MoleculeOS launches as AI operating system for biology R&D, signaling platform shift in drug discovery
The AMW Read
Novelty 2: platformization of AI for biology is an emerging pattern but not yet canonical; significance 2: if validated, this could reshape R&D infrastructure for pharma and biotech, a segment-level shift.
MoleculeOS launches as AI operating system for biology R&D, signaling platform shift in drug discovery
Professor Xu Jinbo's team at MoleculeOS (分子之心) has publicly launched MoleculeOS (MOS), an AI-native operating system for biological R&D, at the 2026 Shanghai State Capital Frontier Forum. The system integrates the company's proprietary model stack — including the NewOrigin (Darwin) multimodal protein foundation model, the MMFold all-atom macromolecule structure prediction model, and the MMDesign generative design model — into a unified platform that takes research intent as input, automatically decomposes tasks, orchestrates model execution, and produces structured, traceable outputs. MMFold achieved 68.6% prediction success on 172 antibody-antigen interfaces in the FoldBench benchmark, significantly outperforming AlphaFold3. The company's antibody de novo design platform achieved over 90% target success rate across 12 targets while testing no more than 50 candidates per target.
Why it matters: This launch exemplifies the platformization pattern in AI-driven drug discovery — the shift from point-solution tools to integrated operating systems that orchestrate the full R&D workflow. MoleculeOS is positioning itself as the infrastructure layer for AI-native biology, analogous to what Cursor and Copilot did for AI coding: moving from single-task assistance to workflow-level orchestration. The company claims to compress multi-week, multi-tool antibody optimization workflows into hours, which if validated would represent a step-change in R&D velocity for biotech and pharma. The launch also updates the competitive landscape in AI for biology, where the race is increasingly about system-level integration rather than individual model benchmarks.