700Scientific computing
Screening 1,761 approved medicines against four drug-resistant bacteria, with every result tied to its evidence.
(57) Abstract
Repurposing asks whether a medicine already approved for something else could act against a drug-resistant superbug. This system ingests real bioactivity and regulatory data, ranks candidates with models, checks molecular fit by docking, and shows what clinical history already exists for each one.
Antimicrobial resistance outpaces new antibiotic discovery. An approved medicine already has known pharmacology and human safety data, so a credible lead from the existing library shortens the road to the lab. The engineering problem is to run the screen at full scale and to never let a prediction read like a proof.
-- src/batchdock/queue.py, claim() (trimmed)
update docking.jobs j set status = 'RUNNING', worker_id = %(w)s,
attempt_count = j.attempt_count + 1, heartbeat_at = now(),
lease_expires_at = now() + make_interval(secs => %(lease)s)
where j.id in (
select q.id from docking.jobs q
join docking.runs r on r.run_id = q.run_id and r.status = 'RUNNING'
where q.status = 'QUEUED'
and (q.not_before is null or q.not_before <= now())
order by q.priority, q.id
limit %(n)s
for update of q skip locked)Every (medicine, target, configuration) is one row in a Postgres queue. Workers claim with FOR UPDATE SKIP LOCKED, hold a lease that a heartbeat extends, and complete in one statement that only succeeds if they still hold it. A result can't be written twice, and a vanished worker's job returns to the queue.
1,761 ligand files and 4 receptor files are prepared once and stored with their SHA-256. A worker on any machine fetches what it lacks and refuses a file whose checksum differs, so every docking run starts from identical inputs.
Scaffold-aware splits keep similar molecules out of both training and test. An audit found enantiomer pairs straddling the split because scaffolds carried stereochemistry; scaffolds are now stereo-free, and leakage across splits is a critical check that fails the pipeline.
A probability is shown only for the four modelled pathogens, and labels that imply clinical benefit raise an error in code. The interface keeps 'no evidence found', 'not yet checked' and 'no effect' apart, and a registered trial is never presented as a successful one.
A screening system in which every docking score is a recorded AutoDock Vina run with its inputs, configuration and command.
The system of claim 1, wherein a Postgres queue with leases guarantees each job is completed at most once.
The system of claim 1, wherein no prediction is presented as evidence that a medicine treats an infection.