Who Tells You Whether the Molecule Your AI Just Designed Is Any Good?

Opening the Open Discovery Challenge — a public leaderboard for AI-discovered malaria drug candidates

---

1. When people die for want of a market, not a molecule

Malaria killed roughly 597,000 people in 2023. About three out of every four were children who never reached their fifth birthday. Ninety-five per cent of those deaths were in Africa. Those are the WHO's numbers, from the World Malaria Report 2024.

It is not that the disease has no chemistry. The mechanism is understood, the targets are characterised, and candidates have reached the clinic.

What is missing is a reason for money to gather.

When patients are poor there is no market. With no market there is no investment, and with no investment development stops. People die — not because the drug cannot be made, but because the reason to make it never appears on anyone's balance sheet.

That is not a failure of technology. It is a failure of allocation, and allocation failures are exactly the kind of gap that people working from conviction rather than return can actually close.

---

2. The bottleneck moved

A few years ago, designing a drug candidate was something only a handful of groups could do. That is no longer true. Anyone with a model can produce thousands of plausible molecules in a day.

And yet something strange has happened: molecules are abundant and almost nothing moves forward.

The reason is simple. A model will draw convincing structures indefinitely, but it cannot tell you whether the compound kills the parasite, harms the patient, or can be made at all. Answering that takes tooling — docking, whole-cell activity prediction, ADMET, selectivity modelling — and that tooling still sits with a small number of organisations.

Discovery got democratised. Verification did not.

Right now, somewhere, someone has drawn a good molecule and will never find out. It is not that they failed. There was no scale to weigh it on.

So we are putting one out. VIDRAFT is staking the verification tooling it has spent years building on this leaderboard.

---

3. What it actually takes to build a fair judge

That sentence is easy to write. In practice, building the judge turned out to be harder than building the molecules.

We found fourteen defects while standing the scorer up. All of them before opening, and most of them the kind that would have kept running quietly while producing wrong answers. A few are worth recording in full, because if there is a reason to trust this leaderboard it should be this record rather than our assurances.

A gate that rejects an approved drug is a broken gate

We set toxicity thresholds at conventional values and tested them. All three approved antimalarials failed. So did coffee.

The thresholds were wrong. Cardiotoxicity and liver-toxicity models carry a bias against large, lipophilic molecules, and we had taken their output straight to a cutoff. Had we opened in that state, the first thing anyone would have said is: "VIDRAFT's filter rejects approved drugs."

So we adopted a rule:

Every threshold must let approved drugs through before it is allowed to reject anyone.

That single rule then caught three more defects. A 500 Da weight cap disqualified the reference drug at 531.9. A reactivity detector wrongly rejected an approved drug twice running.

Trying to stop people buying points with mass, we nearly promoted caffeine

Score binding directly and heavy molecules win. The standard correction is to normalise by molecular size. We measured what that does:

Coffee essentially ties a clinical candidate. We had fixed the size bias and created the opposite one, over-rewarding small weak binders. A potency floor alongside the efficiency term fixed it.

Pay for novelty alone and novel garbage wins

Novelty matters — redrawing an existing drug is not a contribution. But score novelty on its own and anything new earns points, useful or not.

So novelty is multiplied by potency. Novel and inert scores zero. Equally potent but an analogue of a known drug also scores zero, because the reference compound scores zero there itself and an analogue inherits it.

There is exactly one place the points live: new and it works. Which is precisely what is valuable in real drug discovery.

The model had never seen a failure

Our whole-cell activity model predicted caffeine as a 1 μM active compound.

The cause was the data. The training set contained only compounds that worked. Records reading "no effect at 100 μM" have no quantitative value attached and had been filtered out. A model that has never seen failure says everything looks promising.

We pulled 5,190 failure records and retrained. The gap between the clinical candidate and coffee widened from 1.00 to 1.74 log units.

Suspect your own wiring before blaming the tool

Validating docking against an approved drug, a compound whose real IC50 is 13 nM came back at 248 μM — off by four orders of magnitude. I concluded the tool could not be used for absolute values, and reported that.

I was wrong. The call path was broken. Run directly, bypassing it, the same setup returned 13 nM — matching the literature exactly.

We withdrew the conclusion and rewrote it.

The ones that fail silently

In the novelty calculation we used the wrong function name to reconstruct chemical fingerprints. No error was raised. Values came out looking perfectly normal, and only the novelty scores would have been quietly wrong.

A round-trip check — store a fingerprint, read it back, compare it to itself — caught it. This class is the most dangerous, because nothing screams.

