DBPred

DBPred predicts DNA-binding residues in protein sequences to identify protein–DNA interaction sites relevant to transcription regulation, gene expression modulation, and RNA splicing.

Key Features:

• Data-Driven Approach: Uses a benchmark dataset derived from hybridNAP and ProNA2020 comprising 864 proteins for training and 308 for validation, with CD-HIT applied to remove sequences >30% identity yielding a refined set of 646 training and 46 validation proteins.
• Feature Engineering: Computes amino acid binary profiles, physicochemical-properties-based binary profiles, position-specific scoring matrix (PSSM) profiles, and hybrid features combining these profiles.
• Machine Learning Model: Trains a 1D Convolutional Neural Network (1D-CNN) and reports AUROC values of 0.83 (training)/0.74 (validation) for amino acid binary profiles, 0.86/0.73 for physicochemical binary profiles, 0.91/0.74 for PSSM profiles, and 0.91/0.79 for the hybrid feature set.
• Performance Comparison: Demonstrates improved prediction performance relative to existing methods on the provided training and validation datasets.

Scientific Applications:

• Transcription factor binding prediction: Improves identification of DNA-contacting residues to enhance transcription factor binding site analyses.
• Synthetic biology design: Assists design of synthetic constructs by identifying protein residues involved in DNA binding.
• Drug discovery and target identification: Supports identification of potential therapeutic targets by locating DNA-binding residues on proteins.

Methodology:

Derived datasets from hybridNAP and ProNA2020, reduced redundancy with CD-HIT at a 30% sequence identity threshold, generated amino acid binary, physicochemical binary and PSSM profiles and hybrid features, trained a 1D-CNN model, and evaluated performance using AUROC on training and validation sets.