Computational Epistemics of Biological Discovery Part I: The Real Cost of AI Isn’t Generation—It’s Verification

Abstract: Generative artificial intelligence has suppressed the resource cost of producing biological hypotheses by orders of magnitude—from de novo protein sequences to entire metabolic circuits—while wet-lab validation remains bound by strict physical, financial, and temporal constraints. This asymmetry has become the central bottleneck of modern biological discovery. We introduce Computational Epistemics, a formal mathematical framework that quantifies the minimum resources required to establish scientific truth about a biological hypothesis. Adapting the architecture of Probably Approximately Correct (PAC) learning and computational complexity theory to the physical setting, we define Verification Complexity (K), a multi-dimensional resource metric over time, capital, physical samples, and compute. We establish three epistemic complexity classes—EC-P, EC-H, and EC-I—and prove that EC-H contains classes not in EC-P via an unconditional information-theoretic argument grounded in channel capacity. Concretely, we show that single-residue deep mutational scanning lies in EC-P, while verifying the complete higher-order epistatic landscape lies in EC-H. We establish a Cram´er–Rao lower bound on K, prove Technology Monotonicity and Epistemic Frontier Expansion, and formalize the Verification Asymmetry Conjecture: that generation cost becomes asymptotically negligible relative to verification cost for hypothesis classes in EC-H.Significance Statement: AI can now design millions of new protein sequences in hours, but testing each one in the lab remains slow and resource-intensive. This asymmetry—between cheap generation and expensive physical validation—is the central bottleneck of modern biological discovery. We introduce a mathematical framework, Computational Epistemics, that quantifies this bottleneck and proves that certain biological questions require exponentially more lab resources to verify as protein length grows, regardless of future improvements in measurement technology, provided each physical measurement carries finite information. This formalizes why hypothesis generation alone cannot drive scientific progress to completion.