microsoft/bioemu

Inference code for scalable emulation of protein equilibrium ensembles with generative deep learning

What it solves

BioEmu is designed to sample the equilibrium distribution of protein structures for a protein monomer, given only its amino acid sequence. It addresses the challenge of generating physically plausible protein conformations that represent the structural ensemble of a protein rather than just a single static structure.

How it works

The project uses a generative deep learning model that samples from an approximated equilibrium distribution. It integrates with ColabFold and AlphaFold2 for MSA (Multiple Sequence Alignment) retrieval and embedding generation. To ensure physical realism, it employs a steering system that applies potential energy functions during the denoising process to avoid common artifacts like steric clashes and chain breaks. This steering is implemented via two algorithms: Sequential Monte Carlo (SMC) and the Feynman–Kac Corrector (FKC).

Who it’s for

This tool is intended for researchers in structural biology, biochemistry, and computational protein science who need to generate ensembles of protein structures from sequences.

Highlights

  • Physical Steering: Uses potential energy functions to guide the diffusion process toward physically valid structures.
  • Integrated Pipeline: Bundles ColabFold and AlphaFold2 for seamless MSA and embedding generation.
  • enables Ensemble Sampling: Focuses on the equilibrium distribution of structures rather than a single prediction.
  • Post-processing Tools: Includes interfaces for side-chain reconstruction (via HPacker) and molecular dynamics (MD) equilibration.

Related

  • Project
    NVIDIA-BioNeMo/Proteina-Complexa

    A generative model for atomistic protein binder design that unifies generative modeling and hallucination to create binders for proteins and small-molecule ligands.

  • Project
    RosettaCommons/foundry

    A tooling and infrastructure framework for protein design, providing a unified interface for generative models like RFdiffusion3 and structure prediction models like RosettaFold3.

  • Dispatch
    Implicit Generation and Generalization Methods for Energy-Based Models

    OpenAI researchers developed stable and scalable training methods for energy-based models (EBMs), achieving high-quality image generation and superior out-of-distribution generalization compared to likelihood-based models.

  • Project
    OTeam-AI4S/ODesign

    An all-atom generative world model for biomolecular interaction design that generates proteins, ligands, and nucleic acids tailored to bind to specific targets.

  • Project
    aurekaresearch/OpenDDE

    An all-atom biomolecular foundation model that uses co-folding for scalable structure prediction, design, and optimization in drug discovery.