Use AlphaFlow Online

AlphaFlow: Generating Protein Conformational Ensembles with Flow Matching

AlphaFlow (AlphaFLOW) is a sequence-conditioned generative model that samples ensembles of protein conformations. It was developed by Bowen Jing, Bonnie Berger and Tommi Jaakkola at MIT CSAIL and presented at ICML 2024. The method repurposes the AlphaFold and ESMFold structure predictors as flow-matching generators; the ESMFold-based variant is called ESMFlow. Code is available on GitHub.

Proteins are not static. Function often depends on a landscape of conformations, but AlphaFold and similar tools return a single structure. Molecular dynamics (MD) can sample alternatives, though at a high computational cost, and tricks such as subsampling the multiple sequence alignment (MSA) offer limited control over the balance between precision and diversity.

How AlphaFlow Turns AlphaFold into a Flow-Matching Ensemble Generator

1. Denoiser: a small input embedding module, including a C-beta distogram, is attached to the AlphaFold or ESMFold trunk.

2. Harmonic prior: sampling starts from a polymer-like prior over C-beta coordinates and moves toward the data along a linear path.

3. Fine-tuning: the models are fine-tuned from public weights on the PDB, and AlphaFlow-MD is further trained on all-atom MD ensembles from the ATLAS dataset.

4. Sampling and distillation: ten steps by default; fewer steps trade diversity for precision, and distillation compresses sampling into one forward pass.

AlphaFlow Results: Precision, Diversity and MD-Like Flexibility

PDB ensembles (median over 100 targets)

AlphaFlow reaches higher diversity at a given precision than MSA subsampling. On 82 ATLAS test proteins, AlphaFlow-MD reproduced a median per-target RMSF correlation with MD of 0.85, versus 0.52 for MSA subsampling and AlphaFold. With templates, per-sample runtime was about 38 seconds, or 3.8 seconds distilled, and quality held up with roughly ten times fewer samples than replicate MD.

What is Tamarind Bio?

Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.

The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:

• A user-friendly graphical interface for setting up and launching experiments

• A robust API for integration into existing research pipelines

• An automated system for managing and scaling computational resources

Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.

Accelerating Discovery with AlphaFlow on Tamarind Bio

• Flexibility analysis: estimate per-residue fluctuations and positional distributions without running long MD simulations.

• Alternative states: diversify a static PDB structure to explore other plausible conformations.

• Contact and exposure analysis: examine transient contacts and exposed residues across sampled ensembles.

• Faster equilibrium properties: approximate some ensemble observables in less wall-clock time than replicate MD.

How to Use AlphaFlow on Tamarind Bio

1. Open the tool: log in to tamarind.bio and open AlphaFlow.

2. Enter the sequence: provide the protein amino acid sequence.

3. Choose a variant: select the PDB-trained or MD-trained model, or a template-conditioned option, if exposed.

4. Set sampling options: choose the number of samples and sampling steps, if exposed.

5. Run the job: submit and wait for sampling to complete.

6. Download outputs: retrieve the ensemble of sampled structures.

7. Interpret: compare conformations, compute RMSF or clustering, and inspect regions that vary.

8. Use downstream: feed selected conformations into docking, design or further simulation.

Things to Keep in Mind

• Evaluation is limited to phenomena within ATLAS simulation timescales; slower conformational changes were not assessed.

• Neither AlphaFlow nor ESMFlow meaningfully improves aggregate recall over baseline AlphaFold.

• ESMFlow has lower precision than the AlphaFold-based models, and distilled models trade some quality for speed.

• AlphaFlow depends on MSAs, and test sets are small (100 PDB and 82 ATLAS proteins) with MD as the reference.

• Cryo-EM heterogeneity has not been tested.

Source: Jing, B., Berger, B. and Jaakkola, T., AlphaFold Meets Flow Matching for Generating Protein Ensembles. Proceedings of the 41st International Conference on Machine Learning (ICML), PMLR 235, 2024. arXiv:2402.04845. Code: https://github.com/bjing2016/alphaflow