Roadmap
What’s done, what’s next, and what’s further out. Items move up as evidence comes in.
-
sheetgen: schema-driven sample sheets, validated by an nf-core/bacass stub run. -
leanqc: FastQC-equivalent QC including per-tile quality; verdicts match FastQC 0.12.1 on the four bacass test files (40 of 40) and on a synthetic corpus (88 of 90, two deliberate encoding differences). -
dblock: database catalog, disk planning, resumable and streaming fetch, lock file, verify, tool-version check. - Baseline of nf-core/bacass on a laptop (setup, benchmarks).
-
isotable: one graded row per sample from nine tools’ outputs, isolate and MAG (MIMAG) modes, with a note for reduced-genome lineages. - v0.1.0 release with static binaries for Linux (x86_64, ARM64), macOS (Intel, Apple Silicon) and Windows, built by GitHub Actions on each tag.
- Upstream reports: GTDB R232 mock database checksum (GTDBTk#726); FastQC accepting a FASTQ cut off mid-line (FastQC#221, fix in FastQC#222).
- Independent reviews of every tool, with fixes and regression tests (sample-name joins, pairing rules, FastQC fidelity, database integrity, MCP security).
-
msascopefirst version: 100,000 × 30,000 alignments on a laptop (FASTA/gzip, Clustal, Stockholm; overview pyramid, consensus and identity tracks, search, go-to;--typeoverride). -
irit-mcp: typed MCP tools for the irit CLIs, with run records and areplay.shthat reruns a session without the AI; outputs confined to a root folder. - Split GTDB-Tk, step 1: the design and the laptop client
gtdbprep; its bundle classifies identically toclassify_wfon test genomes. - Documentation site (Astro + Starlight) at https://dzulfikar08.github.io/irit/, and an interactive competition demo.
- GTDB placement under an enforced 8.5 GiB memory cap on a 16 GB GitHub runner, with gPlace (write-up): a shared placement server isn’t needed.
- gPlace gives the same placements as pplacer (59/59 genomes) and runs down to
--max_memory1.9G (class-level tree) / 5.2G (full bacterial tree) (write-up). - irit 0.2.0: signed, notarised macOS builds and a curl installer.
-
leanqcvs FastQC-Rust benchmark on a GitHub runner;isotablereads FastQC/FastQC-Rust output. -
dblockkeeps selected folders of an archive and verifies GTDB-Tk folder hashes (gtdbtk-r232-clientforgtdbprep). -
irit-mcpwraps fastp 1.4.0, Bakta 1.12.2, CheckM2 1.1.0 and GTDB-Tk 2.8.0classify_wf, with only the flags of those versions, input and database checks, and the same run records andreplay.sh(0.3.0). Tested with stand-ins for the programs. - More
irit-mcpwrappers (0.3.2): Shovill 1.4.2, mlst 2.35.0 and AMRFinderPlus 4.2.7, so the isolate path from reads to resistance genes has no step outside the record. Checked with the real programs on real S. aureus reads (ST22, mecA) in the real-tools workflow. -
dblockadditions (0.3.2): the AMRFinderPlus database as a pinned entry (amrfinderplus-db-4.2-2026-10-29.1: NCBI’s 38 files with checksums we computed, indexed byamrfinder_indexbefore the folder is sealed), also usable as Bakta’samrfinderplus-db/latestinstead of runningamrfinder_update. - End-to-end run on a laptop through
irit-mcp(write-up): B. subtilis 168 from raw reads to the GTDB R232 species (leanqc, fastp, Shovill/SKESA, CheckM2,gtdbprep ani+ GTDB-Tkclassify_wf, isotable) under a 16 GB and an 8 GB memory cap; largest step 5.82 GiB, correct species (ANI 100.0). The replay reran every command with three unrecorded steps added; gaps filed as #2 to #11. -
seqplan(0.4.0): isolates per sequencing run and cost per isolate at a target coverage, from Illumina’s specifications and reported nanopore yields, each cited (seqplan). Next: measure how low the coverage can go for this kind of work (subsampling the K. pneumoniae reads from the paper reproduction), so a plan can use a measured coverage instead of a habit. -
leanqclong-read views (0.3.2): N50, N90, length bins and quality by length. - Replay gaps from the laptop run closed (0.3.2): output folders (#2),
naming genomes (#3), a recorded
gtdbprep anitool (#4), Shovill and leanqc outputs that change between runs (#5, #6), soreplay.shreruns the isolate path from reads to species on its own. - fieldsheet: an offline phone worksheet for sample metadata (guide, 0.6.0). Next: try it on a low-end Android phone in the field.
