MoleculeOS maker Molecule Heart's QuantaMind simulates 100k-atom chemical reactions on one GPU
量子位 · wechat · 2026-09-15
Molecule Heart (分子之心), founded by protein-folding pioneer Xu Jinbo, published QuantaMind, a reactive machine-learning force field in Science Advances that breaks molecular simulation's accuracy-time-scale trilemma.
Method: (1) train on key transition states from 5,286 chemical reactions so the model sees bonds breaking/forming; (2) fuse data sources of varying precision with quality weighting; (3) iteratively flag uncertain conformations for targeted high-level calculations.
Results:
- Matches experiment on proton hopping, water autoionization, and acid-base neutralization without task-specific training
- Protein titration pKa estimate (5.81) close to NMR value (5.5)
- Headline: a 6ns continuous simulation of the full 17,792-atom PETase system with no preset reaction pathway, running the complete catalytic cycle; energies match experiment and quantum-level spot checks exceed 0.99 agreement
Latest: scale now pushed to 100,000 atoms on a single GPU at 0.25s per step, stable over hundreds of nanoseconds.
Applied: candidate antibodies designed with QuantaMind showed 62x faster dissociation at pH 6.0 vs pH 7.4; the tool has also mapped enzyme mechanisms to guide mutagenesis. The company raised $100M+ Series A in June 2026 and launched MoleculeOS in July.
More from Research
- Domingos: An attention head is just a fast-weight MLP — pmddomingos · 2026-10-24
- Ross Taylor: studying LLM behavior will increasingly resemble social science — rosstaylor90 · 2026-10-24
- Ross Taylor on RLCR: simplicity is the final achievement, details matter — rosstaylor90 · 2026-10-24
- SenseTime open-sources 27B robot model matching GPT-6-Astra at ~1/50 the cost — rohanpaul_ai · 2026-10-24
- Cohere Labs brings five main track papers to EMNLP 2026 in Budapest — Cohere_Labs · 2026-10-24
- Gradient clipping stress test: unclipped RNN diverges in 17 steps, clipped runs converge to 0.043 loss — cneuralnetwork · 2026-10-24