VIDRAFT_LAB

The medieval battlefield belonged to the knight. Armour, a warhorse, a lifetime of training — only the few who could afford all three ever stood on it.

What ended that was not a better knight. It was gunpowder. A gun in a farmhand's hands made generations of inherited skill beside the point. The knights did not disappear; what changed was who was allowed on the field.

Drug discovery has had the same shape. Drawing a molecule is the easy half. Knowing whether it is worth anything is the hard half, and that tool belonged to whoever owned the laboratory and the compute. Everyone else could have the idea and no way to test it.

We spent years building that judgement. Verifying PharmaOS took months on IBM's Heron quantum processor.

And we did not ask anyone to take our word for it.

First, we competed on someone else's leaderboard. On Polaris Hub, the benchmark for AI drug-property prediction, our engine holds first place in 16 ADME and toxicity tracks. The most recent came last week: predicting whether a compound crosses the cell membrane (MDR1-MDCK efflux ratio) from structure alone — correlation 0.802, against 0.725 for second place.

Second, we put a molecule through. A candidate from this engine has been filed as a composition-of-matter patent and taken into preclinical work. The experimental results came back inside the error bars of the prediction.

That engine is what scores your molecules here.

And the method is the same: approved drugs and inert compounds sit in the same table as your entries. We are not asking you to trust the score. You can check it.

Tools spread, and the field changes. That is what gunpowder did.

Progress report

Season #1 Malaria — 42 hours since opening, 919 entries. 222 of them now score above DSM265 (50.880), a clinical-stage antimalarial. Current best: 73.846.

Season #2 Tuberculosis — 12 hours since opening, 224 entries. 73 of them score above delamanid (39.164), the highest-scoring approved TB drug on our ladder. Current best: 58.270.

Season #2 is arriving faster than Season #1 did.

The tuberculosis season opened with its reference ladder published first. Fourteen drugs from the WHO tuberculosis guidelines and five inert controls were put through the same scorer, and entries were only accepted once the approved drugs sat clearly above the compounds that do nothing. Those results are on the leaderboard as reference rows — scored on the same scale, sitting beside your molecules.

Thank you for taking part.

🚀 https://huggingface.co/spaces/FINAL-Bench/open-discovery-challenge

354 entries in a day.

42 entrants, 22 still in the queue. Thank you for the response — it has been well beyond what we expected.

Two numbers worth sharing. 58 entries now score above DSM265 (50.880), a clinical-stage antimalarial. And of the ceiling we measured ourselves — the highest score reachable by simply walking substituents — nine entries have already passed it. The current leader sits well above it.

One of the hard questions in opening this challenge was whether AI can actually find something, or only learn to satisfy a scorer. A day of data has started to answer it, and the answer belongs to the people who entered.

Season #1 closes 30 September. The full measurements are published then.

https://huggingface.co/spaces/FINAL-Bench/open-discovery-challenge

Three things.

The bias is real. But "every result on that board is noise" doesn't match what we measured. We measured the ceiling — the highest score that bias can produce — and the current top entry sits well above it. Yesterday's leader was explained by it. Today's is not. That changed in a day. We will re-measure that trend before any prize is awarded, and if the top of the board has converged on the bias, we will publish how the winner is decided before the deadline, not after. We are not going to award a prize on a basis we know is compromised.

You're right that private structures sit on our servers. We won't argue that. Scoring needs the structure, so it is stored. It is used for nothing else, passed to no one, and we claim no patent or ownership interest. And yes — that is a request to trust us. That judgement is the entrant's to make, which is why private is the default and why we tell people that publishing can cost them patentability before they submit, not after.

You're right that this is in silico. Our scores are computational assessments of candidates, not measurements. We can't promise wet-lab validation, and we won't promise what we can't deliver.

The challenge has been open for a day and has taken 269 entries. The full measurements go public when the season closes. We can argue about the numbers then.

Thank you for the scrutiny. Two of your points are right; one is factually wrong.

You're right that the scorer can be gamed. We suspected the same thing and ran an adversarial probe against ourselves. Without seeing the top entry's structure — it's private — we reached the same molecule by systematically walking substituents. That means the binding axis carries a bias with no chemical basis behind it, and we have the measurements. We also measured the scorer's own reproducibility: ±0.51 points, run to run. Season rules don't change after a season opens, so Season #1 closes under the rubric as published, and the defect is corrected in Season #2. Publishing the specifics now would hand an advantage to whoever read this thread, so we'll release the full measurements when the season closes.

You're right that this is in silico. Our scores are computational assessments of candidates, not measurements. The page says so. We can't promise wet-lab validation today, and we won't promise what we can't deliver.

The patent point is wrong. We state that risk before anyone submits: "Publishing a structure can cost you patentability. If you have commercial intent, keep it private and file first." Private is the default, and most of the current top entries chose it. Private structures never appear on the leaderboard.