CAPO gates protein-model adaptation before training: it refused a Boltz-2 fine-tune and still ranked 11 of 15 targets better
Synopsis
Adapting a protein foundation model to one laboratory's assay is an expensive, failure-prone step, and existing agents treat fine-tuning as a reliable tool call whose feasibility is estimated and whose infeasibility surfaces only after compute is spent. CAPO reframes adaptation as a gated decision: it measures feasibility on the target accelerator, admits or rejects a configuration against a budget, runs a 200-step canary inside the real training loop, and may decline. One unchanged system chose partial fine-tuning of ESM-2 8M, refused to fine-tune Boltz-2 in favor of a calibrated head over frozen representations, and applied rank-4 low-rank adaptation to ESM-2 150M across three protein tasks.
Interpretation
Protein-model adaptation can now be refused before money is spent: CAPO turns whether and how to adapt into an admission check that must pass, rather than a training step taken by default. Earlier tool-using agents treated fine-tuning as a cheap, reliable call, estimating feasibility and cost and discovering infeasibility only after compute had been consumed. On hit-to-lead affinity ranking, the projected cost of a Boltz-2 weight update fell into the rejection branch, so the system kept Boltz-2 frozen, reused 644 cached co-folds and fitted a per-target assay-calibrated affinity head on its outputs.
The same code produced three different adaptation choices across three tasks, two of them departing from the conservative prior keyed on label count. A label-count default would impose one strategy, while measured evidence pushed the choice toward partial fine-tuning, refusal, or low-rank adaptation. Host range partially fine-tuned the last four layers of ESM-2 8M to a species-averaged MCC of 0.826; hit-to-lead refused the fine-tune; variant effect applied rank-4 low-rank adaptation to ESM-2 150M, reaching an MCC of 0.478 on 420 ClinVar variants uncertain in 2020 and since reclassified.
On host range, CAPO tied a general coding agent on prediction accuracy, and the difference appeared in process cost and in data curation. Given identical raw paired-end reads, task text and a USD 50 budget, the general coding agent also trained a model, but failed at the read curation before modelling. CAPO reached a species-averaged MCC of 0.826 against 0.812 for the general coding agent, for which no confidence intervals are available; CAPO used 57.4% less GPU time, 32.5% less cost per successful run and zero failed training trials against four; it agreed with a blinded golden set on 97.6% of 14.4M pangolin sequences and 99.83% of 2.61M possum sequences, where the general agent scored 0.00% on both.
Perspective
The demonstrated scope is three protein-modeling tasks, namely SARS-CoV-2 host-range prediction, hit-to-lead affinity ranking and temporally split variant-effect reclassification, each run end to end by the same unmodified system from a single natural-language request under a USD 50 budget. The intended readers are researchers and engineers who must adapt protein models to their own assay data under a compute budget. There is no closed loop and no wet-lab validation.
Each task is a single end-to-end run, the agentic process itself has no seeds or variance estimates, and the gate is evaluated only as a whole, so zero versus four failed trials is a natural experiment rather than an ablation. The head-to-head against the general agent covers only the host-range task. Each study carries its own bounds: the pre-registered FEP+ hardest-case contrast was empty, the physics anchor mapped no ligands, the variant panel is small and pathogenic-skewed, calibration error is 0.190, and the region-novel interval contains zero. Cost projection overestimates realised GPU cost by 2.1 to 3.2 times yet is optimistic about whole-run cost, projecting USD 4.26 against USD 29.76 realised on the variant-effect study.