VIDRAFT's Open Discovery Challenge: Benchmarking AI Science Agents on Real Drug Discovery Tasks
TL;DR: VIDRAFT has launched the Open Discovery Challenge on Hugging Face — an open-science platform where anyone can use any AI model to discover novel drug candidates for neglected diseases like malaria and tuberculosis. Over 2,000 molecular candidates were submitted within the first three days of launch, producing early data that reveals meaningful performance gaps between frontier AI systems and between different prompting/tooling strategies applied to the same model. If you care about LLMs as scientific reasoning engines rather than benchmark-solvers, this is worth watching.
What it is
The Open Discovery Challenge is a community-run AI drug-discovery competition hosted on Hugging Face by VIDRAFT (비드래프트), a Korean Pre-AGI AI startup. Unlike conventional AI benchmarks — which evaluate models on pre-existing question–answer pairs — this challenge asks participants to generate genuinely novel molecular structures as drug candidates.
Key facts from the announcement:
- Two active seasons are currently open: Season 1 (Malaria) and Season 2 (Tuberculosis).
- Target diseases were intentionally chosen as neglected diseases — conditions where low commercial profitability has historically led to insufficient private R&D investment.
- No pharmaceutical expertise required. VIDRAFT provides an AI-driven drug discovery guide so that developers, researchers, and enthusiasts without a chemistry background can participate.
- Scoring is multi-dimensional. Submitted molecules are evaluated across: drug efficacy potential, toxicity, target binding affinity, ADME properties, and preclinical/clinical feasibility simulation.
The platform's explicit goal is to aggregate global intelligence and computational resources toward public-good scientific problems.
How it works
At a conceptual level, the challenge repositions drug discovery as a generation + evaluation loop rather than a retrieval or classification task:
- Generate: A participant uses any AI model — commercial (OpenAI, Anthropic, Google, DeepSeek) or self-developed — along with custom prompts and external chemistry/search/analysis tools, to propose novel molecular structures.
- Submit: The proposed molecules are submitted to the VIDRAFT platform on Hugging Face.
- Score: VIDRAFT's scoring pipeline evaluates each submission across the multi-dimensional pharmacological rubric described above, then publishes scores and leaderboard rankings publicly.
What makes this architecturally interesting is the framing: the unit of competition is not a bare LLM but a science agent — the composite system of (AI model) + (prompt strategy) + (tool integrations) + (research workflow). This design choice is deliberate and reflects VIDRAFT CEO Minsik Kim's view that the next frontier of AI competition is real scientific discovery, not incremental benchmark improvements.
Benchmarks & results
Early data from the first three days (2,000+ submissions across both seasons) surfaced several notable patterns:
| Model Family | Malaria Season Median Score | Tuberculosis Season Median Score |
|---|---|---|
| Claude family | 43.7 | 39.9 |
| OpenAI family | 31.7 | 30.9 |
| Chinese open models (DeepSeek, Qwen, Kimi) | 37.7 (both seasons combined) | — |
| Gemini family | 17.1 | (low sample count) |
⚠️ Caveat: The Gemini family had only 12 submissions at time of reporting, making per-model conclusions premature.
Key qualitative findings:
- Claude-family submissions led in both seasons by a consistent margin (~12 points over OpenAI family).
- Chinese open-weight models were competitive. DeepSeek, Qwen, and Kimi collectively scored a median of 37.7 — above the OpenAI family — though the difference was not statistically significant in either season at current sample sizes.
- The same model can produce wildly different results. Scores from submissions using the same underlying AI ranged from single digits to ~78 points, demonstrating that prompt design, tool selection, and research strategy are at least as important as the choice of base model.
- The same ranking order appeared independently across two different disease targets, suggesting the trend may be robust rather than disease-specific.
How to try it
The challenge is publicly hosted on Hugging Face. VIDRAFT states that any participant can follow their provided AI drug discovery guide to find and submit candidates using whichever AI they already have access to.
To find the challenge on Hugging Face, search for VIDRAFT Open Discovery Challenge on huggingface.co. No specialized lab equipment, pharmaceutical license, or custom model is required to enter.
No public GitHub repository URL, installable package, or OpenAI-compatible API endpoint was announced in this release. Access is through the Hugging Face platform interface per VIDRAFT's published participation guide.
FAQ
Q: Do I need a chemistry or biology background to participate?
A: No. VIDRAFT explicitly designed the challenge for non-specialist participants and provides an AI-guided drug discovery walkthrough. The scoring pipeline handles the pharmacological evaluation automatically.
Q: Are the scores based on real wet-lab experiments or computational simulation?
A: Based on the source article, scoring is based on computational simulation of drug efficacy potential, toxicity, target binding affinity, ADME properties, and preclinical/clinical feasibility — not physical laboratory experiments. Wet-lab validation is not described as part of the current challenge scope.
Q: Can I use open-weight or self-trained models, or only commercial APIs?
A: Any AI model is permitted, including open-weight models and models you've developed yourself. The early leaderboard already includes submissions from Chinese open models (DeepSeek, Qwen, Kimi) alongside commercial frontier models.
Q: Why malaria and tuberculosis specifically?
A: Both are high-burden neglected diseases where market incentives have historically under-served global health needs. VIDRAFT explicitly frames the platform as an open-science response to this market failure.
Originally reported by AI타임스 (2026-08-18) — source article.
Top comments (0)