FGeneBERT: function-driven pre-trained gene language model for metagenomics.

Area of Science:

• Computational Biology and Bioinformatics
• Genomics and Metagenomics
• Systems Biology

Background:

• Metagenomic data from diverse environments (oceans, soils) are crucial for understanding health and ecosystems.
• Current K-mer-based methods struggle with capturing gene context, biological meaning, and complex gene relationships in metagenomics.
• Existing approaches fail to effectively encode biologically meaningful genes and address inherent data complexities like one-to-many relationships.

Purpose of the Study:

• To introduce FGeneBERT, a novel pre-trained model for enhanced metagenomic data analysis.
• To overcome limitations of K-mer methods in capturing gene context and biological relevance.
• To improve the understanding of inter-gene relationships and gene sequence-function associations in metagenomic datasets.

Main Methods:

• Developed FGeneBERT, a metagenomic pre-trained model utilizing a protein-based gene representation as a tokenizer.
• Incorporated masked gene modeling to capture inter-gene contextual relationships.
• Employed triplet enhanced metagenomic contrastive learning to elucidate gene sequence-function relationships, pre-trained on over 100 million sequences.

Main Results:

• FGeneBERT demonstrated superior performance across four levels: gene, functional, bacterial, and environmental.
• The model effectively analyzed metagenomic datasets with input sizes ranging from 1 to 213k sequences.
• Case studies on ATP synthase and gene operons validated FGeneBERT's functional recognition capabilities and biological relevance.

Conclusions:

• FGeneBERT offers a significant advancement in metagenomic data analysis by providing context-aware and structure-relevant gene representations.
• The model's protein-based approach and advanced learning strategies enhance the understanding of complex gene interactions and functions.
• FGeneBERT shows strong potential for diverse applications in metagenomic research, improving biological insights from complex datasets.