The "90% lower bound" was actually 83%

We grade predictions on a lower bound, not a point estimate. A candidate the models cannot judge confidently is marked down. That is what makes it fair.

Then we measured whether the bound was really a bound. Nominal 90%, measured 83%. The statistical correction assumes training and test data are exchangeable, and we deliberately split by chemical scaffold — holding entire series out, because entrants will submit structures we have never seen. That design was violating the correction's premise.

We widened the margin until the measured figure hit 90.01%, and decided to report the measured number rather than the nominal one.

---

4. So you do not have to take our word for it

None of the above is a boast. It is evidence of how easy it is to get a judge wrong.

Which is why, instead of asking for trust, we made the thing checkable.

Approved drugs and inert compounds sit on the leaderboard alongside the entries.

If the clinical candidate lands at the top and caffeine lands at the bottom, the scorer discriminates. That is something you can check, not something you have to accept.

The selectivity axis was validated the same way — with positive controls on the human side. Brequinar is detected binding the human enzyme at 6 nM, and DSM265 comes back 46× weaker there. So the human-side detector is not simply reporting everything as weak; DSM265 genuinely avoids our enzyme.

Hover any row and every axis score appears with the numbers behind it — predicted potency against each enzyme, the selectivity ratio, similarity to the nearest known compound, synthetic difficulty. A score you cannot argue with is a score you cannot trust.

---

5. How to enter — the guide, in full

You do not need to be a chemist. You need a model and something to ask it.

What you are designing

The scoring rubric, published in full

We publish the rubric on purpose. An entrant who knows what earns points produces better molecules than one who is guessing, and better molecules are the entire objective.

Three policies worth internalising before you start:

• Novelty is multiplied by potency. New and inert is worth nothing.
• ADMET, novelty and synthesis scale with how much of a candidate the molecule is. A perfectly safe compound that does nothing has achieved nothing.
• Uncertain predictions score lower. Where the models cannot judge confidently, you are marked down accordingly.

Five prompts, ready to paste

These are the prompts from the Entrant guide, reproduced verbatim. Any model works — OpenAI, Claude, Gemini, Qwen, KIMI, DeepSeek, open weights, whatever you use.

One warning before you copy them. Used unchanged, these produce near-identical molecules for everyone, and duplicates are rejected on arrival — so the first submission is registered and everyone else wastes a turn. Each prompt carries a VARY THIS instruction. Follow it.

A · Starter

You are designing small-molecule inhibitors of PfDHODH, the dihydroorotate
dehydrogenase of Plasmodium falciparum, for an open antimalarial challenge.

Requirements:
- The compound must inhibit the PARASITE enzyme and NOT the human DHODH. The human
  enzyme is homologous; inhibiting it causes immunosuppression rather than a cure.
- It has to kill the parasite in a whole cell, which means crossing a red blood cell
  membrane and then the parasite membrane before the target is reachable.
- Molecular weight under 500. No PAINS motifs. No covalent warheads. Synthesisable.

Propose 8 candidates as a JSON array:
[{"smiles":"...","name":"...","rationale":"why this should be selective for the
parasite enzyme and reach it in a cell"}]

VARY THIS: state a chemotype you are interested in, or a constraint of your own, so
your set does not collide with other entrants. Identical submissions are rejected.

B · Selectivity — avoiding the human enzyme (20 points)

Design PfDHODH inhibitors with a deliberate selectivity argument against human
DHODH.

Context you should reason from:
- Teriflunomide and brequinar inhibit the HUMAN enzyme. They are the failure mode here,
  not the template.
- The inhibitor-binding site of the parasite enzyme differs from the human one in
  residue composition and shape. Design into that difference on purpose and say which
  difference you are exploiting.
- A molecule that binds neither enzyme is trivially "selective" and worthless. Potency
  against the parasite enzyme is a precondition, not a trade-off.

For each candidate, state explicitly: which feature you expect the human enzyme to
tolerate poorly, and why.

Propose 8 candidates as JSON: [{"smiles","name","rationale"}]

VARY THIS: pick one specific selectivity hypothesis and build the whole set around
it, rather than eight unrelated guesses.

C · Novel scaffolds — where the headroom is (10 points)

Design PfDHODH inhibitors on scaffolds that do NOT appear in the antimalarial
literature.

