Seq2Pocket: Augmenting protein language models for spatially consistent binding site prediction.

Seq2Pocket improves spatially consistent protein-ligand binding site prediction, achieving up to 13% better recall on the LIGYSIS benchmark.

• Why it matters: Accurate pocket-level predictions are crucial for structural biology and drug discovery, but current residue-centric methods often produce fragmented pockets, limiting their utility.
• What they did: The approach combines protein language model finetuning, an extended training dataset called sc-PDBenhanced, and a structure-aware post-processing step using the Pocket Fragmentation Index, evaluated on two benchmarks.
• The result: Seq2Pocket outperforms existing predictors with state-of-the-art DCCtop-N scores, demonstrating that training data quality significantly enhances prediction performance and supporting more reliable ligand binding site evaluation.