Direction Instability LINCS L1000 Drug Mechanisms Cross-Design Evidence
Drug mechanisms that work in one cell line or trial often fail in another. We develop geometric methods to predict which mechanisms will transport and cross-design classification to diagnose why drugs fail.
Direction Instability Predicts Cross-Cell-Line Drug Mechanism Transport
Direction Instability Predicts Cross-Cell-Line Drug Mechanism Transport in LINCS L1000
Zenodo 10.5281/zenodo.21213699
Direction instability (DI) quantifies the directional consistency of a perturbation’s effects across cellular contexts (DI = 1 − mean pairwise cosine similarity). DI is immune to the first-order magnitude confound that dominates Grassmannian geodesic distance, profile variability, and optimal transport cost. Applied to 8,949 LINCS L1000 drugs, DI predicts cross-cell-line mechanism transport with pre-registered, construction-neutral AUROC = 0.945, and is independently validated through Jaccard overlap of perturbed gene sets (AUROC = 0.957), breadth of target gene expression across tissues (ρ = −0.17, p < 0.001), and PRISM cell-viability correlation (ρ = 0.261, p < 0.001).
A Pre-Registered Direction Instability Atlas
A Pre-Registered Direction Instability Atlas Across Three Perturbation Modalities
Zenodo 10.5281/zenodo.21223667
Extends direction instability across three measurement technologies — L1000 transcriptomics, Tahoe-100M single-cell expression, and JUMP Cell Painting morphology — covering ~55,000 drugs and genetic perturbations. Eleven pre-registered hypothesis tests (frozen at commit SHAs 0d07a01, a17c125, 4c34699) reveal that DI is structured by pharmacological mechanism of action (η² = 0.20), concordant across modalities (partial ρ = 0.19), and suppressed by functional constraint (essentiality sign-flip, partial ρ = −0.332). Eight of eleven tests fail their pre-registered criteria, delimiting what the metric captures.
Two Kinds of Drug Failure
Drug Failures Partition into Two Kinds: Cross-Design Evidence Diagnosis Across Ten Disease Domains
Zenodo 10.5281/zenodo.21227354
MR evidence predicts drug success, but when a drug fails despite genetic support — or succeeds without it — the MR screen alone cannot explain why. A deterministic cross-design classification rule comparing observational, MR, and RCT evidence across 41 mechanism families in ten disease domains correctly classifies 24/32 evaluable families (p = 0.004). The eight misclassified families split by MR effect size into two structurally distinct groups: five with null MR (effector-neutralization and mechanism-bypass) and three with causal MR (translation gaps and exposure mismatches). Each group implies a different pipeline response. Pre-registered accuracy: 18/22 (81.8%, CI 61.5–92.7%, p < 0.001). Extension accuracy: 6/10 (60.0%, not significant in isolation). The observational leg carries zero independent predictive information — a structural consequence of the fact that mechanisms reaching Phase III have non-trivial observational support by construction.