DEEPred

DEEPred predicts protein functions using a hierarchical stack of multi-task feed-forward deep neural networks for Gene Ontology (GO)-based annotation.

Key Features:

• Deep Learning Architecture: Utilizes a multi-task feed-forward deep neural network architecture that mitigates overfitting, enables efficient training, and yields improved predictive performance relative to traditional algorithms.
• Hierarchical Stacking: Implements a hierarchical stacking strategy to handle multiple levels of GO terms for nuanced function prediction.
• Hyper-parameter Optimization: Employs rigorous hyper-parameter testing to optimize model performance.
• Diverse Protein Descriptors: Benchmarked using three distinct types of protein descriptors to assess applicability across different feature representations.
• Training with Varied Data Sizes: Trains on datasets of varying sizes, including incorporation of electronically generated GO annotations to evaluate the impact of larger, potentially noisy data.
• Benchmarking with CAFA: Evaluated performance using CAFA2 and CAFA3 challenge datasets for comparative assessment against state-of-the-art methods.
• Extensibility to Other Ontologies: Neural network architecture can be adapted to predict other types of ontological associations beyond GO-based protein functions.

Scientific Applications:

• GO-based protein function annotation: Predicts GO term annotations for uncharacterized protein sequences and can generate novel annotations for experimental consideration.
• Case study on Pseudomonas aeruginosa: Applied to literature-based prediction of the biofilm formation process in Pseudomonas aeruginosa to exemplify predictive capability.

Methodology:

Uses a hierarchical stack of multi-task feed-forward deep neural networks with hyper-parameter testing; benchmarks include three protein descriptor types; training was performed on datasets of varying sizes including electronically generated GO annotations; evaluation used CAFA2 and CAFA3 datasets.