VIDRAFT's Open Discovery Challenge: Claude vs. GPT-4 on Malaria & TB Drug Candidates — Early Benchmark Data
TL;DR: Korean AI startup VIDRAFT launched Open Discovery Challenge on Hugging Face — an open-format drug discovery competition where participants use any frontier LLM (Claude, ChatGPT, Gemini, DeepSeek, etc.) to propose real candidate molecules for malaria and tuberculosis. Early 3-day data shows Claude-family submissions scoring 43.7 vs. OpenAI-family at 31.7 on the malaria track, with the same ranking order repeating on the TB track. Developers and ML engineers can participate directly on Hugging Face right now.
What it is
Open Discovery Challenge is a publicly hosted, open-enrollment drug discovery benchmark created by VIDRAFT and published on Hugging Face. Unlike traditional fixed-dataset benchmarks — where models are evaluated on held-out prediction accuracy — this challenge is generative and competitive: participants freely explore chemical space using whichever AI model they choose, then submit actual candidate molecules for scoring.
Key structural facts (from public reporting):
- Hosted on: Hugging Face (public, accessible to any registered user)
-
Disease targets (initial seasons):
- Season 1 — Malaria
- Season 2 — Tuberculosis (TB)
- Eligible AI tools: OpenAI models, Claude (Anthropic), Gemini (Google), DeepSeek, Qwen, KIMI, or custom/proprietary AI
- Scale reached in 3 days: 2,000+ candidate molecules submitted
- First publicly reported: Jeon Ja Sinmun (전자신문), August 18 2026; VIDRAFT press release from Seoul, August 19 2026
The choice of malaria and TB as initial targets is deliberate. Both diseases carry enormous global burden — particularly in low-income populations — yet face a well-documented structural R&D gap: low commercial returns deter private pharmaceutical investment, a problem the WHO has flagged for decades. VIDRAFT is using open AI competition to lower the barrier to generating novel candidate compounds for these neglected disease areas.
How it works
At a conceptual level, the challenge follows an open-ended generative evaluation loop:
- A participant picks their AI tool (any publicly available frontier model or custom system).
- They use that AI to explore chemical space — reasoning about molecular structures that might satisfy target-binding, safety, and pharmacokinetic constraints for the disease in question.
- They submit candidate molecules to the challenge platform on Hugging Face.
- Submissions are scored against a composite metric that the source describes as encompassing: predicted drug efficacy, toxicity estimates, target-binding affinity, and ADME (Absorption, Distribution, Metabolism, Excretion) simulation outputs.
- A live leaderboard surfaces scores grouped by the AI family used, enabling cross-model comparison at the community level.
This is a meaningful architectural departure from static benchmarks. Rather than testing whether a model can predict a known answer, it tests whether AI-assisted generation can surface novel, plausible preclinical candidates — a far harder and more practically relevant task.
⚠️ Caveat: The exact scoring algorithm, weighting of sub-metrics, and specific model versions used by submitters have not been publicly disclosed by VIDRAFT. The numbers below reflect early-stage data.
Benchmarks & results
These are the publicly reported early figures, captured from the first 3 days of challenge activity:
| Track | AI Family | Median Score |
|---|---|---|
| Season 1 — Malaria | Claude family | 43.7 |
| Season 1 — Malaria | OpenAI family | 31.7 |
| Season 2 — Tuberculosis | Claude family | 39.9 |
| Season 2 — Tuberculosis | OpenAI family | 30.9 |
What to take away — and what to hold loosely:
- The Claude-family advantage (~12 points on malaria, ~9 points on TB) repeated across two independent disease tracks, suggesting the gap is not a single-season artifact.
- However, VIDRAFT itself characterized these as early-stage data. Sample sizes were small relative to a mature benchmark, specific Claude and GPT model versions used by submitters were not disclosed, and the challenge is still ongoing — rankings can and likely will shift.
- Do not read this as a definitive claim that Claude is categorically superior at drug discovery. It is a directional signal from a live, evolving competition.
How to try it
The challenge is publicly accessible on Hugging Face. Based on public reporting, you can find it by searching for VIDRAFT or "Open Discovery Challenge" on the Hugging Face platform:
# Browse or search on Hugging Face directly
# huggingface.co — search: "Open Discovery Challenge" or "VIDRAFT"
No specific dataset slug, model card URL, or API endpoint has been officially confirmed in the public sources available at time of writing. Visit huggingface.co and search for the challenge name to find current submission instructions, scoring criteria, and the live leaderboard.
Participation requires no lab equipment — only an account on Hugging Face and access to at least one LLM (commercial API or self-hosted).
FAQ
Q: What does the challenge score actually measure?
A: Based on VIDRAFT's public description, scores aggregate simulated estimates of drug efficacy, toxicity, target-binding affinity, and ADME properties for submitted candidate molecules. The precise sub-metric weights and scoring algorithm have not been publicly released.
Q: Can I participate using an open-source or self-hosted model?
A: Yes — the challenge explicitly lists custom/proprietary AI as a valid tool alongside commercial frontier models. Any AI-assisted molecular generation approach is in scope.
Q: Are the results statistically conclusive?
A: No, and VIDRAFT acknowledges this. The figures come from the first three days of an ongoing competition with a limited initial sample size. Treat them as early directional data, not a peer-reviewed benchmark conclusion.
Q: Will VIDRAFT expand to other disease targets beyond malaria and TB?
A: The source article indicates this is a possibility under consideration, but no confirmed additional targets or timelines have been announced publicly.
Originally reported by 비젠메디컬 (2026-08-23) — source article.
Top comments (0)