SetBERT: the deep learning platform for contextualized embeddings and explainable predictions from high-throughput...

Area of Science:

• Microbiome bioinformatics
• Computational biology
• Machine learning for genomics

Background:

• High-throughput sequencing (HTS) generates vast microbiome data, offering AI opportunities.
• Current computational models process DNA sequences individually, missing crucial microbial interactions.
• Existing methods risk introducing protocol-specific bias through post-processing.

Purpose of the Study:

• To develop a generalized deep learning methodology for HTS data processing.
• To enable AI models to comprehend functional relationships and interactions within microbial communities.
• To create explainable predictions for downstream microbiome analysis tasks.

Main Methods:

• Introduced SetBERT, a robust pre-training methodology for HTS data.
• Leveraged sequence interactions within microbial communities for model training.
• Developed contextualized embeddings for generalized deep learning models.

Main Results:

• SetBERT achieved 95% accuracy in genus-level taxonomic classification.
• The model significantly outperformed existing computational methods.
• SetBERT autonomously provided biologically relevant explanations for its predictions.

Conclusions:

• SetBERT offers a powerful, explainable approach for microbiome analysis using HTS data.
• This methodology overcomes limitations of individual sequence processing and protocol bias.
• SetBERT enhances understanding of microbial community functional relationships.