Work

New Computational Techniques to Extract and Analyze Sparse Signal from Noisy Metagenomic Datasets

Year: 2026
Type: dissertation
Abstract: This dissertation describes three computational tools that each pull a sparse biological signal out of noisy metagenomic data. The shared difficulty is that metagenomic DNA reports what genes are present in a community rather than which of those genes are doing anything, so each tool combines a machine-learning component with a biological constraint suited to its scale. Prokaregia puts a deep-learning taxonomic filter ahead of assembly and binning, which lets it recover complete symbiont genomes from host-dominated long-read metagenomes without first assembling the host; one benchmark works from only seven percent of the original reads. TAfinder3D adds structure-based search and an operon-aware filter to standard sequence homology and, on a phage dataset, more than doubles the number of detected toxin–antitoxin systems with no obvious degradation in annotation quality. TRLink uses contrastive learning over protein and genomic foundation-model embeddings to predict which transcriptional regulators (TRs) bind which DNA sites, with no prior motif input; on TRs held out from training it reached AUROC 0.8542 and AUPRC 0.8423, and the most sequence-distant regulators in the test set performed about as well as the rest.All three chapters lean on the same idea. A learned model is used to widen what can be detected, and the biology of the problem is then used to keep the output interpretable. The constraint is different in each case—the assembly and binning workflow in Prokaregia, operon organization in TAfinder3D, and held-out paired evaluation in TRLink—but the role it plays is consistent. Together the three tools show that learned representations can be turned into practical metagenomic workflows for genome recovery, gene annotation, and prediction of regulatory interactions.
Language: English
Open Access status: green
OpenAlex ID: W7164706129
Cites: 100
Cited by:
Related to: 10