PHDhtm

PHDhtm predicts the locations of transmembrane helices in integral membrane proteins using a neural network that leverages evolutionary information to improve prediction accuracy.

Key Features:

• Input Data: Processes multiple sequence alignments for each position within a 13-residue window, extracting amino acid frequencies, conservation weights, counts of insertions and deletions, and positional context within the protein chain.
• Additional Inputs: Incorporates overall amino acid composition and total protein length as global input features.
• Neural Network: Uses a trained neural network architecture to interpret evolutionary and compositional inputs to identify transmembrane regions.
• Output: Produces per-residue two-state assignments and predicted transmembrane helix segments.

Scientific Applications:

• Accuracy: In cross-validation on 69 proteins with experimentally determined transmembrane segments, it achieved an overall two-state per-residue accuracy of 95% and approximately 94% of predicted segments were correct.
• Specificity: When tested on known globular proteins as negative controls, it misclassified fewer than 5% as containing transmembrane helices.
• Genome-scale analysis: Applied to 269 open reading frames from yeast chromosome VIII, it predicted that about one-fourth of proteins contain at least one transmembrane helix and roughly 20% have more than one.
• Research uses: Supports genome annotation, structural modeling, and functional studies of membrane proteins by providing transmembrane region predictions.

Methodology:

PHDhtm computes features from multiple sequence alignments in a 13-residue window (amino acid frequency, conservation weights, number of insertions/deletions, positional context), combines these with overall amino acid composition and protein length, and applies a neural network trained to detect evolutionary patterns indicative of transmembrane helices.