PIPENN

PIPENN predicts residue-level protein binding interfaces from protein sequence for protein-protein, protein-small-molecule, and protein-nucleotide (DNA/RNA) interactions.

Key Features:

• Residue-level prediction: Predicts binding interface residues at the amino-acid level using only protein sequence input.
• Interaction types: Covers protein-protein interactions (PPI), protein-small-molecule binding, and protein-nucleotide (DNA/RNA) binding.
• Deep learning architectures: Implements seven distinct deep learning architectures evaluated across datasets.
• Ensemble predictor: Combines multiple architectures into an ensemble that yields superior prediction accuracy compared to single architectures.
• Benchmarking sets: Evaluated across eleven independent test sets, including an independent test set denoted ZK448.
• Training dataset (BioDL): Uses the BioDL dataset comprising PPI data from the Protein Data Bank (PDB) and protein-ligand interactions from the BioLip database.
• Dataset augmentation: BioDL is augmented with curated homo- and heteromeric PPI datasets.
• Model development experiments: Assessed the impact of data features, spatial forms, encoding schemes, network initializations, loss functions, regularization mechanisms, and activation functions on predictor performance.
• Performance metrics: Benchmarked performance using area under the curve (AUC) metrics, reporting AUCs of 0.718 for PPI, 0.823 for protein-nucleotide, and 0.842 for protein-small-molecule interactions on ZK448.

Scientific Applications:

• Interface mapping: Identification of residue-level interfaces for studying signaling pathways, transcription, receptor activation, and enzymatic activity.
• Interaction characterization: Prediction of protein-protein, protein-small-molecule, and protein-nucleotide binding sites to support molecular interaction studies.
• Experimental prioritization: Prioritization of candidate interface residues for experimental validation or mutational analysis.
• Method benchmarking: Comparative evaluation and benchmarking of sequence-based protein interface predictors.

Methodology:

Seven distinct deep learning architectures were trained and evaluated, combined into an ensemble; evaluations used the BioDL dataset (PDB for PPI and BioLip for ligands) augmented with curated homo- and heteromeric PPI datasets, tested across eleven independent test sets including ZK448, with performance measured by AUC; development experiments varied data features, spatial forms, encoding schemes, network initializations, loss functions, regularization mechanisms, and activation functions.