Seq2Pocket: augmenting protein language models for spatially consistent binding site prediction
- 1Department of Software Engineering, Faculty of Mathematics and Physics, Charles University , Prague, 118 00,
- 2Department of Cell Biology, Faculty of Science, Charles University , Prague, 128 43,
Abstract
Motivation: Protein–ligand binding site prediction (LBS) is important for many areas of structural biology and molecular modeling, where, as in other tasks, protein language models (pLMs) have shown great promise. In their application to LBS, the pLM classifies each amino acid as binding or not, but translating these predictions into three-dimensional binding pockets remains challenging; in particular, residue-centric predictions tend to produce spatially fragmented pockets.
Results: We present Seq2Pocket, a methodology for pocket-level LBS prediction that combines pLM finetuning, data enhancement, and structure-aware post-processing. First, we introduce sc-PDBenhanced, an extended training dataset that augments an existing sc-PDB dataset with additional small-molecule and ion-binding sites, improving coverage of non-obvious and small pockets. Second, we employ an embedding-supported smoothing classifier to refine residue-level predictions. Third, we define the Pocket Fragmentation Index and use it to select a clustering approach that preserves a consistent mapping between predictions and ground-truth pockets. We evaluate Seq2Pocket on two tasks: general binding site prediction using the LIGYSIS benchmark and cryptic binding site prediction using CryptoBench benchmark. Across both benchmarks, Seq2Pocket achieves state-of-the-art DCCtop-N and DCCtop-(N + 2) performance. In particular, for general binding site prediction on the LIGYSIS benchmark, it improves distance-center-to-center recall by up to 13%, outperforming existing predictors. This improvement originates from training data quality rather than model architecture, as demonstrated by an ablation study. We believe that these findings contribute to more reliable evaluation practices in ligand binding site prediction, highlight the importance of training data curation, and provide pocket-level prediction tools that are better suited for downstream applications such as drug discovery.
Availability: The Seq2Pocket web interface is available at https://seq2pocket.projekty.ms.mff.cuni.cz, running the ESM-2 650M variant. The sc-PDBenhanced is published using Zenodo platform at https://zenodo.org/records/18271517. The models are published at HuggingFace: https://huggingface.co/skrhakv/seq2pocket.
Related
- Augmenting patient safety surveillance in radiation oncology with large language model-based root cause analysis: A proof-of-concept study
- Tractable and Thermodynamically Consistent pKa Prediction via First-Dissociation Aggregation
- Integrating Spatially Adjusted Protein Summaries for Survival Prediction in Spatial Proteomics
- Limitations of general protein language models and public TCRpMHC data in training specificity prediction models
- Benchmarking Large Language Models for Assay-Aware Drug-Target Binding Affinity Prediction
BibTeX
@article{skrhak2026seq2pocket,
title = {{Seq2Pocket: augmenting protein language models for spatially consistent binding site prediction}},
author = {Vít Škrhák and Lukáš Polák and Marian Novotný and David Hoksza},
journal = {Bioinformatics},
year = {2026},
month = oct,
doi = {10.1093/bioinformatics/btag778},
url = {https://academic.oup.com/bioinformatics/article/doi/10.1093/bioinformatics/btag778/8931593},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Seq2Pocket: augmenting protein language models for spatially consistent binding site prediction
AU - Vít Škrhák
AU - Lukáš Polák
AU - Marian Novotný
AU - David Hoksza
PY - 2026
DA - 2026/10/16
JO - Bioinformatics
DO - 10.1093/bioinformatics/btag778
UR - https://academic.oup.com/bioinformatics/article/doi/10.1093/bioinformatics/btag778/8931593
AB - Motivation: Protein–ligand binding site prediction (LBS) is important for many areas of structural biology and molecular modeling, where, as in other tasks, protein language models (pLMs) have shown great promise. In their application to LBS, the pLM classifies each amino acid as binding or not, but translating these predictions into three-dimensional binding pockets remains challenging; in particular, residue-centric predictions tend to produce spatially fragmented pockets. Results: We present Seq2Pocket, a methodology for pocket-level LBS prediction that combines pLM finetuning, data enhancement, and structure-aware post-processing. First, we introduce sc-PDBenhanced, an extended training dataset that augments an existing sc-PDB dataset with additional small-molecule and ion-binding sites, improving coverage of non-obvious and small pockets. Second, we employ an embedding-supported smoothing classifier to refine residue-level predictions. Third, we define the Pocket Fragmentation Index and use it to select a clustering approach that preserves a consistent mapping between predictions and ground-truth pockets. We evaluate Seq2Pocket on two tasks: general binding site prediction using the LIGYSIS benchmark and cryptic binding site prediction using CryptoBench benchmark. Across both benchmarks, Seq2Pocket achieves state-of-the-art DCCtop-N and DCCtop-(N + 2) performance. In particular, for general binding site prediction on the LIGYSIS benchmark, it improves distance-center-to-center recall by up to 13%, outperforming existing predictors. This improvement originates from training data quality rather than model architecture, as demonstrated by an ablation study. We believe that these findings contribute to more reliable evaluation practices in ligand binding site prediction, highlight the importance of training data curation, and provide pocket-level prediction tools that are better suited for downstream applications such as drug discovery. Availability: The Seq2Pocket web interface is available at https://seq2pocket.projekty.ms.mff.cuni.cz, running the ESM-2 650M variant. The sc-PDBenhanced is published using Zenodo platform at https://zenodo.org/records/18271517. The models are published at HuggingFace: https://huggingface.co/skrhakv/seq2pocket.
SN - 1367-4803
SN - 1367-4811
ER -