AI Folding

OVERVIEW

AI Folding is a tool in Bench that provides a common interface to AI-based protein structure prediction models. Protein structure prediction has been revolutionized by the power of AI, trained on massive experimental datasets like the Protein Data Bank. These models have been benchmarked on existing structures and found to produce atomic-level resolution.

Which model should I use?

AlphaFold2 is the most experimentally validated model for protein prediction. AlphaFold2 Monomer uses evolutionary homology to achieve high confidence structures for monomers. The SingleSeq mode of AlphaFold2 skips evolutionary alignment and is appropriate for de novo sequences which have little homology to proteins in nature.

AlphaFold-Multimer is provided for the prediction of protein-protein interactions since it was trained on multi-chain structure in the Protein Data Bank. AlphaFold-Multimer is available under AlphaFold2 in Multimer mode.

OpenFold has very similar behavior, architecture, and success rates compared to AlphaFold and was developed by a nonprofit organization dedicated to open source protein engineering tools. OpenFold is provided in monomer mode.

ESMFold is a structure prediction model based on Evolutionary Scale Modeling (ESM), a language model of protein sequences. It is provided because it can be 10-fold faster, making it useful for extremely high throughput needs that don’t require advanced accuracy. However, it performs at lower success rates than AlphaFold2 and OpenFold. ESMFold is provided in monomer mode.

RUNNING AI Folding

  1. Begin by opening a project in Bench and selecting the AI Folding button located to the right under Structure Prediction.

  2. Click the AI Folding Action to open a job setup tab. Using the dropdown menu, choose which model is best for your use.

  3. In AlphaFold2 monomer mode, OpenFold, or ESMfold, enter your sequence into the box for monomer structure prediction.

  4. For multiple sequences, select multimer mode and click the plus button to add boxes for additional sequences. Enter each chain sequence in its own box.

  5. In AlphaFold2 singleseq mode, enter your sequence into the box and enter a reference structure from a Bench collection for calculating RMSD.

  6. You can also open a Sequence collection and select one or more sequences, click the Run dropdown menu to open an AI Folding setup tab prefilled with the selected sequence(s).

  7. At the bottom of the setup tab, select the Run button to start the job(s).

  8. The prediction(s) will appear in the Structures panel when it is ready. You can view it in Bench, select it for further Actions or Analysis or download the PDB file or FASTA file.

At Levitate Bio, we are dedicated to providing the best tools for your goals. We are working to implement newer models and look forward to bringing them to Bench soon.

Looking to use the newest structure prediction model? We provide Boltz1, RoseTTAFold All Atom, RoseTTAFold Nucleic Acid, and the models above in our API, Levitate Engine.

Updated: