VIDRAFT's PharmaOS Closes the Loop: AI Drug Discovery Gets Cell-Level Experimental Validation
TL;DR: VIDRAFT has partnered with Pyo Biotech to connect its AI-driven drug candidate discovery platform, PharmaOS, to real wet-lab cell experiments — bridging the gap between in silico prediction and in vitro biological validation. Experimental results will feed back into the AI models, creating an iterative predict–validate loop. Developers and researchers working on computational drug discovery should take note of how this architecture addresses one of the field's most persistent blind spots.
What it is
VIDRAFT (비드래프트), a Korean Pre-AGI AI startup resident at Seoul AI Hub, operates PharmaOS — a computational drug research platform designed to score and prioritize small-molecule drug candidates. PharmaOS currently handles:
- Molecular binding prediction: estimating the likelihood that a candidate molecule binds to a target protein
- Drug-likeness assessment: evaluating ADMET-adjacent properties computationally
- Toxicity risk scoring: flagging candidates with potential safety issues before any wet-lab work begins
Separately, VIDRAFT runs the Open Discovery Challenge (ODC), a publicly accessible drug discovery competition platform. Participants submit AI-designed molecular structures, and VIDRAFT evaluates them using PharmaOS and related internal tools across dimensions including predicted efficacy, selectivity, safety, novelty, and synthetic accessibility.
As of September 6, 2026, the ODC has accumulated 13,067 submitted candidate molecules across four active disease tracks: malaria, tuberculosis, Chagas disease, and non-opioid analgesia.
The newly announced MOU with Pyo Biotech (표바이오텍), a cell-based assay and validation company, extends PharmaOS's reach from pure computation into experimental biology.
How it works
The collaboration is structured as a two-stage, feedback-driven pipeline:
Stage 1 — In silico (VIDRAFT side):
- Disease target identification using AI
- Candidate molecule design and property prediction via PharmaOS
- Computational pre-screening to rank candidates before any experiment is run
Stage 2 — In vitro (Pyo Biotech side):
- Cell-line and disease-model-based assays using Pyo Biotech's existing biological infrastructure
- Measurement of actual efficacy and cytotoxicity at multiple concentration levels
- Mechanism of action (MoA) analysis on selected candidates
The key engineering insight here is the feedback loop. Wet-lab results from Stage 2 are shared back with VIDRAFT to retrain and refine the AI models in Stage 1. This is a classic active learning architecture applied to drug discovery: the model generates predictions, experiments generate ground-truth labels, and those labels improve the model's future predictions.
This design directly addresses a well-known limitation of purely computational drug discovery pipelines — predictions may rank candidates well relative to each other, but without experimental calibration, absolute accuracy on real biological systems can drift significantly.
The parties plan to co-design research projects, jointly pursue government and private R&D funding, and collaborate on regulatory strategy and technology transfer. A specific shared research program on ovarian aging therapeutics is also included in the agreement, with detailed scope and timelines to be defined in a separate binding contract.
Important note for developers: The MOU itself is non-binding (except for confidentiality and IP clauses) and represents a statement of intent. The technical integration timeline and deliverables will be governed by subsequent agreements.
Benchmarks & results
No quantitative model performance metrics (e.g., AUC, enrichment factor, hit rate) are disclosed in this announcement. What the source does confirm publicly:
- 13,067 candidate molecules submitted to ODC as of September 6, 2026, across four disease indications
- ODC scoring covers: predicted efficacy, selectivity, safety, novelty, and synthetic accessibility
- VIDRAFT was also selected for the South Korean Ministry of SMEs and Startups' "2026 Public-Private Open Innovation Support Program — Everyone's Challenge Bio", under which it is co-validating AI-discovered drug candidates with domestic pharmaceutical companies
The ODC platform's outputs are explicitly described as research prioritization scores, not experimentally confirmed activity data — a distinction the collaboration with Pyo Biotech is designed to address.
How to try it
The Open Discovery Challenge (ODC) is described as a publicly accessible platform where participants can submit AI-designed molecular structures for evaluation. Developer access details for ODC are not specified in this announcement. No Hugging Face model weights, GitHub repositories, API endpoints, or SDK packages for PharmaOS are referenced in the source article.
If you are interested in participating in ODC or accessing PharmaOS capabilities, monitor VIDRAFT's official channels for public access announcements.
FAQ
Q: Is PharmaOS a generative molecular design model, a scoring/property-prediction model, or both?
A: Based on the source, PharmaOS functions primarily as a scoring and prioritization system — it evaluates molecular binding likelihood, drug-likeness, and toxicity risk for candidate structures. The ODC framing suggests that candidate generation is left to participants, while PharmaOS handles downstream evaluation. Whether VIDRAFT also offers generative capabilities within PharmaOS is not stated in this article.
Q: How does this predict–validate feedback loop differ architecturally from standard active learning in ML?
A: Conceptually it follows the same pattern: a model generates candidates, an oracle (here, a cell assay rather than a human annotator or cheaper proxy) labels them, and those labels improve future model iterations. The pharmaceutical context adds complexity — wet-lab cycles are slow and expensive, making candidate selection at the in silico stage especially high-stakes. The MoA data from Pyo Biotech also provides richer signal than a simple binary hit/miss label, potentially enabling more informative model updates.
Q: What disease areas does the ODC currently cover?
A: As of the publication date, four disease tracks are active: malaria, tuberculosis, Chagas disease, and non-opioid analgesia.
Q: Is VIDRAFT's work here limited to small molecules?
A: The article focuses on small-molecule drug candidates evaluated for target binding, drug-likeness, and toxicity. No mention is made of biologics, peptides, or other modalities in this specific collaboration.
Originally reported by IT조선 (2026-09-09) — source article.
Top comments (0)