DEEPred: Automated Protein Function Prediction with Multi-task Feed-forward Deep Neural Networks.

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning

Background:

• Automated protein function prediction is essential for annotating unknown protein sequences.
• Deep learning methods show promise, outperforming traditional algorithms in related fields.
• Accurate prediction tools are needed to address the growing volume of sequence data.

Purpose of the Study:

• To introduce DEEPred, a deep neural network model for Gene Ontology (GO) based protein function prediction.
• To evaluate DEEPred's performance using various protein descriptors, dataset sizes, and GO term levels.
• To assess the impact of including large, potentially noisy electronic GO annotations on predictive performance.

Main Methods:

• Developed DEEPred, a hierarchical stack of multi-task feed-forward deep neural networks.
• Optimized DEEPred through extensive hyper-parameter testing.
• Benchmarked DEEPred against state-of-the-art methods using CAFA2 and CAFA3 challenge datasets.

Main Results:

• DEEPred demonstrated strong predictive performance on benchmark datasets.
• The model showed significant potential, especially when trained on large datasets, including electronically generated GO annotations.
• A case study on Pseudomonas aeruginosa biofilm formation validated novel annotations.

Conclusions:

• Deep learning algorithms, like DEEPred, offer substantial potential for protein function prediction, particularly with large-scale data.
• The DEEPred neural network architecture is adaptable for predicting other ontological associations.
• The study provides open-source code and datasets for reproducibility and further research.