A Korean AI startup's open science platform is turning AI model selection into a measurable public-health experiment.
TL;DR: VIDRAFT, a Korean deep-tech AI startup, launched an open drug discovery competition called the Open Discovery Challenge on Hugging Face, attracting more than 2,000 candidate molecule submissions within three days of going live. Early scoring data from two active disease seasons — malaria and tuberculosis — shows Claude-family models producing higher median scores than OpenAI-family models across both diseases. The platform is designed to let anyone, not just professional researchers, use AI tools to discover new drug candidates for neglected tropical diseases.
VIDRAFT, the Korean deep-tech AI and scientific research company, has transformed a critical gap in global pharmaceutical research into a live, real-time AI competition. Its publicly accessible Open Discovery Challenge, hosted on the global AI platform Hugging Face, drew over 2,000 candidate molecule submissions in just three days after launch — and the early performance data is already sparking debate about which AI models are best suited to genuine scientific discovery.
The Open Discovery Challenge is not a conventional AI benchmark. Rather than measuring how accurately a model answers predefined test questions, VIDRAFT's platform asks participants to use whichever AI they choose — including OpenAI, Claude, Gemini, DeepSeek, Qwen, KIMI, or their own custom models — to actively propose novel molecular structures as potential drug candidates. Each submitted molecule is then evaluated on a composite score covering pharmacological potential, toxicity, target-binding affinity, ADME properties, and simulated preclinical and clinical viability. Scores and rankings are published openly.
Two disease seasons are currently active: Season 1 targets malaria, and Season 2 targets tuberculosis. VIDRAFT selected both diseases deliberately because they disproportionately affect low-income populations in developing countries, and because the profit motive for large pharmaceutical companies to pursue these treatments has historically been limited — a textbook example of pharmaceutical market failure that WHO has long flagged.
The platform is designed to be accessible to non-specialists. VIDRAFT provides an AI drug discovery guide, meaning anyone comfortable using a modern AI tool can participate in the search for new candidate molecules.
The early scoring data coming out of the Open Discovery Challenge is arguably more interesting than the concept itself. In the malaria season, submissions made using Claude-family models achieved a median score of 43.7, while OpenAI-family submissions scored a median of 31.7 — a gap of roughly 12 points. In the tuberculosis season, the same ordering held: Claude models scored a median of 39.9 versus 30.9 for OpenAI models. The fact that an identical ranking emerged across two distinct diseases makes the trend harder to dismiss as statistical noise, though VIDRAFT's own reporting notes the need for more accumulated data before drawing firm conclusions.
Chinese open-source models — including DeepSeek, Qwen, and KIMI — also performed notably well, posting a combined median score of 37.7, which placed them above the OpenAI-family median. Crucially, the source reports that the difference between Chinese open models and OpenAI-family models was not statistically significant in the current sample across either season, suggesting that freely accessible or open-weight models can produce competitive drug candidates alongside top-tier commercial models.
Gemini-family models currently sit at the lowest median score of 17.1, but with only 12 total submissions recorded at the time of reporting, that figure does not yet carry enough statistical weight to serve as a reliable indicator of model capability.
Perhaps the most consequential insight to emerge so far is this: within a single model family, individual submission scores ranged from single digits all the way up to the high 70s. That variance makes it clear that the choice of AI model is only one factor. The quality of the prompt, the selection of specialized chemistry and analysis tools, and the broader research strategy each play a significant role in the final outcome. As VIDRAFT frames it, the real competitor in this arena is not a bare language model but a full scientific agent — an integrated system of model, prompt design, external tooling, and research methodology.
VIDRAFT CEO Minsik Kim stated that the next frontier for AI competition lies not in answering more benchmark questions correctly but in whether AI can genuinely discover new molecules and scientific answers. He described the Open Discovery Challenge as VIDRAFT's effort to grow an open global science platform where anyone can use their own AI to participate in finding drugs that humanity needs.
Q: What is VIDRAFT's Open Discovery Challenge?
A: It is an open AI drug discovery competition hosted on Hugging Face where participants use any AI model to propose novel molecular drug candidates for diseases like malaria and tuberculosis, with each submission scored on pharmacological potential, toxicity, target binding, ADME, and simulated clinical viability.
Q: Which AI model family performed best in the early results?
A: Based on early median scores across both the malaria and tuberculosis seasons, Claude-family models led the rankings, followed by Chinese open-source models (DeepSeek, Qwen, KIMI), and then OpenAI-family models. Gemini-family models had too few submissions to draw reliable conclusions.
Q: Why did VIDRAFT choose malaria and tuberculosis as the first target diseases?
A: Both diseases carry a heavy global burden but primarily affect low-income populations with limited purchasing power, creating a market incentive gap that discourages private pharmaceutical R&D. VIDRAFT designed the platform as a public-interest open-science initiative to address that gap using AI and crowd participation.
Source: 전자신문 (2026-08-18) — original article