Integrated Machine-learning for PRotEin Structures at Scale (IMPRESS)

IMPRESS is a high-performance computational framework to enable the inverse design of proteins using Foundation Models such as AlphaFold and ESM2.

IMPRESS pipeline design

• The user submits the C-terminal peptide sequence, which they want to target for binding.
• The ProfAff database is queried to generate design starting points.
• ProteinMPNN is used to create a design sequence.
• The designed sequences are submitted to AlphaFold-Multimer.
• AlphaFold-predicted structures design structures are fed to a structure-aware machine learning model to predict binding specificity and affinity. The cycle continues until convergence.
• Best designs are tested in wet-lab experiments. Measured and predicted properties are used as feedback for model refinement and, if successful, added to the input set to serve as starting templates.

Current implementation is capable of running ProteinMPNN on an input two-chain complex to redesign the entire receptor, creating new interactions to the substrate and solubilizing the protein. These designs are then submitted to AlphaFold-Multimer for structure prediction, where AlphaFold’s intrinsic discriminatory power and confidence metrics can be leveraged to determine if the input protein has improved. The predicted structures are then resubmitted to ProteinMPNN to continue optimizing the receptor. Over the course of many iterations, we expect the receptor to gradually improve, both in terms of stability and substrate binding affinity. The pipeline monitors this process, distributing the sequence generation, structure determination, and design analysis tasks evenly across the specified GPU and CPU allocations.

Middleware foundation

RADICAL-EnTK (Ensemble Toolkit) is used as a workflow management system, that enables the use of high-performance computing (HPC) platforms and features. EnTK-based designed pipeline provides the integrated ability to efficiently and effectively train sophisticated models. That requires advances in HPC workflow methodology that brings together the ability to “evaluate as you go”.

Resources

To learn more, please visit the project website at https://radical-project.github.io/impress/