iProDNA-CapsNet

iProDNA-CapsNet predicts DNA-binding residues from protein sequences by applying an ensemble of Capsule Neural Networks to position-specific scoring matrix (PSSM) profiles for residue-level identification of protein–DNA interfaces.

Key Features:

• Ensemble of Capsule Neural Networks: Uses an ensemble of Capsule Neural Networks (CapsNets) to model complex spatial and hierarchical relationships in protein sequence features.
• Position-Specific Scoring Matrix (PSSM) Input: Operates on PSSM profiles to incorporate evolutionary conservation information of amino acid positions into predictions.
• Benchmark Training and Evaluation: Trained on the PDNA-543 dataset and evaluated on the PDNA-TEST dataset introduced by Hu et al. (2017).
• Comparative Performance at FPR ≈ 5%: At a decision threshold corresponding to a false positive rate of approximately 5%, reported improvements versus TargetDNA (≈+2.0% accuracy, ≈+2.0% sensitivity, ≈+14.0% precision, ≈+5.0% MCC) and versus BindN+ (≈+1.0% accuracy, ≈+75.0% sensitivity, ≈+45.0% precision, ≈+77.0% MCC).
• Robust MCC Gains Across Thresholds: Demonstrated higher Matthews's correlation coefficient (MCC) across decision thresholds, with reported increases ranging from approximately 8.0% up to 220.0% compared to other models.

Scientific Applications:

• Gene regulation studies: Identifying DNA-binding residues to inform analyses of regulatory protein–DNA interactions involved in transcriptional control.
• Transcription factor binding characterization: Mapping residue-level contacts to support studies of transcription factor–DNA specificity.
• Mechanistic studies of genetic disease: Locating DNA-binding residues to aid investigation of molecular mechanisms underlying genetic disorders.

Methodology:

Uses PSSM profiles as input to an ensemble of Capsule Neural Networks; trained on PDNA-543, evaluated on PDNA-TEST, and compared to TargetDNA and BindN+ using decision thresholds including one with FPR ≈ 5%.