#2 · ML / AI methods

Benchmarks and learned models expose gaps in metagenomic analysis software

The common thread is analysis software that operates directly on sequencing data and is judged against simulated, curated, or benchmark datasets. MAG-E benchmarks assemblers, binners, and QC tools for metagenome-assembled genome pipelines in the human gut microbiome, and a metagenomic binning benchmark on CAMI, custom-simulated and real datasets finds that depth, complexity and sample number drive genome recovery. The trnL study compares OBITools3/ecoPCR, RESCRIPt and MetaCurator for reference database curation, and MicroSSNet provides an R package for single-sample and aggregated microbial co-abundance networks. LOCOM2 addresses differential abundance, and DIANA predicts host, community type and material of ancient metagenomic samples from unitig abundances across 2,597 runs. What is new is the shift toward learned models operating on raw reads and unitigs, alongside benchmarks that expose pipeline-level performance gaps. The set is heterogeneous, so this describes the dominant pattern rather than a property shared by every item.