ABLE: Choosing Perturbation Experiments to Recover Gene Logic

Scientific knowledge requires claims stated formally, checked against evidence, and paired with what remains undecided. Perturb-seq and CRISPR screens promise genome-scale interventional data, yet current machine-learning tools return ranked edge lists rather than executable regulatory logic. We address this for Boolean gene regulation under an idealized in silico Boolean intervention oracle, a setting where three capabilities are each necessary and none suffices alone. A neural proposer amortizes candidate-rule search at biological scale, replacing combinatorial enumeration with a sub-second forward pass. A symbolic verifier, LiftCert, issues support-conditional uniqueness certificates on the declared regulator support and abstains explicitly when the data do not determine a rule. A coverage-guided active loop then closes the remaining gaps by naming the exact missing input combination. ABLE (Active Boolean Learning Engine) realizes these capabilities as a single propose-verify-query loop. Trained once on synthetic data, ABLE delivers support-conditional certified recovery across published biological Boolean models and certifies every rule in curated biological networks from modest warm starts. ABLE thus instantiates neuro-symbolic methodology for AI4Science, showing that the verifier and active loop, not the neural architectural prior, earn certification and yield a recoverability map where every rule is a checkable claim or a named missing observation.