There is a familiar shape to how a field opens up. For a long time the barrier is not the idea; it is the instrument. Anyone can imagine a molecule. Almost nobody could find out whether it was worth anything, because judging it required a laboratory, a compute budget, and years of training. The idea was cheap and the verdict was expensive.
Something worth watching is happening to that asymmetry.
An open drug-discovery challenge is currently running on Hugging Face, and the interesting part is not the prize money. It is that the evaluation has been made public.
Two seasons are live
Season 1 — Malaria. Target: PfDHODH in Plasmodium falciparum, with human DHODH as a counter-target. Closes 30 September 2026.
Season 2 — Tuberculosis. Target: InhA in Mycobacterium tuberculosis, counter-target human FASN (ER domain). Closes 31 October 2026.
Entry is a molecule, submitted as SMILES or InChI. There is no restriction on how you get there: Claude, GPT, Gemini, Qwen, KIMI, DeepSeek, or a model you trained yourself. The challenge scores the structure, not the pedigree of whoever produced it.
The scoring is where it gets interesting
100 points across six axes — activity, binding, selectivity, ADMET, novelty, synthesizability — with weights that differ per season. Season 2 loads weight onto binding and selectivity, which is the right instinct: an InhA inhibitor that also shuts down the human homolog is not a candidate, it is a toxin.
That counter-target design is the detail most worth noticing. It is easy to build a scoreboard that rewards binds the target. Rewarding binds the target and leaves the human protein alone is harder, and much closer to what actually decides whether a compound survives.
The anchors are in the table
Approved drugs and inert compounds are scored on the same rubric as the entries.
That is an unusual thing for an organizer to do, because it means anyone can see where the scale actually sits — including when a known drug does not score the way you would expect. A leaderboard you can audit is a different object from a leaderboard you are asked to believe.
Private submission is supported, so entrants who want to keep a structure unpublished can still compete; the board shows a masked SMILES. The standings are also exposed as JSON, which means the whole thing can be pulled and checked programmatically rather than read off a webpage.
The honest caveat, which the organizers state themselves
A high score is a research hypothesis, not a drug. Everything here is computational. It is not clinical validation and does not become one by ranking well.
Still — the direction is the encouraging part. Instruments spread, and fields change when they do. Whether this particular challenge produces anything durable is unknown. That it exists in a form outsiders can inspect is already worth something.
Challenge: https://huggingface.co/spaces/FINAL-Bench/open-discovery-challenge
Disclosure: the challenge is operated by the FINAL-Bench organization on Hugging Face.
Top comments (0)