DeepREx

DeepREx predicts per-residue solvent exposure to characterize protein–solvent interactions and identify regions important for maintaining functional integrity during surface engineering when structural data are incomplete or absent.

Key Features:

• Deep learning-based residue classification: A deep learning algorithm classifies each residue in a protein sequence as buried or exposed.
• Training dataset from Protein Data Bank: The model is trained on a high-quality, non-redundant dataset derived from the Protein Data Bank comprising 2332 monomeric protein chains.
• Independent benchmarking: Performance is evaluated against an independent test set of 200 unrelated sequences, reporting state-of-the-art accuracy for solvent exposure prediction.
• Integrated sequence- and structure-derived features: Predictions incorporate residue conservation from multiple sequence alignments, local sequence hydrophobicity, residue flexibility predicted by MEDUSA, secondary structure prediction, and disordered region prediction from MobiDB-Lite3.0.

Scientific Applications:

• Protein engineering: Identification of exposed and buried regions to guide targeted mutations that modulate stability while preserving functional integrity.
• Analysis of proteins without experimental structures: Characterization of solvent exposure for sequences lacking complete structural data to inform experimental design and interpretation.
• Variant pathogenicity and functional annotation: Assessment of single-residue variations (SRVs) in human proteins, supporting analyses that link lower solvent accessibility to disease-associated variations.

Methodology:

Per-residue classification via a deep learning algorithm trained on a non-redundant PDB-derived dataset of 2332 monomeric chains and benchmarked on 200 unrelated sequences; integration of features from multiple sequence alignments (residue conservation), local sequence hydrophobicity calculations, residue flexibility predicted by MEDUSA, secondary structure prediction, and disordered region prediction using MobiDB-Lite3.0.