A shape-constrained regression and wild bootstrap framework for reproducible drug synergy testing

Abstract

High-throughput drug combination screens require methods to identify synergistic pairs, yet widely used synergy scores lack statistical inference and can fail when parametric dose-response fits do not converge. We present SIR (Synergy via Isotonic Regression), a nonparametric framework that defines interaction as deviation from a monotone-additive null, fit by 2D isotonic regression. A degrees-of-freedom-corrected wild bootstrap yields calibrated p-values for each dose-response matrix. On DrugCombDB, SIR interaction surfaces achieve higher replicate concordance (median correlation 0.91 across 1,839 replicate pairs) than all baselines (0.53-0.74), while avoiding Loewes 20.9% and ZIPs 3.6% failure rates. The fitted surface also predicts missing wells (median holdout RMSE 0.040). By replacing heuristic scores with calibrated effect sizes and p-values, SIR enables principled hit calling and error-rate control in large screens.