Deep learning program to predict protein functions based on sequence information.

Area of Science:

• Bioinformatics
• Computational Biology
• Protein Science

Background:

• Predicting protein function is crucial for understanding biological systems.
• Current in silico methods often require extensive protein data, limiting their use for novel or poorly characterized proteins.

Purpose of the Study:

• To develop a deep learning model, FUTUSA (function teller using sequence alone), for predicting protein functions solely from sequence information.
• To improve the efficiency and accuracy of protein function prediction, especially for uncharacterized proteins.

Main Methods:

• Designed a binary classification deep learning program (FUTUSA) utilizing only protein sequence data.
• Implemented a convolution neural network with sequence segmentation for feature extraction.
• Trained the model to identify regional sequence patterns and their relationships.

Main Results:

• FUTUSA achieved a 49% improvement in predictive performance compared to full-length sequence processing.
• Demonstrated superior performance in predicting oxidoreductase, acetyltransferase, and demethylase activities over baseline methods.
• Successfully predicted the functional impact of point mutations on phenylalanine hydroxylase, linking it to PKU.

Conclusions:

• FUTUSA offers a novel, sequence-only deep learning approach for in silico protein function prediction.
• Sequence segmentation significantly enhances prediction efficiency and accuracy.
• The model has potential applications in predicting the clinical significance of mutations and polymorphisms.