Introduction
irit is a set of small tools for the parts of sequencing analysis that pipelines leave to the researcher: building sample sheets, checking read quality, and getting the right reference databases onto disk. They are written in Rust, ship as single binaries, and are meant to run on the laptop a student actually owns, not only on a cluster.
The name is the Indonesian and Javanese word for frugal: getting the most out of modest resources. The project was called leanbio for its first day, until it turned out a biologics manufacturer already uses that name.
Two things came out of the research behind this project (landscape):
- The heavy science is mostly solved; the glue isn’t. Assemblers, classifiers and annotators are mature. What still costs researchers hours is the manual work around them: writing sample sheets by hand, fetching and matching database versions, collecting QC into one table.
- The standard tooling assumes big machines. Running the nf-core/bacass test profile on a laptop needed only 3.1 GB of RAM, but pulled about 16 GB of container images to produce 78 MB of results (setup notes). FastQC alone used 483 MB of RAM for one small file. GTDB-Tk now needs about 140 GB of RAM (details).
So each tool here does one manual step, does it on modest hardware, and is checked against the established tool it replaces or feeds.
Principles
Section titled “Principles”- Match the reference, then beat it on resources.
leanqcreproduces FastQC’s verdicts before claiming to be faster.sheetgen’s output is validated by running the real pipeline. - Refuse rather than guess.
sheetgenstops on values the pipeline would reject;dblockdeletes a download whose checksum doesn’t match and refuses databases that won’t fit on disk. - Record what was used.
dblock.lockand the JSON outputs are there so an analysis can be repeated exactly. - Measured claims only. Every number in these docs comes from a run described alongside it.
Status
Section titled “Status”Early. The tools work and are tested on real data (see benchmarks), but they have one test machine so far. Real-data runs on that laptop, each in an 8 GB memory cap: an AI agent redoing two published genome papers through irit-mcp alone (first, second), nanopore isolates, metagenome to MAGs and how much sequencing depth is enough. The roadmap lists what comes next.