VIDRAFT has opened the first season of the Open Discovery Challenge, a public challenge aimed at designing new drug candidates against malaria and evaluating them through a verifiable leaderboard. The first season focuses on the PfDHODH enzyme found in the malaria parasite, while excluding inhibition of its human counterpart, Human DHODH, because targeting the human enzyme is associated with immunosuppression rather than treating the infection, according to the material published on the Hugging Face blog.
The season is titled Open Discovery Challenge #1 — Malaria and closes on September 30, 2026. The highest-ranked entry at the end of the season will receive a $1,000 prize. VIDRAFT says the prize is not full compensation for the time or model usage involved, but rather a way to record contributions to neglected-disease drug discovery, with the possibility of increasing the prize value if sponsors join the initiative.
Why does the initiative focus on verification?
The material begins from the premise that AI models are now capable of producing thousands of candidate molecules in a single day, but that this abundance in design has not been matched by a comparable expansion in verification. A model can propose a chemical structure that appears convincing, but it cannot by itself determine whether it kills the parasite, harms the patient, can be manufactured, or actually reaches its target inside the cell.
VIDRAFT says the challenge is to make verification tools available at a broader scale through computational evaluations covering molecular binding, intracellular activity prediction, absorption, distribution, metabolism, and toxicity properties, selectivity, and manufacturability. The leaderboard displays reference drugs and inactive compounds alongside new entries, allowing participants to examine the classifier's ability to distinguish between them rather than relying solely on the organizer's claims.
Tests before opening the leaderboard
The material explains that building the classifier revealed 14 defects before the challenge was launched. Among the problems was that the initial toxicity thresholds rejected approved antimalarial drugs, as well as coffee, because heart and liver toxicity models were biased against large, lipophilic molecules. VIDRAFT therefore adopted a rule requiring approved drugs to be allowed through before applying any threshold to exclude participants.
The organization also found that calculating binding efficiency based solely on molecular size could disproportionately reward small molecules with weak binding; coffee nearly tied with the clinical candidate DSM265. A minimum potency threshold was added to address this. In another test, the intracellular activity model predicted coffee to be an active compound at 1 micromolar because the training data lacked sufficient records of failure. After adding 5,190 failure records and retraining, the gap between DSM265 and coffee widened from 1.00 to 1.74 logarithmic units.
The material also notes that the docking test returned an incorrect value for a known compound, before it was discovered that the invocation pathway itself was broken. In addition, using an incorrect function name when calculating chemical fingerprints produced results that appeared normal without raising an error, until a periodic verification process revealed the problem. In the confidence-interval test, measured coverage was 83% instead of 90%, so VIDRAFT expanded the correction margin until the measured rate reached 90.01%.
How are entries evaluated?
The challenge asks participants to design compounds that inhibit PfDHODH and reach the target inside the parasite while avoiding Human DHODH. The evaluation framework presents dimensions including intracellular activity, target binding, selectivity, ADMET properties, novelty, and manufacturability. Intracellular activity is awarded 30 points, target binding 20 points, ADMET 15 points, novelty 10 points, and manufacturability 5 points, while the material presents selectivity as an independent dimension without specifying its points in the published table.
Novelty is affected by potency; a new inactive compound receives no practical value in this dimension, and a close analog of a known drug loses novelty points. Scores for uncertain predictions are reduced, while compounds with predicted liabilities or severe insolubility are ranked below clean entries rather than being automatically rejected.
Participation and ownership rules
Any model may be used, including OpenAI, Claude, Gemini, Qwen, KIMI, DeepSeek, and open-weight models, followed by submission of the chemical structure in SMILES or InChI format. Compounds exceeding 550 in molecular mass, containing covalent reactive groups or PAINS patterns, or matching an entry previously registered are rejected. Molecular formulas alone are also rejected because they may represent multiple isomers.
VIDRAFT says ownership of the submitted molecule remains with the participant, and that it does not take patent rights or ownership, transfer the structure to third parties, or use it in its own pipeline. A participant may keep the structure private, but it remains stored with VIDRAFT for evaluation purposes. Publishing it publicly may affect patent eligibility because publication constitutes disclosure; therefore, the initiative warns those with commercial intentions not to publish before filing.
VIDRAFT emphasizes at the conclusion of the material that the displayed results are computational evaluations of candidates, not experimental measurements or claims about their efficacy or safety, and that they do not represent a ranking compared with approved drugs. Experimental validation remains a separate stage.