Loading…
In silico directed evolution that designs peptide binders against a chosen protein interface from sequence alone, scored by a frozen AlphaFold2.
No providers recorded yet. Browse all providers
Designing a peptide against a specific patch of a protein surface used to require a structure to start from. Scaffold-based pipelines graft a known motif onto an existing backbone or redesign the sequence of one — the route ProteinMPNN enabled — so the reachable designs are bounded by the backbones already in hand. Phage display escapes that bound experimentally, but returns binders whose binding site stays unknown until someone solves the complex.
EvoBind removes the scaffold. Given the receptor's sequence, the residues to target and a desired peptide length, it starts from a random peptide sequence and evolves it in silico: each step mutates one amino acid, re-predicts the receptor–peptide complex with AlphaFold 2, and keeps the change only if a composite score improves. No weights are trained or fine-tuned at any point — the released AlphaFold2 parameters serve unmodified as a fitness function, and the whole search happens in peptide sequence space. Because the complex is re-predicted from scratch each iteration, the receptor interface rearranges around whatever peptide the search currently holds, rather than the peptide being docked into a fixed holo conformation.
The protocol came from Patrick Bryant and Arne Elofsson at Stockholm University and SciLifeLab, posted to bioRxiv in July 2022. It founds a design line the lab has extended repeatedly since: EvoBind-multimer carries the same mutate-and-score loop to macrocyclic peptides bridging two proteins, and RareFold extends binder design to noncanonical amino acids.
The predictor is AlphaFold2 v2.0 adapted per the FoldDock protocol, run as a single ensemble of model_1 with the receptor represented by an HHblits 3.1.0 alignment against uniclust30_2018_08 and the peptide by a single sequence. A sweep of one to ten recycles fixed eight as the operating point: on 96 non-redundant protein–peptide complexes, 12.5% were predicted within 2 Å interface RMSD at eight recycles against 6.25% at one, and peptide pLDDT separates accurate predictions from the rest at ROC AUC 0.94. A run starts from a Gumbel-initialised random sequence and performs 1,000 single-residue mutations, the loss typically converging before iteration 500 — roughly 46 seconds per iteration on a single 40 GB NVIDIA A100, or near 13 hours per trajectory.
Across the 12 complexes the predictor places accurately, designs reached a median contact similarity of 56% with the native interface, a median peptide pLDDT of 70 and a median Cα RMSD of 4.3 Å; binding-motif recovery spanned 3% to 75%. Used as a filter rather than a generator, the same loss ranked 5,578 experimentally measured miniprotein sequences against FGFR2, TrkA, IL7Ra and VirB8, selecting 47% of true binders at a 20% false-positive rate.
The practical use is producing starting sequences for wet-lab optimisation against interfaces with no known ligand and no co-crystal structure — the situation for most protein–protein interactions considered undruggable by small molecules. The same score doubles as a triage filter on libraries of already-designed binders, cutting the constructs that need expression and assay. The authors position the platform as a front end for laboratory work, not a source of final drug candidates: solubility, aggregation propensity and immunogenicity are outside what it sees.
EvoBind established a pattern now common in biomolecular design: hold a structure predictor fixed and evolve the sequence against its confidence rather than training a generator. Its descendants and independent binder-design efforts follow the same shape. The paper is candid about the ceiling this inherits from the predictor — AlphaFold2 placed only 12 of 96 native peptides within 2 Å, and across 234 peptides against nine peptide-recognition modules its confidence tracked measured ELISA ratios only weakly, so relative affinity could not be ranked. Optimising pLDDT alone recovered just one of two known binding sites on PDB entry 2z9i, which is why interface residues must be specified rather than discovered. The 2022 preprint was never revised or published in a journal. Its code is an AlphaFold2 derivative under the Apache 2.0 licence, with the AlphaFold2 parameters used unmodified under CC BY 4.0; the shared GitHub repository's default branch now distributes the successor protocol EvoBind2, and the code line described here is preserved under the repository's v1 tag.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.