DeepTMInter

DeepTMInter predicts interaction sites in α-helical transmembrane (TM) proteins using deep learning to annotate residues involved in molecular interactions.

Key Features:

• Architecture: Ultra-deep residual neural networks combined with a stacked generalization ensemble technique for final predictions.
• Input features: Uses topological, physicochemical, and evolutionary properties of amino acid sequences in TM proteins.
• Performance metrics: Reported area under the ROC curve (AUC) of 0.689 and area under the precision-recall curve (AUCPR) of 0.598.
• Predicted coverage: Estimates that typically 10%–25% of amino acid sites are involved in interactions, with values up to 30% in ion channels.

Scientific Applications:

• Interaction-site annotation: Identification and annotation of interaction residues in α-helical TM proteins.
• Comparative analysis across families: Comparative prediction of interaction-site prevalence across functional families of human transmembrane proteins.
• Ion channel analysis: Detection of interaction-rich regions in ion channels where predicted interacting-site fractions can reach ~30%.

Methodology:

Employs ultra-deep residual neural networks with a stacked generalization ensemble trained on topological, physicochemical, and evolutionary sequence features and evaluated using AUC and AUCPR.