- GTDB-Tk on a desktop: measure before building. Peak memory and time for GTDB-Tk with and
without
--scratch_dir, and with gPlace and--max_memory, on a rented large-memory machine. See GTDB-Tk and memory. GTDB plans to move GTDB-Tk to gPlace “in the coming weeks”, placing directly on the full trees and dropping the backbone/class-level split, so no--placeroption (GTDBTk#728). Wait for that release, then measure it; until then thepplacershim stays. The placement mirror andirit-mcp’ssplit/check will need updating for the new layout. - GTDB placement data without the 60.8 GB download: an unofficial, credited mirror
(280 MB download) on GitHub Releases (
dzulfikar08/irit-data) and thedblockentriesgtdbtk-r232-placementand-lite(0.2.1);dblockstores identical files once as hard links (0.3.0), so the tenfasttree.logcopies take 2.7 GB instead of 7.4 GB. Waiting on GTDB GTDB has no plans for per-folder downloads and is fine with the mirror as long as it stays credited and links back (GTDBTk#727). - The ANI species check on a laptop:
gtdbprep aniscreens locally and range-reads only the reference sketches a genome needs, then runs GTDB-Tk’s own ANI screen on them; identical output to GTDB-Tk on the full database (write-up). The screen index is on irit-data (dblockentrygtdbtk-r232-ani-screen). Sketches come from an unpackedsketches.db, or, opt-in, from GTDB’s archive through a seek index (2–6× the download), with a per-user cache; GTDB won’t publish the skani folder unpacked and asked for exactly these two safeguards (GTDBTk#727). Built since (0.3.4): parallel range requests with retries (377 sketches from GTDB’s server in 474–562 s instead of 1,167–2,275 s), and the opt-in--species-only, which fetched 47 instead of 882 sketches for six test genomes with the same calls (fewer unassigned-hits rows than GTDB-Tk’s; measurements). On five genomes from crowded genera (Streptococcus, B. cereus, Burkholderia, Pseudomonas, Klebsiella)--species-onlyfetched 542 instead of 2,694 references with the same calls and the same related references (measurements). -
msascopeafter the redesign (built in 0.3.0–0.3.2: themes, names, hover card, colour schemes, hidden insert and gappy columns, tree panel, selection, export, sort and filter, motif search, annotation tracks, reference row, remembered between sessions, and annotation tracks and the pinned row in figures; 0.5.0: gzip decoded on its own thread, BGZF in parallel): alignments larger than RAM. - Nanopore isolates on a laptop:
flyeandmedakatools inirit-mcp(0.4.0). One carbapenemase-producing K. pneumoniae gave the same ST, the same 49 AMRFinderPlus elements, the same completeness and the same GTDB species from ONT reads alone as from Illumina, in an 8 GB cap (write-up). isotable reads long-read QC, the record hashes Medaka’s model file, and Medakabatch_size50 or 25 keeps it under 5 GiB with the same consensus (#21). Three more isolates with small resistance plasmids: same STs and every resistance element, once Flye’s--metacaught a 5 kb plasmid default mode missed; deep reads subsampled withrasusa; irit-mcp warns about plasmids Flye closed as several copies. Next: a ligation-kit dataset. Read files can now be checked against ENA’s MD5s (sheetgen verify, 0.6.0). -
dblockadditions: GTDB-Tk split download (waits for GTDB-Tk’s gPlace layout). Parallel downloads are built (0.5.0,fetch --jobs). - Check the
irit-mcpwrappers on a large machine: Bakta with the full database, SPAdes through Shovill (the real-tools workflow uses SKESA). GTDB-Tkclassify_wfhas now run with R232 data for a genome the ANI screen assigns, through the small data foldergtdbprep anibuilds (laptop run); genomes that need tree placement still go through the launcher and gPlace outside irit-mcp (#4).
