Seq2Bind webserver for binding site prediction from sequences using fine-tuned protein language models.

Area of Science:

• Computational Biology
• Structural Biology
• Drug Discovery

Background:

• Protein-protein interactions (PPIs) are fundamental to cellular processes.
• Understanding PPIs at the residue level is key for therapeutic development.
• Current methods for predicting binding affinity often require 3D structures, limiting their applicability.

Purpose of the Study:

• To develop a sequence-based computational framework, Seq2Bind, for predicting protein-protein binding affinity and identifying critical binding residues.
• To evaluate the performance of fine-tuned protein language models (PLMs) for this task.
• To provide a webserver tool for rapid screening of PPIs.

Main Methods:

• Fine-tuned four PLM architectures (ProtBERT, ProtT5, ESM2, Bidirectional LSTM) on the SKEMPI 2.0 dataset.
• Evaluated models on 6063 dimer proteins from the Protein Data Bank using alanine mutagenesis data.
• Assessed performance using N-factor metrics and compared against structural docking (HADDOCK3) and mutation-based baselines on human protein complexes.

Main Results:

• ESM2 and ProtBERT demonstrated high interface-residue recovery rates (67.4% and 68.2% at N-factor=3, respectively).
• Seq2Bind, utilizing ESM2 and ProtBERT, outperformed HADDOCK3 on human protein complexes (37.2% and 35.1% vs. 32.1% at N-factor=2).
• The sequence-based approach successfully handled disordered proteins.

Conclusions:

• Seq2Bind offers a powerful, sequence-based alternative for predicting PPIs and identifying binding residues.
• The webserver enables rapid screening and can guide structural methods like blind docking.
• This approach has significant implications for accelerating drug discovery and understanding cellular mechanisms.