Loading…
Blind peptide binder design from a protein sequence alone, evolving linear or cyclic binders against a frozen AlphaFold2 with no binding site given.
No providers recorded yet. Browse all providers
Every binder design pipeline must answer three questions before it starts: what the target looks like, which patch of it to bind, and how large the binder should be. Scaffold routes such as RFdiffusion followed by ProteinMPNN take the first from a structure and demand the other two from the user. For a genuinely novel target all three are unknown.
EvoBind2 removes all three. Its only input is the target's amino acid sequence. A Gumbel-initialised random peptide is then evolved: at each of 1,000 iterations one residue is mutated, a modified single-chain AlphaFold 2 predicts the receptor-peptide complex, and the mutation is kept only if a loss — inverse peptide pLDDT times the mean shortest distance from each peptide atom to the target — improves. Nothing is trained; the released AlphaFold2 parameters serve unmodified as a fitness function, and because the complex is re-predicted from scratch each step, the receptor interface reshapes around whatever peptide the trajectory holds. A cyclic offset in the positional encoding makes the same loop emit head-to-tail macrocycles. A second, orthogonal pass with AlphaFold-Multimer then discards adversarial designs: sequences that score well under EvoBind2's own loss but do not bind.
EvoBind2 comes from Qiuzhen Li, Efstathios Nikolaos Vlachos and Patrick Bryant at SciLifeLab and Stockholm University, posted to bioRxiv in June 2024, revised as v2 that October and published in Communications Chemistry in July 2025. It supersedes the lab's EvoBind, which required the user to name interface residues and fix the peptide length; siblings EvoBind-multimer and RareFold carry the same mutate-and-score idea to ternary complexes and to noncanonical amino acids.
The design network is the single-chain AlphaFold2 folding pipeline run with model_1, one ensemble and eight recycles. The receptor is represented by an HHblits 3.1.0 alignment against UniClust30_2018_08; the peptide is a single sequence with no alignment. Optimisation runs 1,000 mutation rounds, repeated five times from independent random starts. The adversarial check uses AlphaFold-Multimer v2.1.0 with dropout and 20 recycles under early stopping, scored by a second loss adding a centre-of-mass term. Survivors pass a solubility filter: mean pLDDT above 90, more than 25% charged residues, fewer than 25% hydrophobic.
Validation targeted a semi-synthetic ribonuclease (PDB 1SSC), against which designs of 8-20 residues were generated and the best sequence per length synthesised. Five of 13 linear designs (38%) bound at µM or better affinity by surface plasmon resonance, the strongest at Kd 7.5 nM — 162-fold tighter than the 1.2 µM native 11-mer positive control. One of four cyclic designs (25%) bound, at 6.48 µM, near the control. Among in-silico metrics the AlphaFold-Multimer loss tracked measured Kd best (Spearman R = 0.96) and EvoBind2's own loss barely at all (R = −0.14); designs drawn deliberately from the adversarial region bound in only 2 of 13 cases, a 2.5-fold hit to success rate. Designed binders shared at most 32% of their receptor contacts (21 of 65) with the native peptide.
The practical case is a target with no ligand, no co-crystal structure and no characterised pocket, where scaffold- and motif-based design has nothing to work from. Cyclic output matters here: macrocycles are the route to intracellular targets. The protocol has already produced a single-shot cyclic peptide inhibitor of HIV membrane fusion and blindly designed dual cyclic agonists of GCGR and GLP1R.
EvoBind2 makes the case that the binding site is itself designable rather than a required input, and that a frozen predictor plus an orthogonal second opinion suffices to reach nanomolar affinity without a scaffold library. The evidence base is narrow and the paper says so: one target protein, expressed as an MBP fusion the authors acknowledge may affect the SPR responses, and 17 synthesised peptides. No in-silico metric separated true from false binders among the evaluated designs, and the solubility criteria did not correlate with binding success, so wet-lab triage remains unavoidable. The code inherits AlphaFold2's Apache 2.0 licence and the parameters are used unmodified under CC BY 4.0, but the EvoBind2 design protocol carries CC BY-NC 4.0, excluding commercial use. It shares a GitHub repository with its predecessor, whose code line is preserved under the v1 tag while the default branch distributes this protocol.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.