is per-year $\rho$ decay for mRNA-only versus regulon-augmented models. Experiment three re-runs the TCGA survival stratification per time window, checking whether the 72%-versus-28% five-year survival separation between predicted responders and non-responders persists or collapses as training data ages. The single-cell translation check from Precily carries over directly: regulon scores should transfer across sequencing resolutions, and if the median-of-targets score transfers better than raw genes, that is independent evidence the anchor is platform-stable.

The decisive checks, in order. First, does the regulon-anchored model decay significantly slower than every baseline under rolling backtesting — the criterion is a statistically significant difference in per-window slope, not a static-set win. Second, does the random-grouping ablation fail to match curated regulons, isolating biology as the active ingredient. Third, does post-hoc recalibration recover lost discrimination for no model, confirming that the anchor must act at the feature level. If all three hold, the deliverable is twofold: the first temporal-drift benchmark for molecular diagnostic and drug-response models, and evidence that the microRNA-regulon layer — the same layer the therapeutic-targeting literature wants to drug — is also the layer that keeps predictive models honest as cohorts age. On the therapeutic side, the backtest yields a bonus no static study can: any regulon whose activity systematically drifts across collection eras in a direction consistent with the known resistance mechanisms (miR-136 loss, miR-125a loss, miR-21 gain) surfaces as a time-stable, mechanistically motivated candidate for mimic or antagomiR development, directly connecting the drift analysis back to therapeutic target identification.