-
subprep: validate NCBI/ENA submission metadata offline before upload. - Split GTDB-Tk, step 2 (parked): a placement service for
gtdbprepbundles. The gPlace test showed a 16 GB laptop can place on its own, so this is only worth building for larger batches. - Laptop benchmark for AI agents: can an agent with these tools produce a defensible microbial
genomics result on an 8 GB laptop, compared with a hand-built workflow (accuracy,
reproducibility, time, cost, human intervention)? A possible paper. First data point: one
isolate through
irit-mcpfits an 8 GB cap and gets the right species (laptop run), with the steps called by a scripted client rather than an agent. Second data point: an agent with onlyirit-mcp(no shell, no web) redid Table 1 of a December 2025 K. pneumoniae genome announcement on the laptop. AMR genes and MLST matched exactly, GTDB confirmed the species, and it found six irit bugs (#14–#19) (write-up). Third data point: a second paper, three E. coli and three Salmonella Kentucky from chickens in Bangladesh. All six STs matched, genome size within 0.13 %, and the agent caught twelve blaTEM alleles listed for one truncated gene (write-up). Still to do: more papers and lineages, and the hand-built comparison. - Gaps the paper redo showed: QUAST-style assembly stats (
assembly_stats), coverage by read mapping (coverm_genome) and Klebsiella typing (kleborate) are built (0.5.0). Still missing (both redos): virulence genes for other species (VirulenceFinder/VFDB), plasmid replicons (PlasmidFinder), ResFinder, Salmonella serotyping (SISTR/SeqSero2), a Kraken2 tool and clonal complexes. (Checking read files against ENA MD5s:sheetgen verify, 0.6.0.) - How much depth is enough (write-up): on the paper’s three
K. pneumoniae, every Table 1 call held down to 15× (20× is a practical target, 2.5 times as
many isolates per run as 50×); below that Shovill needs
gsize, and irit-mcp warns when it didn’t get one (#22). - A MAG route for metagenomes on 8 GB (laptop test): MEGAHIT,
CoverM and MetaBAT2 in irit-mcp, then CheckM2 (
lowmem), the GTDB ANI screen and isotable’s MAG mode. A CPR-rich groundwater sample gave 23 bins (4high*, 13 medium), 8 of them Patescibacteria, every step within 8 GB. - Next for MAGs: built (0.6.0): tree placement of unassigned bins on placement-lite
(
gtdbprep place, with gPlace), SemiBin2 (semibin2: pretrained models fit 1.5 GiB, its self-supervised training doesn’t fit 8 GB), dereplication (galah), one genome × sample abundance table (coverm_abundance) and co-occurrence statistics (cooccur). SemiBin2’s multi-sample mode is built too (0.6.1; 4.3 GiB on three samples, no better than the pretrained model there). Still open: co-assembly within 8 GB for diverse, shallow samples; a MAG-based, strain-level oral or water series forcooccur. A first real test (write-up) on 128 plaque and 136 tongue profiles did not recover TM7’s cultured host: the association came out negative (rho −0.25), partly through uneven zeros. - Instrument connection: lab instruments run vendor software on dedicated (often older Windows) PCs, and labs usually connect them through a LIMS, which parses instrument exports or uses vendor drivers. So the practical route is the LIMS or the files instruments drop for it. Open standards exist (SiLA 2; LADS, an OPC UA companion specification for lab devices). A small agent that watches sequencer run folders or speaks these protocols could start QC as soon as a run finishes. Check real instrument support first. Note that current Rust supports Windows 10+; Windows 7 needs a lower-tier target.
- Microbial observatory: environmental sensor time series plus periodic sequencing metadata. Needs field partners; long-term.