How a Korean AI startup is turning open science into a global race for new medicines
TL;DR: VIDRAFT, a Korean AI startup, launched the Open Discovery Challenge on Hugging Face — a public competition inviting anyone to use AI to propose novel drug candidates for neglected diseases. Within just three days of launch, the platform had received more than 2,000 candidate molecule submissions. Early data from the challenge suggests that prompting strategy and tool selection may matter as much as the choice of AI model itself.
VIDRAFT (비드래프트), the Korean AI startup led by CEO Minsik Kim, announced on August 18, 2026 that its newly launched Open Discovery Challenge had already attracted over 2,000 candidate molecule submissions within three days of going live on Hugging Face. The project invites researchers, developers, and curious generalists alike to use AI — whether commercial frontier models or their own custom systems — to identify promising new drug candidates for diseases that have long been neglected by private R&D.
The Open Discovery Challenge is structured as a living, multi-season competition rather than a conventional static AI benchmark. Instead of testing models against a fixed set of correct answers, participants submit AI-generated molecular structures that are then evaluated across multiple dimensions: potential therapeutic efficacy, toxicity profile, target-binding affinity, ADME properties, and simulated pre-clinical and clinical viability. Each submission receives a composite score and is ranked on a public leaderboard.
Two disease seasons are currently open for participation. Season 1 targets malaria and Season 2 targets tuberculosis — both conditions where commercial R&D incentives have historically been insufficient relative to global disease burden. VIDRAFT provides an AI-assisted drug discovery guide so that participants without specialized pharmaceutical training can still contribute meaningful submissions using whatever AI tools they prefer.
Early score data from the challenge has already produced notable findings. In the malaria season, submissions generated using Claude-family models achieved a median score of 43.7 points, outperforming those from OpenAI-family models, which posted a median of 31.7 points — a gap of 12.1 points. The tuberculosis season followed the same pattern, with Claude-family submissions scoring 39.9 and OpenAI-family submissions scoring 30.9. The fact that the same ordering appeared across two distinct disease targets gives the trend added weight, though VIDRAFT notes that data continues to accumulate.
Chinese open-source models — including DeepSeek, Qwen, and Kimi — also performed competitively, with a combined median of 37.5 points that outpaced OpenAI-family submissions. Importantly, the difference between the Chinese open models and OpenAI-family models did not reach statistical significance in either season based on current sample sizes, suggesting that freely accessible AI models are capable of producing drug candidates that compete with those from top commercial systems. Gemini-family submissions currently show the lowest median score, though with only 12 entries recorded so far, VIDRAFT cautions against drawing firm conclusions about model capability at this stage.
Perhaps the most striking finding is the sheer range of scores among submissions that used the same underlying AI model — from single digits to the high 70s. VIDRAFT interprets this as evidence that the entire scientific workflow surrounding the model — the prompt design, the external chemistry and search tools integrated with the model, and the overall research strategy — can be just as decisive as which model was chosen. As VIDRAFT framed it, the true unit of competition is not a bare LLM but the complete "scientific agent" composed of model, prompt, tools, and research approach.
CEO Minsik Kim stated: "The next competition in AI is not about getting a few more benchmark questions right — it's about whether AI can actually discover new substances and scientific answers." He added that VIDRAFT intends to develop the Open Discovery Challenge into a global open-science platform where anyone can contribute to the discovery of medicines humanity needs, using the AI available to them.
The Open Discovery Challenge reframes what an AI benchmark can be. By evaluating AI systems on real-world molecular discovery rather than curated test sets, VIDRAFT is generating data that may become a meaningful indicator of AI's capacity for genuine scientific contribution — distinct from standard language or reasoning benchmarks.
The project also addresses a structural problem in pharmaceutical research. Malaria and tuberculosis primarily affect populations in lower-income regions, where market incentives for drug development are weak. By lowering the barrier to participation and aggregating global computational effort, VIDRAFT is attempting to apply the logic of open-source software development to drug discovery for underserved diseases. That even hobbyist participants with no lab access can submit competitive candidates underscores how dramatically AI is redistributing the ability to do frontier science.
Q: What is VIDRAFT's Open Discovery Challenge?
A: It is an open public competition hosted on Hugging Face where participants use AI tools to propose novel drug candidate molecules for neglected diseases, currently malaria and tuberculosis, with submissions scored on multiple pharmaceutical criteria.
Q: Who can participate in the Open Discovery Challenge?
A: Anyone can participate. VIDRAFT provides an AI drug discovery guide so that people without specialized pharmaceutical research backgrounds can use their preferred AI model to generate and submit candidate molecules.
Q: Which AI models have performed best in the challenge so far?
A: Based on early data, Claude-family models achieved the highest median scores in both the malaria and tuberculosis seasons. Chinese open-source models including DeepSeek, Qwen, and Kimi also outperformed OpenAI-family models on median scores, though the difference was not statistically significant in current sample sizes.
Source: AI타임스 (2026-08-18) — original article