New here? Read ARCHITECTURE.md first — what each file is for, how the agent picks which one to read, and how much of it has actually been tested on the cluster.
Claude Science skills for IGB Biocluster 4 at the University of Illinois Urbana-Champaign. Import them and a new session already knows the cluster: which partitions that account can use, which databases are already mirrored, which module versions pair with which, and which failures are silent.
Start here: INSTALL.md — three steps, about ten minutes.
Rebuilt for Biocluster 4 on 2026-11-01. Biocluster 4 replaced Biocluster 3 in October 2026: new login node (
biologin-0), new hardware, Slurm 25.05.1, and a module stack that shrank from 927 names to 173. Every number in this repository was re-measured on the new cluster. If you are reading a fork or a cached copy that mentionsbiologin-3, theIGB-gcc-8.2.0toolchain, or/scratch, it is out of date.
| Skill | Load it for |
|---|---|
igb-cluster |
always — partitions, site policy, module rules, job skeleton, audit commands |
igb-reference-data |
before downloading any reference genome, annotation or database |
igb-structure-prediction |
AlphaFold 2/3, docking, molecular dynamics |
igb-gpu-ml |
anything needing a GPU; PyTorch, and the foundation models installed here |
igb-single-cell |
Cell Ranger, Space Ranger, nanopore and PacBio |
igb-workflows |
Nextflow, Apptainer/Singularity, conda |
igb-ngs-pipelines |
ordinary sequencing work, plus an index of all 173 modules |
igb-cluster is the hub and is always loaded first; the other six assume it. It also
ships kernel.py, which turns the site rules into callable helpers — GitHub import
excludes sidecar files, so that file is not installed by the steps in
INSTALL.md and the skill documents are written to stand alone without it.
They are the site layer only — which module, which database, which partition,
measured wall-clock, known failures. For the science itself the agent loads the tool's
own skill (alphafold2, boltz, scvi-tools, proteinmpnn, …) alongside.
Reference databases are already here, as modules. 29 of them — BLAST nt/nr,
GenBank, RefSeq, UniProt, Pfam, InterPro, GTDB, SILVA, Kraken2, CheckM, BUSCO, the GATK
bundle, the AlphaFold 2 and 3 databases. None appears on the IGB Applications wiki
page. Compute nodes have no internet, so loading the module is not the convenient
route — it is the only one. The variable name is not guessable either: EBROOT plus the
module name uppercased with every - replaced by MIN, so alphafold3-db sets
$EBROOTALPHAFOLD3MINDB. Four of the 29 modulefiles point at a directory that does not
exist; the corrected paths are in igb-reference-data.
The software stack shrank by 80% in the migration. 927 module names became 173.
SPAdes, Trinity, salmon, scanpy, snakemake, rosetta, pymol, rdkit,
ViennaRNA, RepeatMasker, cromwell and toil are all gone. A workflow that ran
on Biocluster 3 may have nothing to load. Always module avail <name> before promising a
tool — and say so plainly when it is absent rather than substituting something silently.
TensorFlow cannot see the GPU. Measured on 2026-10-31: TensorFlow/2.21.0 reports
GPUs visible: 0 both alone and with CUDA/13.1.0 + cuDNN/9.17.1.4 loaded. The
cluster ships only CUDA 13.x and TF 2.21 is built against CUDA 12, so no module pairing
fixes it. A TensorFlow job will train on the CPU, slowly, with no error. Use PyTorch,
which works correctly against the A40.
Every claim is tagged, so the agent knows what to trust:
- A — verified on this cluster; the recipe is the one that worked, with timings.
- B — documented: the module exists and the invocation comes from
module showor IGB's docs, but nobody has run it here. - C — listed: the module is present. Nothing further is claimed.
As of 2026-11-01, tier A: the hub, the partition map, all 29 reference-database paths, GTDB-Tk against r226, the GPU framework results, the AlphaFold 3 recipe, and the four foundation-model installs (ESM-2, Segment Anything, DINOv2, Evo 2). Tier B: the 10x pipelines and Nextflow, where the modules and invocations are verified but no full run has been done here. Tier C: most of the 88 NGS modules.
What remains at C cannot be promoted from a desk. Those claims are about tools producing correct output on real input, and the only way to settle one is to run it on the cluster with your own data. If you do, open a PR adding the command you used and the runtime you saw — that is how this stops being a catalogue and becomes a manual.
Verification sweeps have repeatedly found claims that were wrong, not merely
unverified: reference databases whose module variable points at a path that does not
exist, an nvidia-smi preflight that would have aborted every GPU job, and a PyTorch
build that passed torch.cuda.is_available() and then crashed. Assume the tier C rows
contain errors of the same kind.
Rebuilt from live measurement on biologin-0.igb.illinois.edu (Biocluster 4,
Slurm 25.05.1) on 2026-10-31 and 2026-11-01: module avail (173 names, 212
versions), scontrol show partition against an account's Unix groups, all 29 database
modules opened and their contents listed, a GPU framework probe on an A40, and an
AlphaFold 3 monomer run. The full module inventory ships inside the repo at
skills/igb-ngs-pipelines/references/module-index.csv.