Fitting dynamics is not identifying causal edges: a white-box masked ODE benchmark for trans-omics digital twins of drug action mechanisms
Multi-component drug regimens, with traditional Chinese medicine formulas as the hardest case, act across signalling, transcriptional, proteomic and metabolic layers, and elucidating their mechanisms requires dynamic models that predict molecular trajectories rather than static association networks. Trainable ordinary differential equation (ODE) systems fitted to time-series omics are…
The white-box masked ODE benchmark for trans-omics digital twins of drug action mechanisms reveals that fitting dynamics alone does not identify causal edges. This benchmark, centered on the insulin action in mouse liver, comprises a ground truth of 2,106 molecular species and 4,912 edges. Four instruments were used to assess identifiability: an oracle-perturbation basin curve, a held-out-layer corruption assay, a saturation audit, and an ideal-budget ceiling test.
A static baseline (FD + LASSO) performed at chance, and even from-scratch training without noise, fully observed data, and dense sampling yielded similar results with per-layer AUROC scores ranging from 0.48 to 0.53. Held-out layers functioned as corruption sinks, failing due to three factors: an information floor, a scale-mismatch amplifier, and an edge-gradient drag, which could only be mitigated jointly.
Tanh saturation was observed to silence entire regulator columns. A 12-knockout validation battery was employed to dissect intervention reliability based on network distance. This benchmark, along with its code and audit tools, is slated for open release post-publication. The critical factor for structural identifiability is not fitting but rather structural identifiability.
Written by urgent.news from bioRxiv's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
