PREMISE: A Quality-Aware Probabilistic Framework for High-Resolution Source Assignment in Metagenomic Sequencing

Metagenomic applications, from influenza A virus (IAV) subtype tracking to bacterial strain resolution, frequently require sub-species resolution.However, traditional $k$-mer-based methods fragment reads into short $k$-mers and discard per-base quality scores.By omitting both intra-read linkage and error information, these methods confound closely related genomic variants, thereby impeding real-time pathogen surveillance and containment.We introduce \premise (Pathogen Resolution via Expectation-Maximization In Sequencing Experiments), a probabilistic, alignment-based framework implemented in Rust for high-resolution source assignment against near-complete reference databases.\premise aligns whole reads using a bidirectional run-length FM-index and quantifies reference abundances with a quality-aware Expectation-Maximization algorithm mapped over a taxonomy.This framework applies an optional sparsity penalty to resolve ambiguous assignments, which is particularly important when sparse samples are mapped against reference databases rich in highly homologous candidate references.On ten dense CAMI II strain-madness samples, \premise achieves the highest species-level read classification F1 score ($0.992$) among eight evaluated methods, with no false-positive or missed species.At the strain level, \premise maps reads to their source genome with top precision ($0.80$ vs. at most $0.64$ for competing methods), achieves 100\% recall, and estimates abundances as accurately as the best competing tool.These results demonstrate \premise's ability to provide rigorous, uncertainty-aware source resolution across both dense and sparse metagenomes, accurately resolving targets ranging from viral subtypes to bacterial strains.\\textbf{Availability:} \url{https://github.com/sriram98v/premise} under an MIT license.\\textbf{Contact:} \href{sriramv@iastate.edu}{sriramv@iastate.edu}\\