MicrobeRT: A Framework and Suite of Fine-tuned Language Models for Genomic Sequences

In the realm of infectious disease surveillance, accurate and reproducible metagenomic analysis is paramount. Traditional diagnostic methods like culturing, antigen assays, and multiplex PCR panels are constrained by their focus on predefined organisms, often overlooking atypical or novel infections. Shotgun metagenomics, while providing hypothesis-free detection, faces hurdles of workflow complexity, reliance on mutable reference databases, and high computational costs.

To address these obstacles, researchers have developed MicrobeRT, an integrated framework and suite of fine-tuned genomic language models designed to classify metagenomic sequencing reads across three key tasks: (i) species-level taxonomic identification, (ii) antibiotic resistance gene detection, and (iii) pathogenicity assessment.

This innovative approach builds upon three pre-trained genomic language models - Nucleotide Transformer, DNABERT, and HyenaDNA - capitalizing on their expansive genomic pre-training, adaptability to downstream tasks, and the unique capability of HyenaDNA to capture long-range sequence dependencies. These models were fine-tuned using extensive, curated datasets encompassing microbial, viral, fungal, and human genomes, along with over 500,000 genomes and 250,000 antibiotic resistance gene sequences.

To assess the efficacy of MicrobeRT across various classification tasks, a benchmarking framework was established, pitting its performance against state-of-the-art (SOTA) bioinformatics tools. The results indicate that MicrobeRT not only delivers robust performance but also significantly reduces computational overhead. The framework is now available as open-source software, allowing for further fine-tuning and the seamless integration of alternative models and classifiers.

Moreover, MicrobeRT showcases the practicality of employing its taxonomic classification models as an optional enhancement within the open-source TaxTriage pipeline. By demonstrating the feasibility of three clinically relevant multi-label analyses, MicrobeRT paves the way for scalable, reproducible metagenomic diagnostics, promising broader adoption of AI/ML-enhanced infectious disease testing.