A Semi-Supervised Autoencoder-Based Approach for Protein Function Prediction.

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in genomics

Background:

• Next-generation sequencing yields vast protein data, but functional annotation lags.
• A significant gap exists between available protein sequences and their known functions.
• Current methods for protein function determination are often expensive and time-consuming.

Purpose of the Study:

• To develop an advanced deep-learning approach for accurate protein function prediction using only protein sequences.
• To address the growing challenge of annotating protein functions in the post-genomic era.
• To provide a scalable and efficient method for classifying protein functions.

Main Methods:

• A semi-supervised deep-learning strategy utilizing a set of autoencoders, each trained for a specific protein function.
• Training 932 autoencoders for biological processes and 585 for molecular functions separately.
• Employing reconstruction losses from each autoencoder as features for sequence classification.

Main Results:

• The proposed autoencoder-based model demonstrated promising performance in predicting protein functions on test datasets.
• The method successfully classified protein sequences into their corresponding biological processes and molecular functions.
• Achieved high accuracy in protein function prediction, highlighting the efficacy of the deep learning approach.

Conclusions:

• The developed deep-learning method offers an effective solution for protein function prediction from sequences.
• This approach can be readily extended to predict a larger number of protein functions with sufficient data.
• The availability of code and models facilitates further research and application in bioinformatics.