Molecularizing Modern AI Models

Molecular computation offers chemical parallelism and direct operation in biochemical environments, but modern neural networks are not written for resource-limited molecular substrates. We study molecularization: a learning-and-compilation transformation of pretrained neural modules into resource-aware molecular programs. A molecularized student adapts to non-ideal primitives; molecular architecture search (MAS) proposes structures under compiler-level budgets; an independent validation split selects between MAS and matched saliency candidates; hard commitment is followed by recovery; and the committed program is lowered toward autonomous CRN/DSD/DNA execution. Across Qwen3-0.6B, 1.7B, and 4B, one frozen protocol over 27 model–budget operating points yields zero observed selector regret relative to a post-hoc evaluation oracle, with two switches away from MAS at extreme 1.7B budgets. Frozen WikiText choices also match the post-hoc candidate oracle at all six evaluation-only LAMBADA probes. Matched SmolLM2 ablations separate substrate-adaptation and post-commit-recovery gains, while Mamba transfers the adapt–commit–recover principle beyond Transformers. The compiler further lowers phased blocks to one-pot DSD/DNA control, where sequence-resolved mass-action simulations quantify autonomous clock cost. Overall, molecularization jointly optimizes task fidelity, molecular resources, structure, and autonomy rather than treating chemistry as post-processing.