srtpred

srtpred predicts whether mammalian protein sequences are secretory or non-secretory, including proteins lacking N-terminal signal peptides, using support vector machines, artificial neural networks, and sequence-similarity searches.

Key Features:

• Machine learning techniques: Uses artificial neural networks (ANN) and support vector machines (SVM) for classification.
• Training dataset: Trained on 3321 secretory and 3654 non-secretory mammalian protein sequences.
• Feature types: Employs 33 physico-chemical properties, amino acid composition, and dipeptide composition as input features.
• ANN performance: ANN-based modules achieve accuracies of 73.1%, 76.1%, and 77.1% across different feature sets.
• SVM performance: SVM-based modules achieve accuracies of 77.4%, 79.4%, and 79.9% across corresponding feature sets.
• Sequence-similarity search: Incorporates BLAST and PSI-BLAST modules with accuracies of 23.4% and 26.9%, respectively.
• Hybrid approach: Integrates amino acid and dipeptide composition-based SVM modules with PSI-BLAST to yield 83.2% accuracy, 60.4% sensitivity, and a 5% false positive rate.
• Signal-peptide independence: Capable of predicting secretory proteins even when N-terminal signal peptides are absent or misannotated.

Scientific Applications:

• Protein secretion pathway analysis: Distinguishes classical and non-classical secreted proteins to support studies of secretion mechanisms.
• Genomic and proteomic annotation: Enables prediction of secretory proteins in large-scale genome sequencing projects where N-terminal signal peptides or annotations may be missing or erroneous.

Methodology:

Predictions combine artificial neural networks and support vector machines trained on 3321 secretory and 3654 non-secretory mammalian sequences using 33 physico-chemical properties, amino acid composition, and dipeptide composition, supplemented by BLAST and PSI-BLAST sequence-similarity searches and a hybrid SVM+PSI-BLAST integration.