BioEmu · azureml v4

# Key capabilities ## About this model The BioEmu model is intended for prediction of protein equilibrium structure ensembles. The model is used for predicting diverse protein structures that emulate…

azureFoundry/azureml/BioEmu@4

Key capabilities

About this model

The BioEmu model is intended for prediction of protein equilibrium structure ensembles. The model is used for predicting diverse protein structures that emulate the thermodynamic ensemble (i.e., Boltzmann distribution) given an amino acid sequence. An interface to side-chain reconstruction and MD equilibration is provided.

Key model capabilities

  • prediction of structural ensembles
  • prediction of folding free energies
  • providing mechanistic hypotheses
  • sampling of conformational changes related to protein function (specifically local unfolding, domain motion and the formation of cryptic pockets)
  • emulation of molecular dynamics (MD) equilibrium distributions
  • prediction of protein stabilities

Use cases

See Responsible AI for additional considerations for responsible use.

Key use cases

The model is used for predicting diverse protein structures that emulate the thermodynamic ensemble (i.e., Boltzmann distribution) given an amino acid sequence. Examples of direct usages include but not limited to prediction of structural ensembles, prediction of folding free energies, providing mechanistic hypotheses. The model is intended for research and experimental purposes.

Out of scope use cases

The model only supports predictions of protein structures in backbone frame representation. Any attempt and interpretation beyond that should be avoided. The model does not support generation of new protein sequences as it is designed for the above purpose only. The current model has low prediction quality for protein-protein interactions, including multi-chain proteins, and does not feature explicit interactions with other chemical entities like small molecules. Further testing/development are needed before considering its application in real-world scenarios. We advice against predicting entities that are not considered by the used embeddings or represented in the training data, including but not limited to multi-chain proteins.

Pricing

Pricing is based on a number of factors, including deployment type and tokens used. See pricing details here.

Technical specs

Developed by: Sarah Lewis, Tim Hempel, José Jiménez-Luna, Michael Gastegger, Yu Xie, Andrew Y. K. Foong, Victor García Satorras, Osama Abdin, Bastiaan S. Veeling, Iryna Zaporozhets, Yaoyi Chen, Soojung Yang, Adam Foster, Arne Schneuing, Jigyasa Nigam, Federico Barbero, Vincent Stimper, Andrew Campbell, Jason Yim, Marten Lienen, Yu Shi, Shuxin Zheng, Hannes Schulz, Usman Munir, Roberto Sordillo, Ryota Tomioka, Cecilia Clementi, Frank Noé

Model type: We only release the model fine-tuned with amber data (BioEmu 1.1).

Model Parameters: 31M

Training Data Size: 161k structures (AFDB), 216 ms MD simulations, 502k dG measurements

Training cut-off date

The provider has not supplied this information.

Training time

The provider has not supplied this information.

Input formats

{
    "input_data": 
        {
            "sequence": "GYDPETGTWG",
            "num_samples": 10
        }
}

Output formats

{
    "status": "success",
    "results": 
        {
            "sequence.fasta": "<binary file content>", 
            "topology.pdb": "<binary file content>", 
            "samples.xtc": "<binary file content>"
        }
}

Supported languages

The provider has not supplied this information.

Sample JSON response

{
    "status": "success",
    "results": 
        {
            "sequence.fasta": "<binary file content>", 
            "topology.pdb": "<binary file content>", 
            "samples.xtc": "<binary file content>"
        }
}

Model architecture

BioEmu-v1 model is DiG architecture (https://www.nature.com/articles/s42256-024-00837-3) trained on a variety of datasets to sample systematically diverse structure ensembles. In the pretraining phase, we use denoising score matching to match the distribution of flexible protein structures curated from AFDB. In the fine-tuning phase, we use a combination of denoising score matching objective for molecular dynamics data and property prediction fine-tuning (PPFT) for matching the experimental folding free energies.

Long context

The provider has not supplied this information.

Optimizing model performance

The provider has not supplied this information.

Additional assets

Training disclosure

Training, testing and validation

Our model has been trained on a large variety of structurally resolved proteins, so it inherits the biases of this data. Besides experimental data, the model is trained on synthetic data, which is predictions of AlphaFold2 and molecular dynamics simulations. We expect that the approximations of these models are propagated to BioEmu. For testing conformational changes, sets of structures exhibiting different phenomena (local unfolding, domain motion and formation of cryptic pockets) were curated based on PDB reports and published literature. To test emulation of MD equilibrium distributions, an in-house dataset of molecular dynamics simulation based on the CATH classification of proteins was generated. Protein stability predictions were evaluated using a combination of published experimental folding free energies.

Distribution

Distribution channels

The provider has not supplied this information.

More information

Please refer to the BioEmu manuscript for more details on the model.

License: MIT License

Funded by: Microsoft Research AI for Science

Specifications & capabilities

Input
text
Output
text
Max output · tokens
4096

Provider specifications & access

1–1 / 1

BioEmu · azureml v4 ↗

source model id
azureml/BioEmu@4
hosting
foundry-managed-compute
platforms
  • Microsoft Foundry
license
MIT
max output tokens
4096
architecture
input_modalities
  • text
output_modalities
  • text
capabilities

    Source specification observations

    1–1 / 1

    azureFoundry ↗

    field
    max_output_tokens
    value
    4096
    normalized value
    4096
    scope
    source-model
    source id
    azureFoundry
    confidence
    1

    Original sources

    1–3 / 3

    Microsoft Foundry · Model page ↗

    Microsoft Foundry · Model documentation ↗

    Microsoft Foundry · Catalog ↗

    Source links

    1–1 / 1

    Explore more models

    Models

    Loading
    EVIDENCE / COMPARISON

    Benchmarks

    Explore every benchmark

    Loading

    These charts contain model-linked results. The charts above also include unlinked source profiles. Compare scores within each evaluation. Model quality, usage, transparency and system performance measure different things.

    Install TensorTrace