A Korean AI startup is crowdsourcing neglected-disease drug candidates — and the results are already revealing which AI models perform best.
TL;DR: VIDRAFT, a Korean AI-for-science startup, launched the Open Discovery Challenge on Hugging Face and collected more than 2,000 participant submissions within three days of going live. The project asks researchers worldwide to use any AI model of their choice to propose novel molecular structures for malaria and tuberculosis drug candidates. Early data show meaningful performance differences across AI models and, more strikingly, across prompting and analysis strategies within the same model.
VIDRAFT, the Korean AI startup focused on scientific research, announced on August 18, 2026 that its publicly available drug-discovery initiative — the Open Discovery Challenge — surpassed 2,000 participation submissions in just three days after launching on the global AI platform Hugging Face. The milestone signals unusually rapid traction for an open-science project targeting diseases that have long been overlooked by commercial pharmaceutical R&D.
The Open Discovery Challenge is a departure from conventional AI benchmarking. Rather than asking participants to solve a fixed set of predefined problems, it invites researchers to freely propose entirely new molecular structures using whichever AI system they prefer — including OpenAI models, Anthropic's Claude, DeepSeek, Qwen, or any other tool. Submitted molecules are then evaluated against a composite score that factors in estimated drug efficacy, toxicity, and ADME (absorption, distribution, metabolism, and excretion) indicators drawn from preclinical simulation data.
The challenge is structured in thematic seasons. Season 1 targets malaria, and Season 2 focuses on tuberculosis — both classified as neglected tropical or poverty-related diseases where private-sector investment has historically been thin due to limited commercial upside.
VIDRAFT's own early-stage analysis of the first wave of submissions turned up some striking patterns. In the malaria track, submissions generated with Claude-family models recorded a median score of 43.7, compared with 31.7 for OpenAI-family models. The tuberculosis track showed a similar gap: Claude-family submissions scored a median of 39.9 against 30.9 for OpenAI-family entries. Notably, open-source models such as DeepSeek and Qwen were not far behind, with a combined median score of 37.7 — competitive with the leading commercial offerings.
VIDRAFT was careful to flag that submission counts vary across model categories, and that these early figures should not be treated as definitive rankings. More data will be needed before firmer conclusions can be drawn.
Perhaps the most telling finding from the initial dataset is the sheer spread of scores produced by participants using the same underlying AI model. Submissions clustered under a single model family ranged from the low single digits all the way to the high 70s on the scoring scale. According to VIDRAFT, this wide variance suggests that the choice of AI tool matters far less than how a researcher uses it — specifically, the quality of prompts they write and the analytical frameworks they layer on top of the model's raw output.
CEO Minsik Kim framed the project's ambition in a statement reported by Dong-A Ilbo: "The real competition for AI lies not in solving predefined problems but in finding genuinely useful molecular structures and scientific answers." He added that VIDRAFT intends to grow the Open Discovery Challenge into a global open-science platform where researchers everywhere can contribute to finding medicines that humanity truly needs.
The pharmaceutical industry has long struggled to attract investment in neglected diseases. Conditions like malaria and tuberculosis disproportionately affect lower-income populations, which translates into limited revenue potential for drug developers — and, consequently, chronic underfunding of early-stage research. Open-science approaches have tried to fill this gap for decades, but they have typically required access to expensive laboratory infrastructure or large institutional budgets.
VIDRAFT's model changes that equation. By running the discovery process through AI models that are freely accessible online, the challenge lowers the barrier to entry to near zero for individual researchers, independent scientists, or resource-constrained institutions. If the crowd-sourced approach can reliably surface viable molecular candidates, it could represent a meaningful shift in how early-stage neglected-disease research gets done.
The early engagement numbers — 2,000-plus submissions inside 72 hours — also point to a latent global appetite for participatory drug discovery that has not previously had a clear outlet.
Q: What is VIDRAFT's Open Discovery Challenge?
A: It is a crowdsourced, open-science project hosted on Hugging Face where participants propose novel drug candidate molecules using AI models of their choice, with submissions evaluated against simulated preclinical metrics including efficacy, toxicity, and ADME indicators.
Q: Which AI models performed best in the early submission data?
A: Based on median scores from the first wave of entries, Claude-family models led both the malaria and tuberculosis tracks, though VIDRAFT cautioned that submission volumes differ across model categories and results may shift as more data accumulates.
Q: Do participants need laboratory access or institutional affiliation to join?
A: No. The challenge is designed so that anyone with access to an AI model can participate, making it accessible to independent researchers and individuals without large research budgets or specialized facilities.
Source: 동아일보 (2026-08-18) — original article