ACTIVEGENE: REWARD-FREE, HOMEOSTASIS- ALIGNED CONTROL FOR CLOSED-LOOP GENE REGULATION VIA ACTIVE INFERENCE

mujtaba hasan

Abstract: Reinforcement learning (RL) is increasingly used to frame closed-loop genomics as sequential decision-making, but its reliance on scalar rewards makes biological control brittle: minor specification errors can induce reward-hacking–like solutions and require extensive, context-specific reward shaping (Amodei et al., 2016; Weng, 2024). We introduce ActiveGene, a conceptual framework and benchmark specification for reward-free gene regulation that replaces engineered utilities with prior preferences over future assay outcomes/states - a distributional definition of “healthy” aligned with biological homeostasis. ActiveGene selects intervention policies by minimizing Expected Free Energy (EFE), which trades off reaching preferred outcomes (risk/pragmatic value) with resolving uncertainty (epistemic value) under partial observability, avoiding ad-hoc exploration bonuses and handtuned penalty terms. To make the proposal operational without wet-lab access, we propose ActiveGeneBench: a POMDP-style virtual-cell environment separating latent cellular state from noisy single-cell observations and supporting sequential perturbations (e.g., CRISPRi/a/KO, dosing). We outline method-agnostic evaluation metrics - target attainment, safety-violation probability, intervention cost, and sample efficiency—and argue that planning under interventions is a missing axis in current static perturbation-prediction evaluations (Wu et al., 2024).

Submission Number: 102