VIDRAFT's Open Discovery Challenge Hits 2,000 Submissions in 3 Days — AI-Driven Drug Discovery Goes Open Science

TL;DR: VIDRAFT has launched the Open Discovery Challenge, an open-science drug discovery platform hosted on Hugging Face where participants use any LLM of their choice to propose novel molecular structures for neglected diseases. Unlike static benchmark leaderboards, this is a generative, real-world task scored on preclinical simulation criteria — and it crossed 2,000 submissions within 72 hours of launch.

---

What it is

The Open Discovery Challenge is a community-driven, participatory drug-discovery project built by VIDRAFT, a Korean AI-for-science startup. Rather than asking participants to solve a fixed test set, it invites researchers, engineers, and hobbyists to:

• Choose any AI model (commercial or open-source) as their reasoning engine
• Use that model to propose novel molecular structures as drug candidates
• Submit those structures to an automated scoring pipeline that evaluates preclinical viability

The challenge is structured in seasons, each targeting a specific neglected disease:

• Season 1 — Malaria
• Season 2 — Tuberculosis

Neglected diseases were deliberately chosen because low commercial profitability has historically made them under-resourced in private R&D. By crowdsourcing AI-assisted molecular design, VIDRAFT aims to redirect global AI capability toward high-need, low-investment therapeutic areas.

The project is hosted publicly on Hugging Face, making it accessible to anyone with an internet connection and access to an AI model — no wet lab or institutional affiliation required.

---

How it works

At a conceptual level, the pipeline has three stages:

1. Proposal generation. Participants use an LLM or any AI model of their choice, guided by their own prompting strategy and supplementary cheminformatics tools, to generate candidate molecular structures.

2. Automated preclinical scoring. Each submitted structure is evaluated against a composite score that integrates:

   • Efficacy potential — predicted likelihood of biological activity against the target pathogen
   • Toxicity — computational safety estimates
   • ADME (Absorption, Distribution, Metabolism, Excretion) — pharmacokinetic simulation metrics

3. Leaderboard and data aggregation. Scores are aggregated publicly, allowing the community to observe which models and prompting strategies produce higher-quality molecular candidates.

The key design insight is that this is not a reading-comprehension or reasoning benchmark — it is a generative science task. The evaluation signal comes from domain-specific preclinical simulation, not human annotation or held-out labels.

---

Benchmarks & results

Early data from the first 2,000+ submissions (as of August 18, 2026) reveals several publicly reported findings:

Model-level median scores (malaria track):

• Claude-family models: 43.7 (median)
• OpenAI-family models: 31.7 (median)

Model-level median scores (tuberculosis track):

• Claude-family models: 39.9 (median)
• OpenAI-family models: 30.9 (median)

Open-source models (DeepSeek, Qwen, and similar): 37.7 (median, aggregated across both tracks)

Critically, the score distribution within a single model family spanned from approximately 1 to 78 points. This enormous intra-model variance is the most technically significant finding so far: it strongly suggests that prompting strategy and tool integration matter more than model selection alone. Two participants using the same underlying LLM can produce results separated by over 70 score points depending on how they structure their queries and what cheminformatics tooling they combine with the model.

⚠️ Note: Submission counts vary significantly across model families at this early stage, so cross-model comparisons should be interpreted cautiously. The ranking may shift as more data accumulates.

---

How to try it

The Open Discovery Challenge is publicly accessible on Hugging Face. You can navigate directly to VIDRAFT's Hugging Face profile to find the challenge dataset, submission interface, and leaderboard.

# Browse VIDRAFT's public Hugging Face presence
huggingface-cli whoami   # confirm you're logged in
# Then visit: https://huggingface.co/VIDRAFT

No specific model, endpoint, or SDK is mandated — participants bring their own AI tooling. The workflow is intentionally model-agnostic:

1. Access the challenge page on Hugging Face
2. Review the task specification and scoring rubric
3. Use any LLM (API-based or local) to generate molecular structure proposals
4. Submit your structures through the platform's submission interface
5. Track your score on the public leaderboard

---

FAQ

Q: Do I need a chemistry or biology background to participate?
A: Not necessarily. The platform is designed so that an AI model can assist with domain knowledge. However, early results suggest that participants who understand how to combine AI outputs with cheminformatics tools (e.g., for structure validation or ADME pre-filtering) achieve substantially better scores. Basic familiarity with molecular representations like SMILES strings would be an advantage.

Q: Which AI model should I use for the best results?
A: Early aggregate data shows Claude-family models posting higher median scores in both malaria and tuberculosis tracks, but the intra-model score spread (1–78 points) dwarfs the inter-model differences. The data strongly implies that how you prompt and what tools you pair with the model is the dominant performance factor — not the model itself. Open-source models like DeepSeek and Qwen are competitive with commercial options.

Q: Is this a benchmark or a real research contribution?
A: It is framed as both. Submitted molecular structures are evaluated via preclinical simulation pipelines, meaning high-scoring submissions represent genuinely plausible drug candidates — not just correct answers to trivia. VIDRAFT's stated goal is to build a global open-science platform for neglected disease drug discovery, implying that strong submissions may feed into downstream research workflows.

Q: Can individual developers or independent researchers participate, or is this for institutions only?
A: The platform explicitly targets individual researchers without access to large lab budgets or institutional infrastructure. Access to an AI model is the only hard requirement.

---

Originally reported by 동아일보 (2026-08-18) — source article.