Tools for exploring large metagenomic datasets, and the unusual proteins they turn up.

I built Sharur to explore large metagenomic datasets with AI agents. It brings genomic neighborhoods, sequence and structure search, and protein language model embeddings into one queryable system. Sharur passes those embeddings directly to ELSA’s conserved synteny search; see below.

Sharur combines system calls from TXSScan and DefenseFinder; PFAM, KOFAM, and HydDB annotations; MinCED CRISPR arrays; and biosynthetic gene cluster results from antiSMASH and GECCO.

Evidence and replay instructions stay attached to each finding. Independent AI agents can systematically rerun analyses, test falsification criteria, and vet claims before they move forward.

I’m studying unusually large proteins from Omnitrophota. I use comparative genomics and structure prediction to work out how they grew, how their repeated architectures are organized, and what their genomic neighborhoods reveal about their evolution.

A large part of the work is separating evolutionary signal from assembly, gene-calling, and structure-prediction artifacts.

I built Talea to read patterns in a protein language model’s attention and find solenoids: elongated protein structures made from repeating units.

Starting from sequence alone, it locates the repeated region, estimates the length of its repeating unit, and measures whether the units form an open path or close into a loop. That makes it possible to search metagenomes for unusual structures, then choose promising proteins to predict and inspect.

I built ELSA to search for conserved gene neighborhoods using protein language model embeddings as anchors, then join collinear hits into syntenic blocks.

I co-first-authored Gaia, a platform that uses genomic context to improve protein sequence search and annotation.