Cross-Model Uncertainty-Aware Minimal Editing of Cis-Regulatory Elements for Cell-Type-Selective Design
TL;DR
By separating search from cross-model audit and enforcing explicit edit budgets and sequence–domain safeguards, the framework provides a reproducible and auditable approach to computational CRE prioritization, regulatory-grammar studies, and synthetic regulatory-element design.
Abstract
Background/Objectives: Cis-regulatory elements (CREs) help mediate cell-type-selective gene expression, yet sequence-to-activity models are rarely converted into compact, experimentally tractable variant panels that retain natural regulatory scaffolds. We developed SafeEdit-CRE to select minimal edits that preserve cross-model specificity while reducing predictive uncertainty, yielding an auditable CRE variant library for reporter-screening prioritization. Methods: SafeEdit-CRE combines the published Malinois sequence–activity predictor with an architecture-distinct reviewer ensemble, validation-calibrated uncertainty, sequence-domain controls, and constrained beam search. Natural 200 nt CREs were edited under fixed substitution budgets of 1, 5, 10, or 20 nucleotides and evaluated by a separately trained multi-kernel ensemble kept entirely separate from candidate generation. Results: Across 600 held-out CRE parents in K562, HepG2, and SK-N-SH cells, SafeEdit-CRE achieved a mean cross-model specificity-margin gain of 0.877 versus 0.822 for greedy editing (paired difference 0.055; 95% CI 0.040–0.071). Relative to a compute-matched primary-model beam, SafeEdit-CRE yielded a similar cross-model specificity-margin gain (difference −0.004; 95% CI −0.011–0.004) with lower ensemble uncertainty (0.314 vs. 0.335). In a locked 90-parent benchmark, 60.6% of designs passed all nine prespecified sequence and model-agreement checks, compared with 38.8% for greedy and 13.1% for random substitution (paired difference 21.8 percentage points; 95% CI 18.3–25.3). Conclusions: SafeEdit-CRE reduced 3240 designs to 74 Tier A computational candidates prioritized for reporter screening across three cell types and four edit budgets. By separating search from cross-model audit and enforcing explicit edit budgets and sequence–domain safeguards, the framework provides a reproducible and auditable approach to computational CRE prioritization, regulatory-grammar studies, and synthetic regulatory-element design.
