EvoPro

What is EvoPro?

EvoPro is a genetic algorithm-based pipeline for designing protein binders through in silico evolution. Developed by the Kuhlman Lab at the University of North Carolina, it combines iterative structure prediction with AlphaFold2 and sequence design with ProteinMPNN to evolve protein sequences that bind tightly to a target protein.

The approach differs from traditional computational design methods by allowing backbone plasticity during optimization. As sequences evolve across generations, their predicted structures can undergo conformational changes favorable for binding—something difficult to encode in physics-based design methods like Rosetta.

In published work, EvoPro generated autoinhibitory domains for a PD-L1 antagonist, with four designs achieving sub-150 nM binding affinity and the best reaching 0.9 nM without any experimental optimization.

How does EvoPro work?

EvoPro runs a genetic algorithm that maintains a population of candidate binder sequences and evolves them through repeated cycles:

1. Scoring: Each sequence is evaluated by predicting its structure bound to the target using AlphaFold-Multimer
2. Selection: Sequences are ranked by fitness; the bottom half is discarded
3. Diversification: New sequences fill the population through mutation, crossover, or ProteinMPNN redesign
4. Iteration: The cycle repeats for the specified number of generations

Fitness function

The fitness score combines three components from AlphaFold2 predictions:

Component	What it measures
Placement confidence	Interface quality based on sidechain contacts weighted by PAE (predicted aligned error)
Fold confidence	Binder stability from average pLDDT across the designed protein
Conformational stability	RMSD between bound and unbound structures to minimize binding-induced changes

Lower scores indicate better designs. The conformational stability term encourages rigid binders with fast association kinetics.

Sequence diversification

New sequences are generated through two strategies:

• Random mutagenesis + crossover: Introduces point mutations (~12.5% of residues) and recombines sequences from surviving parents
• ProteinMPNN optimization: Redesigns sequences from AlphaFold-predicted backbones during the evolutionary cycle when enabled

How to use EvoPro online

ProteinIQ provides GPU-accelerated EvoPro runs without local installation, making binder design accessible through a browser interface.

Inputs

Input	Description
Target Protein	PDB file, RCSB PDB ID, raw amino-acid sequence, or FASTA sequence for the target used in AlphaFold2 co-folding. For multi-chain target sequences, separate chains with : or provide multiple FASTA records; these become chains B, C, and so on
Starting Scaffold	Required starting binder structure or FASTA sequence. EvoPro evolves this single binder chain as chain A

Evolution parameters

Setting	Range	Default	Description
Population size	4-100	20	Candidate pool size. Larger pools explore more diversity but increase runtime
Number of generations	5–200	50	Evolutionary cycles. More generations improve optimization at the cost of time
Mutation rate	0.05-0.5	0.125	Fraction of mutable scaffold residues changed during random mutagenesis

Design settings

Setting	Description
Use MSA	Generate AF2 multiple sequence alignments instead of single-sequence predictions. This can substantially increase runtime and is usually unnecessary for de novo binder optimization
Mutable residues	Restrict which binder positions can mutate. The binder is chain A. Use A*, a range such as A1-A65, or positions and ranges like A5,A10,A20-A30. Leave empty to allow all binder positions
Target contact residues	Optional target residues the binder should contact, rewarded in the EvoPro contact score. Use target chain IDs such as B34 or B30-B40,B55
Target avoid residues	Optional target residues the binder should avoid, penalized in the EvoPro contact score. Use target chain IDs such as B6 or B10-B15
Target chain	Select one chain from the target PDB. Leave empty to use the first protein chain
Scaffold chain	Select one chain from a scaffold PDB. Leave empty to use the first protein chain; FASTA scaffolds are assigned binder chain A
Enable ProteinMPNN	Toggle ProteinMPNN sequence design during evolution (recommended)

Runtime scaling

EvoPro runtime scales with the number and size of AlphaFold2 evaluations:

Driver	Effect
Target residues	Larger targets increase each bound-complex prediction
Scaffold residues	Larger scaffolds increase both bound-complex and scaffold-only predictions
Population size	Multiplies the number of sequences evaluated per generation
Number of generations	Multiplies the total evolutionary evaluations
ProteinMPNN	Adds sequence redesign work during the evolutionary cycle
MSA generation	Adds AF2 sequence-search work and can make each prediction much slower

As a rough guide, EvoPro evaluates each candidate in both a bound target-scaffold state and a scaffold-only state, so AlphaFold2 prediction count is approximately population size x number of generations x 2.

Results

EvoPro returns a ranked list of designed binders with:

Column	Description
Rank	Position in the ranked output (1 = best)
Binding Score	EvoPro composite fitness score label and value from the score table (lower = better within the same run)
Sequence	Designed amino acid sequence
Download	PDB file of the predicted complex

The 3D viewer displays selected designs bound to the target protein. The files tab includes ranked PDB files, the full EvoPro score table, confidence plots when available, and a compressed archive of the raw result directory.

Interpreting binding scores

EvoPro scores depend on the target, scaffold, score function, and generated AlphaFold2 predictions. Treat the rank order and score differences within one job as the primary signal. Do not compare raw score values across unrelated targets or scaffolds as absolute binding-affinity thresholds.

Limitations

• Runtime: Full evolutionary trajectories can range from a few minutes for tiny smoke tests to much longer jobs as target size, scaffold size, population size, and generation count increase
• Target size: Very large target proteins increase AlphaFold prediction time per generation
• No explicit binding energy: Fitness scores correlate with but don't directly predict experimental binding affinity
• Single binding mode: The algorithm optimizes for one binding interface; alternative binding sites aren't explored