The full 300-drug screen hammered the public ColabFold MSA server (one
redundant query per drug for the same HDAC2 sequence) -> timeouts, and
the default return_exceptions=False risked aborting the whole run on a
single failure.
Corrected:
- cache_msa(): compute the target MSA ONCE via the server, cache the a3m
on the Volume; doubles as the weight/CCD warmup.
- build_boltz_yaml(msa_path): protein reuses the cached a3m.
- cofold(msa_path): when given, skip --use_msa_server (no server query).
- screen(): cache MSA once, then cofold.starmap(..., return_exceptions=True)
so a bad drug is skipped, not fatal.
Turns a fragile ~2.5 hr run into a fast, robust one. Validation pilot
(6 drugs) running; full 300 to run tomorrow.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Junior B.
066e0096d7