CAPO: Compute-Aware Automated Protein-Model Optimization

Protein models like Evolutionary Scale Modeling (ESM), AlphaFold, and Boltz have revolutionized computational protein science. However, these pretrained models often need adaptation to specific laboratory assays, which is technically challenging and consumes significant computational resources. Existing automated systems can train these models, but they often estimate feasibility and cost, discovering infeasibility only after spending computational resources.

Our study presents CAPO (Compute-Aware Automated Protein-Model Optimization), a multi-agent system that decides whether and how to adapt a model before investing expensive computation.

CAPO employs a single rule: gate before you train. It evaluates feasibility on the target accelerator, estimates cost against the available budget, runs a canary within the training loop, and maintains an auditable artifact trail. Crucially, CAPO can refuse an adaptation if it is unaffordable or opt for a cheaper alternative. We tested CAPO on three protein-modeling tasks: SARS-CoV-2 host-range prediction, hit-to-lead affinity ranking, and temporal variant-effect reclassification.

For SARS-CoV-2 host-range prediction, CAPO fine-tuned an 8-million-parameter ESM-2 model to a species-averaged MCC of 0.826, surpassing a hand-designed expert transformer's 0.743. It achieved this using 57.4% less GPU time and eliminated failed training trials entirely. In hit-to-lead affinity ranking, CAPO opted to train a lightweight head over frozen representations instead of fine-tuning Boltz-2.

This approach improved ranking on 11 out of 15 held-out targets in just 1.68 GPU-hours. Lastly, for temporally held-out ClinVar variants, CAPO-based rank-4 low-rank adaptation of ESM-2 achieved an MCC of 0.478, a significant improvement over the 0.252 MCC of a one-hot baseline and the 0.000 MCC of the frozen zero-shot score.

These results, obtained from single runs without a gate ablation and with one generalist-agent comparison, demonstrate that measured admission control can make protein-model adaptation not only automated but selective. CAPO decides what to compute, what not to compute, and why, making the process more efficient and cost-effective.