Protein-Peptide Docking with ESMFold Language Model

Abstract

Designing peptide therapeutics requires precise peptide docking, which remains a challenge. We assessed the ESMFold language model, originally designed for protein structure prediction, for its effectiveness in protein-peptide docking. Various docking strategies, including polyglycine linkers and sampling-enhancing modifications, were explored. The number of acceptable-quality models among top-ranking results is comparable to traditional methods and generally lower than AlphaFold-Multimer or Alphafold 3, though ESMFold surpasses it in some cases. The combination of result quality and computational efficiency underscores ESMFold's potential value as a component in a consensus approach for high-throughput peptide design.

Figure 1

ESMFold docking results with default settings and enhanced sampling variants. (A, B) DockQ vs pLDDT scatter plots for different ESMFold simulation variants: (A) default settings and (B) using a masking approach. Blue dots indicate peptides in contact with the receptor (within 8 Å), while red dots indicate those not in contact (over 8 Å). pLDDT values are the average for each peptide, with (B) showing the weighted mean pLDDT for the top-ranked model out of eight generated per complex. (C) Distribution of high-quality (DockQ ≥ 0.8), medium-quality (0.5 ≤ DockQ < 0.8), and acceptable (0.23 ≤ DockQ < 0.5) docking models for different ESMFold simulation approaches (default, random masking) across 111 complexes, with colors indicating quality.

Figure 2

Comparison of ESMFold and AlphaFold-based docking tools on Dataset 1 (A) and Dataset 2 (B). The upper panels show DockQ scatter plots, comparing ESMFold with (A) AlphaFold Multimer (AFM) using enhanced sampling and (B) AlphaFold 3 (AF3). The lower panels present the distribution of high (DockQ ≥ 0.8), medium (0.5 ≤ DockQ < 0.8), and acceptable (0.23 ≤ DockQ < 0.5) quality models, comparing ESMFold with (A) AFM with enhanced sampling and (B) AFM, AF3, and ColabFold (CF). Data for AF-based tools are taken from previous studies.^,

Figure 3

Example high-quality protein–peptide docking results using ESMFold. Protein receptor shown as surface: experimental peptide in magenta; predicted peptide in lime.