Improving protein domain classification for third-generation sequencing reads using deep learning.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Third-generation sequencing (TGS) yields long DNA reads (10s to 100s of kb).
• TGS enables protein domain annotation without assembly, offering biological insights.
• High error rates in TGS data challenge existing domain analysis methods, reducing accuracy.

Purpose of the Study:

• To develop a computational method for accurate protein domain prediction in long, noisy TGS reads.
• To address the limitations of current domain analysis pipelines with TGS data.

Main Methods:

• Introduction of ProDOMA, a deep learning model for protein domain classification.
• Utilizes deep neural networks with 3-frame translation encoding to capture conserved features.
• Formulates the problem as an open-set to enable rejection of reads lacking targeted domains.

Main Results:

• ProDOMA demonstrates superior performance in protein domain classification compared to HMMER and DeepFam.
• Experiments on simulated and real TGS human genome data validate ProDOMA's effectiveness.
• The model successfully classifies domains in long, noisy reads.

Conclusions:

• ProDOMA is an effective end-to-end tool for protein domain analysis on long, noisy reads.
• The tool operates without the necessity of prior error correction.
• ProDOMA enhances the utility of TGS data for biological function studies.