FunFOLD5: Combining Localised Templates and Structural Homology for Accurate Predictions of Biologically Relevant Protein-Ligand Interactions.

Abstract

MOTIVATION A central challenge in protein-ligand modelling is to predict not only the presence of an interaction, but also the correct ligand identity, binding-site location, and biologically relevant bound conformation. While recent deep learning approaches such as AlphaFold3 and Boltz-2 have demonstrated impressive accuracy in modelling protein-ligand complexes, these methods typically require prior knowledge of the ligand, focus on a single dominant binding pose, and rely heavily on large training datasets. Template-based methods offer a complementary solution by leveraging experimentally resolved protein-ligand complexes to transfer both binding-site architecture and ligand chemical identity, enabling function inference even for uncharacterised or engineered proteins. However, conventional template-based methods are limited by their dependence on global fold similarity, which can miss local binding features critical for ligand specificity. Localised template search strategies may overcome this limitation by capturing conserved pocket-level interactions, enabling more accurate and functionally relevant ligand-binding predictions across diverse protein families. RESULTS We present FunFOLD5, a template-based method for predicting protein-ligand binding sites. Whereas its predecessor FunFOLD4 relied on global fold similarity to identify templates, FunFOLD5 combines local and global searches, integrating rapid structural template retrieval with a local targeted protocol to improve interaction accuracy. The pipeline combines Foldseek-based template searching, curated ligand annotations from BioLiP2, and ligand superposition followed by spatial clustering to identify associated ligands and consensus binding sites. Benchmarking on blind targets from CASP15 and CASP16 demonstrates robust performance across a wide range of biologically relevant ligands, including nucleotides, metal ions, cofactors, and small metabolites. FunFOLD5 achieved high-accuracy predictions for 81.2% of ligand instances (56 of 69). Importantly, the local targeted protocol consistently improved ligand identification, residue-level recall, and binding-pose accuracy. AVAILABILITY FunFOLD5 is freely available as a Dockerised pipeline at https://hub.docker.com/r/radiyaman/funfold5_template, providing a scalable and reproducible framework for large-scale protein-ligand interaction analysis and functional annotation. SUPPLEMENTARY INFORMATION Supplementary data are available at Bioinformatics online.