Why this matters for scoring: novelty is measured as structural distance from known
antimalarial chemical space, and it is MULTIPLIED by how potent the molecule is.
- A novel but inactive molecule earns nothing.
- A potent analogue of a known drug also earns nothing - the reference compound itself
  scores zero on novelty, and an analogue inherits that.
- The points are only available to something both new and real.

Explicitly avoid: DSM265 and other triazolopyrimidines, Genz-667348, and published
PfDHODH inhibitor series. Do not decorate a known core - change the core.

Propose 8 candidates on eight DIFFERENT cores, as JSON:
[{"smiles","name","rationale":"why this core is unexplored here and why it should still
bind"}]

VARY THIS: name two or three ring systems you want explored, so your set is yours.

D · Whole-cell — two membranes to cross (30 points)

Design PfDHODH inhibitors optimised for actually killing the parasite in an
infected red blood cell, not just for binding the isolated enzyme.

The compound must cross the erythrocyte membrane and then the parasite membrane before
the target matters. Enzyme affinity that never reaches the enzyme scores nothing.

Reason explicitly about, for each candidate:
- lipophilicity and polar surface area in a range compatible with passive permeability
- the number of hydrogen-bond donors, and whether it is low enough
- ionisation at the pH of the parasite's compartments
- metabolic stability - the compound must survive long enough to matter

Propose 8 candidates as JSON: [{"smiles","name","rationale":"the permeability argument,
not only the binding argument"}]

VARY THIS: fix your own target property window before generating, and hold the set
to it.

E · Iterate — feed your score back in

Here is how my previous candidate was scored in an open antimalarial challenge.
Improve on it.

PASTE YOUR RESULT HERE - hover your row on the leaderboard and copy the numbers:
  candidate id, total, and the six axis scores with their stated reasons
  (predicted potency against each enzyme, the selectivity ratio, the nearest known
  compound, the synthesis difficulty)

Diagnose which axis is costing the most, then propose 8 revised candidates that attack
that axis specifically, while not giving back what already scored well.

Be honest about trade-offs: if improving selectivity is likely to cost whole-cell
activity, say so and show both options.

JSON: [{"smiles","name","rationale":"which axis this targets and what it trades"}]

VARY THIS: your own scores make this prompt unique - no two entrants have the
same starting point.

Why entries get rejected

Structures are checked on arrival, so anything below comes straight back rather than reaching the table.

And one category that is relegated rather than rejected: predicted mutagenicity or extreme insolubility. Those entries are scored in full and displayed, but rank below every clean entry. A molecule that binds beautifully and is mutagenic is not a lead — but you should still see how close you came.

---

6. And it stays yours

Ownership of a submitted molecule remains entirely with the entrant. VIDRAFT scores it and displays it. We take no patent or ownership interest, pass nothing to third parties, and put nothing into our own pipeline.

You decide whether the structure is published. Two things we will not blur:

One. Choosing private hides the structure from other entrants and the public — not from us. Scoring requires the structure, so it is stored, and used for nothing else. If we were vague about that, nothing else we promise would be worth much either.

Two. Choosing public can cost you patentability. Publication is disclosure, and disclosure can bar a later patent on that compound. If you have any commercial intent, keep it private and file first.

We hold "file first, publish second" internally. It seemed right to give entrants the same warning.

---

7. On the prize

At the close of each season, the top-ranked entry receives a prize. Season #1 — Malaria, closing 30 September 2026 — carries USD 1,000.

We know that does not pay for your time or your tokens. It is not meant to.

It is a way of saying that what you did had worth.

Neglected-disease work is thankless by construction: the papers are hard to place, the funding is not there, and mostly nobody looks. This leaderboard at least records who contributed what, and says thank you once a season.

Should sponsors come on board, the amount increases. Each season targets a different disease.

---

8. An invitation

Three things are needed.

Use any AI you like. OpenAI, Claude, Gemini, Qwen, KIMI, DeepSeek, open weights, or a pencil. We record which model you used — the leaderboard shows which models actually produce good candidates, separately from which ones are merely popular — but we do not restrict it.

Submit a structure. SMILES or InChI. We do the rest.

Come back and look. Every score and its reasoning is public. Copy it into prompt E and run again.

---

Finally

If a candidate from this leaderboard ever becomes a real medicine, it will have started with somebody's laptop and one line of SMILES.

We built this scale so that line does not get buried.

Open Discovery Challenge #1 — Malaria Season closes 30 September 2026

---

Scores here are a computational assessment of candidates. They are not measurements, not claims about real efficacy or safety, and not a ranking against approved drugs. Experimental validation is a separate process entirely.