Deciphering enzymatic potential in metagenomic reads through DNA language models

Area of Science:

• Microbiology
• Bioinformatics
• Genomics

Background:

• Microbial communities perform vital global functions, but their full potential is largely unknown.
• Metagenomics sequences microbial DNA from environments, but analysis often relies on comparing to known sequences.
• Current methods limit the exploration of microbial "dark matter" and novel functional genes.

Purpose of the Study:

• To develop novel language models (LMs) for reference-free analysis of metagenomic reads.
• To introduce REMME (Read EMbedder for Metagenomic Exploration) for understanding DNA context.
• To present REBEAN (Read Embedding-Based Enzyme ANnotator) for predicting enzymatic functions of genes within reads.

Main Methods:

• Development of a pretrained foundation LM, REMME, for metagenomic read embedding.
• Fine-tuning REMME into REBEAN for predicting enzymatic potential and molecular functions.
• Utilizing LMs to identify function-relevant gene segments and annotate orphan sequences.

Main Results:

• REBEAN successfully predicts enzymatic functions, even for previously uncharacterized (orphan) sequences.
• The models identify function-relevant gene parts without explicit training for this task.
• Novel enzymes were discovered, expanding the understanding of enzymatic capabilities in microbial communities.

Conclusions:

• Language model-based approaches offer a paradigm shift for reference-free metagenomic analysis.
• REMME and REBEAN significantly expand the annotation of enzymatic potential in unassembled metagenomic reads.
• These models advance the exploration of microbial "dark matter" and functional diversity.