Microsoft Open-Sources BioEmu2, Claiming 100,000x Speedup in Biomolecular Dynamics

Frank Noé's team (Free University of Berlin, with a Microsoft Research AI for Science background) has released BioEmu2, a biomolecular foundation model focused on dynamics prediction: a 5-millisecond protein simulation that takes traditional molecular dynamics (MD) about 10 GPU years now takes BioEmu2 roughly 40 minutes — an over 100,000x speedup. The team calls it the only SOTA biomolecular foundation model capable of predicting dynamics.

Confirmed

• Dynamics prediction: built on implicit transfer operators, it can propagate known or predicted protein structures forward to nanosecond and microsecond timescales, or push directly to the equilibrium distribution
• Equilibrium sampling accuracy: free-energy surface prediction error of about 1 kcal/mol, reproducing in GPU minutes the structural distributions that would take GPU years of MD
• Functional cases: simulating the folding of an intrinsically disordered peptide and its binding to a protein receptor (via intermediates); adenylate kinase sampling both open and closed conformations, with the main pocket closing upon ATP binding and both pockets closing with added AMP; a weak binder (2r58, KD=1 mM) showing a broad, divergent ensemble of binding poses, while a strong binder (2am9, IC50=3.9 nM) settles stably into the pocket
• Large-system capability: multi-GPU prediction of Mega-Dalton-scale (10,000+ residue) protein/RNA/DNA/small-molecule complexes, with the ability to incorporate experimental data to aid structure prediction of large complexes
• Structure prediction benchmarks: competitive with Protenix v1/v2 and ESMFold2 on the high-resolution protein–ligand complex benchmark Runs N' Poses

Why it matters

Noé also published a roadmap for scalable molecular biology: sequencing and structure prediction have been "solved" by Nobel-level work (DNA sequencing in 1980, next-generation sequencing in 2022, protein structure prediction in 2024), but the third pillar — biomolecular function prediction — remains unsolved. BioEmu2 unifies structure prediction with dynamics and equilibrium sampling, aiming squarely at that gap; if the accuracy and speed claims hold, computational costs for tasks like distinguishing strong vs. weak binders in drug design and studying large-complex function could drop dramatically.

2026-10-20 ~ 2026-10-20 · 15 related posts

• Sequencing and structure prediction are solved; biomolecular function prediction remains the frontier — FrankNoeBerlin · 2026-10-20
• BioEmu2 matches SOTA folding models on Runs N' Poses and FoldBench — FrankNoeBerlin · 2026-10-20
• BioEmu2 handles Mega-Dalton complexes with 10k+ residues on multiple GPUs — FrankNoeBerlin · 2026-10-20
• BioEmu2 integrates experimental data for large-complex structure prediction — FrankNoeBerlin · 2026-10-20
• BioEmu2 predicts free-energy surfaces within ~1 kcal/mol in GPU-minutes — FrankNoeBerlin · 2026-10-20
• BioEmu2 separates weak vs strong binders by pose ensemble convergence — FrankNoeBerlin · 2026-10-20
• BioEmu2 samples adenylate kinase open/closed states driving ATP catalysis — FrankNoeBerlin · 2026-10-20
• [source] BioEmu2 propagates structures in time via implicit transfer operators — FrankNoeBerlin · 2026-10-20
• [source] BioEmu2 compresses a 5 ms protein simulation from 10 GPU-years to 40 minutes — FrankNoeBerlin · 2026-10-20
• Microsoft open-sources BioEmu2, a foundation model for biomolecular structure and dynamics — mmbronstein · 2026-10-20
• Microsoft Research releases BioEmu2, an open-weight MIT-licensed foundation model for biomolecular dynamics — FrankNoeBerlin · 2026-10-20

4 near-duplicate retellings: FrankNoeBerlin · FrankNoeBerlin · FrankNoeBerlin · FrankNoeBerlin