Microsoft Research AI for Science Releases BioEmu2 Foundation Model for Biomolecular Structure and Dynamics

Microsoft Research AI for Science has published BioEmu2, a foundation model designed for biomolecular structure prediction, equilibrium sampling, and time-lagged trajectory generation. The software package aims to support molecular systems containing proteins, nucleic acids, ligands, and other chemical components.

Core Capabilities and Features

• Structure Prediction and Equilibrium Sampling: Predict folded structures and sample equilibrium ensembles from the command line or using Python interfaces.
• Dynamic Trajectory Generation: Propagate structures forward in time using implicit transfer operator models.
• Steered and Parallel Sampling: Apply contact conditioning, perform inference-time steering toward specific properties, and utilize context-parallel sampling across multiple GPUs for structures that exceed single-device memory limits.

System Requirements and Installation

BioEmu2 is reported to require a Linux operating system and a supported NVIDIA GPU; macOS and Windows environments are not supported.

• GPU Compatibility: The software requires NVIDIA hardware with compute capability 7.0 or newer. Specific installation extras depend on the GPU model, separating requirements for CUDA 13 configurations and older architectures like Volta (compute capability 7.0).
• Python Environment: The model is tested with Python 3.12, while versions 3.13 and 3.14 are noted to pass test suites.

Practical Implications and Verification

Users can interact with BioEmu2 via command-line utilities or Python APIs, and run interactive workflows using Jupytext and Jupyter notebooks or VS Code. Developers interested in verifying the original announcement, reviewing complete documentation, or checking license terms can inspect the official Microsoft repository on GitHub (`microsoft/bioemu2`).