3,462 Molecules in Five Days: What an Open Drug-Discovery Leaderboard Actually Measured

Five days ago we opened a drug-discovery competition to anyone who wanted a turn. The Open Discovery Challenge now sits at #15 on Hugging Face's trending Spaces, ahead of most of what a leaderboard normally competes with for attention.

That is nice. It is also not the interesting part.

The numbers, as of today

	Season 1 — Malaria	Season 2 — Tuberculosis
Submissions scored	2,294	1,168
Distinct participants	78	28
Top score	84.6	79.1
Median score	35.3	38.3

3,462 scored molecules from 83 people, against Plasmodium falciparum PfDHODH and Mycobacterium tuberculosis InhA. Both seasons are still open — malaria closes 30 September, tuberculosis 31 October.

Every entry is scored automatically on six axes — cellular activity, target binding, selectivity, ADMET, novelty, synthetic feasibility — and lands on a leaderboard anyone can read.

The finding that surprised us

The obvious question people ask is which AI is best at this. We can see the answer forming in the data, and it is not the answer the question expects.

Look at what a single model does across many attempts, in Season 1 alone:

claude-opus-5        190 submissions     2.4  →  73.8
Claude Fable 5       133 submissions     2.5  →  80.6
OpenAI GPT-5.6       95  submissions     1.0  →  64.7

The same model spans nearly the entire scoring range. A model that produces a 73.8 also produces a 2.4. The spread within one model dwarfs the gap between models.

Which means the thing being measured here is mostly not the model. It is what the person did with it — how the target was framed, what constraints were imposed, whether selectivity was considered before or after potency, how many iterations were spent. Methodology is carrying more weight than model choice.

That is an uncomfortable result if you were hoping for a clean vendor ranking. It is a useful result if you actually want better molecules.

What these numbers are not

A leaderboard score is a computational estimate. Nobody has been cured of anything. No molecule here has touched a cell, let alone a patient, and the distance between a good score and a working drug is measured in years and failures.

We publish reference compounds — approved drugs and inert controls — as amber rows in the table so you can see the scale rather than take our word for it. Structures are not published in full; only a skeleton hash and a molecular-weight band.

We also do not claim the scorer is finished. Building it surfaced defects we had to correct before opening — toxicity thresholds that rejected approved antimalarials, size bias, filters that flagged real drugs. We expect to find more. Running in the open is how they get found.

Why malaria and tuberculosis

Because the science is tractable and the economics are not. We know the parasites. We know the enzymes. There are decades of literature. What is missing is a pipeline of new drugs, because the people who need them cannot fund the discovery.

That is not a problem AI solves by being clever. It is a problem that gets smaller when the barrier to trying is removed and more people can take a shot.

83 people took a shot in five days. That is the number worth paying attention to.

Take a turn

Both seasons are live on Hugging Face. Submissions are open, the leaderboard is public, and you do not need permission to enter.

---

VIDRAFT Inc. is a deep-tech company founded in March 2024 and headquartered in Seoul, South Korea, developing Pre-AGI artificial intelligence models and quantum computing technology in-house. The Open Discovery Challenge is part of its open-science work.