Use EvoPro Online

EvoPro: Genetic Algorithm Binder Design with AlphaFold2 and ProteinMPNN

EvoPro is a genetic algorithm pipeline that evolves high-affinity protein binders by alternating deep learning structure prediction with sequence diversification. It was developed in Brian Kuhlman's laboratory at the University of North Carolina School of Medicine and described in PNAS in 2023 by Odessa J. Goudy, Amrita Nallathambi, Tomoaki Kinjo, Nicholas Z. Randolph and Brian Kuhlman. Scripts and input files are available on GitHub.

Computational protein design has advanced steadily, but traditional physics-based workflows usually require screening hundreds or thousands of designs and then experimental affinity maturation to reach dissociation constants below 100 nM. EvoPro asks whether AlphaFold2 and ProteinMPNN can close that gap in silico.

The authors tested the approach by designing autoinhibitory domains (AiDs) that mask a PD-L1 antagonist until tumor-enriched proteases cleave a linker, a route to an anticancer agent that is inactive until it reaches the tumor.

How EvoPro Evolves Protein Binders with AlphaFold2 and ProteinMPNN

Each iteration scores a pool of sequences, discards the worse-scoring half, and refills the pool with new variants.

1. Predict: every sequence is run through a runtime-optimized AlphaFold2 as a monomer and as a complex with the target.

2. Score: a custom function sums placement confidence (interface contacts weighted by predicted aligned error), fold confidence (pLDDT of the binder) and conformational stability (how little the binder changes on binding).

3. Select and diversify: the best half is kept; new sequences come from random point mutations, crossovers, or ProteinMPNN run on the predicted complexes.

4. Repeat: after the user-defined number of iterations, the final pool and its AlphaFold2 results are written out.

Benchmarking insights

Small sequence pools often got stuck in local minima, while larger pools gave better scores at higher compute cost. Using ProteinMPNN in the refill step lowered the minimum score reached, and later benchmarking showed that using it more often than every 10 iterations often lowers scores further. The score function is customizable for other design goals.

Results: Designing PD-L1 Antagonist Autoinhibitory Domains Without Affinity Maturation

Starting from 51 de novo miniprotein scaffolds, the authors ran five trajectories per scaffold (pool of 50 sequences, 60 iterations), producing 6,375 designs. After filtering with Rosetta metrics and OmegaFold cross-checks, 23 designs were tested experimentally.

Protease-treated masked antagonists generally regained PD-L1 binding with KDs below 10 nM by SPR, and constructs with an MMP-2 cleavage site retained protease-dependent activity. Alanine mutations of predicted interface residues raised KD more than 10-fold in the three designs tested, supporting the predicted binding modes.

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 EvoPro on Tamarind Bio

• Binder design without affinity maturation: evolve miniprotein binders against a target of interest.

• Conditionally active therapeutics: design autoinhibitory masks cleaved by tumor proteases such as TEV-site or MMP-2 linkers.

• Custom objectives: adapt the score function to other design requirements.

• Competitive inhibitors: target known protein-protein interaction sites, where the authors found EvoPro works best.

How to Use EvoPro on Tamarind Bio

1. Log in and open the tool: sign in at tamarind.bio and select EvoPro.

2. Provide the target: supply the target protein sequence or structure, as the interface defines.

3. Provide starting binders: choose scaffold or starting sequences, such as miniprotein scaffolds, if exposed.

4. Set run parameters: adjust pool size and number of iterations if exposed; the paper used 50 sequences over 60 iterations.

5. Run the job: launch it and wait for the evolution to finish.

6. Download results: retrieve the final sequence pool and predicted complexes.

7. Rank and filter: sort by ipTM or interface energy, and cross-check with an independent predictor.

8. Validate experimentally: express and test top candidates.

Things to Keep in Mind

• Designs are predictions; AF2 confidence and Rosetta energies could not predict that AiD5 would bind more than 40-fold tighter than the others.

• EvoPro tends to place binders at known interaction sites and struggles to target other regions of a target surface.

• Each trajectory converges to highly similar sequences (about 81% average identity), so run many independent trajectories.

• The tightest binder was not the best switch, since AiD5 stayed bound after cleavage.

Source: Goudy O.J., Nallathambi A., Kinjo T., Randolph N.Z., Kuhlman B. "In silico evolution of autoinhibitory domains for a PD-L1 antagonist using deep learning models." PNAS 2023, 120(49), e2307371120. DOI: 10.1073/pnas.2307371120. Code: github.com/Kuhlman-Lab/evopro.