End-to-End Differentiable Design of Cardiac-Safe Therapeutics via Physics-Informed Neuro-Symbolic Learning

Current computational cardiac safety pipelines predict ion channel affinity and simulate electrophysiological risk in isolation, with no gradient signal connecting physiological outcomes to molecular structure. We present Nabadat-Q1, the first end-to-end differentiable platform that propagates arrhythmia risk gradients from a population of virtual hearts back through a biophysical cardiac simulator to the molecular encoder. This enables gradient-based inverse molecular design for cardiac safety: starting from 45 known HIGH-risk drugs, physics-informed optimization successfully de-risked 41 (91%) by following the gradient through a JAX-based O'Hara–Rudy simulator. The dominant modification identified autonomously—reducing hERG potency while introducing protective ICaL co-blockade—recapitulates the pharmacological principle underlying verapamil's safety. Five neuro-symbolic axioms implemented via Logic Tensor Networks enforce biophysical consistency, achieving >85% satisfaction across all constraints. On 115 FDA-approved drugs, the system reaches 71.3% three-class TdP accuracy (90.9% safety rate) with transparent, per-prediction explanations. Nabadat-Q1 transforms cardiac safety from passive screening into physics-guided generative design.