Screen qc
Legible, deterministic QC/qualification for bio-AI tool outputs — named-reason contracts + agent skills that compose with NVIDIA BioNeMo.
npx -y skills add Curtisflo/karyon --skill screen-qcAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- no licenseNo license file was found in the repository. Code published without one is not open source by default, so using it at work is a question for whoever answers licensing questions where you are.
- 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its author says it does
Copied from the file, not written here
Qualify a pooled CRISPR screen's non-hits with karyon — flag genes called "no effect" that the screen was simply under-powered to detect. Use for CRISPR screen QC, under-power detection, silent-failure / false-negative qualification, guide-count reliability, NT/NEG-calibrated screen analysis, or deciding whether a non-hit is trustworthy or just low-powered.
The file declares its own license as Apache-2.0 AND CC-BY-4.0. That is the author’s claim about this one file, and it is not the same thing as the license GitHub reports for the repository, which is listed with the other numbers below.
SKILL.md
2.9 KB, as published. Nobody here has run it
CRISPR Screen QC
A pooled-screen pipeline emits a per-gene q-value: hit / non-hit. That scalar throws away the within-gene guide structure, so a gene the screen had no power to call looks identical to a gene that genuinely has no effect. This skill reads the structure back — purely from the counts, calibrated on the non-targeting / negative controls, never from a gold standard — and qualifies every non-hit with a named reason (too few usable guides, dispersion too high, effect below the null band).
On the Wang-2014 leukemia dropout screen it flags ~53% of gold-standard silent failures at a 3% false-flag rate, and the flag is non-redundant with the FDR (|ρ|≈0.29) — it catches what the q-value misses. (Honest: the value here is legibility / qualification, not a recovery-accuracy lift.) Full method, pre-registered Q1–Q4 evaluation, and limits: docs/screen-power.md.
Install
pip install karyon
Usage
From the command line (karyon audit screen) — the Wang-2014 reference screen:
karyon audit screen --json # Q1 recall / Q2 false-flag / Q3 non-redundancy, as a JSON report
Run the reference analysis from Python:
from karyon import screen_qc
screen_qc.run() # default seeds=50; prints the qualification + the non-redundancy guard
Qualify your own screen's non-hits (counts → reasons):
from karyon import screen_qc as sq
null = sq.null_band(calib_lfcs, direction="deplete") # null band from NT/NEG controls
verdicts = sq.qualify(non_hit_genes, gene_guides, null=null) # GeneReliability per gene, with reasons
reliability_contracts(...) is the underlying contracts.ContractSet; each GeneReliability names why a
non-hit is (un)trustworthy.
Composition with NVIDIA BioNeMo
Runs downstream of an accelerated genomics pipeline (e.g. Parabricks → guide counts → a MAGeCK-style gene summary): the model/pipeline calls hits, this skill qualifies the non-hits so the agent reports "gene X: non-hit, but under-powered (4 usable guides, dispersion above threshold)" instead of a bare "no effect."
Scope (honest)
Qualification, not accuracy: it tells you which non-hits to distrust, it does not re-rank hits. Calibrated on controls, so it needs a screen with non-targeting / negative-control guides.