CHARR does a mean of per-site estimates of contamination; for low coverage, it might be more robust to sum counts across sites and divide at the end. It's trivial to implement and might work well for your case
ah, ok. that will be too low for somalier.
the writeup is here: github.com/brentp/somal...
the methods match conpair and CHARR pretty faithfully
if you give somalier a try for contamination and it doesn't do what you want, I'd like to hear about it. ๐
New version of somalier is up: github.com/brentp/somal...
Has a CHARR contamination estimator, a CONPAIR pair-wise inter-sample contaminatoin estimator, an AB/depth plotter and other improvements to make it of greater utility in cancer samples.
somalier is a pretty cool idea and implementation. it's one of my tools that I think deserves more use. ;)
the next version will have better contamination estimates (CHARR-like) and pairwise contamination (CONPAIR-like) and concordance of homozygous sites. It will also have a tool to plot sketches like this as interactive html where we can see large chromosomal events.
as part of my work with isabl genomics, I've been updating somalier with more info for cancer samples (contamination and concordance). somalier extracts small "sketches" of ~20K sites from a BAM/CRAM/VCF and then does rapid all-vs-all kinship checks. (1/n)
the codex desktop app was only available as a Mac DMG. A while ago, I downloaded it and told codex cli to make it work on linux. It did. Now I use that nearly exclusively.
gist.github.com/brentp/b6bd8...
yes, that was the reasoning
I don't think so. But a skill or MCP could work as well.
Again, I think making a custom GPT (or similar) as I did for bedder will be the default for most software shortly. If it can't answer common questions, improve the docs and refresh and the maintenance burden of answering user questions goes down.
Brent Pedersen, Mitchell Vollger and I have "posted" our preprint of the manuscript describing bedder, a complement to the functionality of bedtools. The "preprint server" we have chosen is google docs because it was rejected by biorxiv.
docs.google.com/document/d/1...
Here is the reply "bioRxiv is intended for complete research papers with new data/analysis rather than announcements of tools/resources/reagents and your submission falls into this category"
I would guess that soon, every tool will have one of these. It is more interactive than just documentation and better than waiting for a grumpy maintainer to answer the same question in a github issue or personal email.
Feeding my markdown docs to this, then using it in anger shows where the docs need to be bolstered and provides a nice reference for users:
help.openai.com/en/articles/...
here's one for bedder: chatgpt.com/g/g-69a6027f...
I had a similar, less principled way of doing this, but this does seem to work quite well after trying a couple of large changes with it: boristane.com/blog/how-i-u...
(I use codex, but same idea)
AGENTS.md > skills
vercel.com/blog/agents-...
it does support more than 2 sets of intervals, but now, a single --b-piece controls all of them. I'll have a look at this and your other comments. thank you!
scratch that. better to use gemini-cli, turn on preview features and then get gemini 3.0 pro and flash for free for a decent amount each day.
cerebras is so fast it makes the human <-> agent loop (or one-shot) very quick
and crush will also detect your GEMINI/OPENAI/OPENROUTER/CLAUDE env API vars as well
crush coding agent with a free plan from cerebras using glm4.7 model is an easy way to start with coding cli tools.
inference-docs.cerebras.ai/support/rate...
github.com/charmbracele...
crush is static binary that detects your CEREBRAS_API_KEY. GLM 4.7 isn't as good as Opus, but it is very solid.
another reminder about htsvcf which allows either
1. rust library to support javascript expressions (filtering/modification) on variants from VCF/BCF
2. javascript (node/bun) library for reading/writing/modifying variants from vcf/bcf.
both very performant.
github.com/brentp/htsvcf
a post about how, with LLMs, I wrote a new set of libraries.
Wrapping htslib is my hello-world but doing it in rust to allow javascript expressions handled by V8 was more substantial project.
The tools and LLMs are now getting quite good:
brentp.github.io/latest/blog/...
let me know what you think
my LLM-coded version of the single-page year calendar that's been going around. This one can be scripted with javascript to highlight cells, change the borders and texts. It also allows saving and sharing via url:
brentp.github.io/single-page-...
with code here:
github.com/brentp/singl...
exactly! and with Plan mode in opencode (or architect in kilo code) this becomes even more mind-blowing
so Opus 4.5 is excellent for coding. I've been using these LLMs for coding for a while and it felt like slow improvement with each new model and sometimes more work to guide the model. Opus 4.5 just works. It solves hard problems and has good taste.
Anyone else notice this?
I'll write more about this later, but as an experiment in LLM coding (GPT-5.2 and Opus 4.5), I wrote a JS wrapper for rust-htslib to read/write modify VCFs:
www.npmjs.com/package/htsvcf
works on linux (x64/aarch64) and OSX.
brentp.github.io/htsvcf/lates...
Get bedder v0.1.8 and r๐ฒ๐ฎ๐ฑ ๐๐ต๐ฒ ๐ฑ๐ผ๐ฐ๐๐บ๐ฒ๐ป๐๐ฎ๐๐ถ๐ผ๐ป here: github.com/quinlan-lab/...
We are looking for researchers to kick the tires, integrate bedder into their pipelines, and provide feedback on the Python functions, performance, and overall user experience! Please share and give us your feedback.
We are thrilled to announce the first official release (v0.1.8) of #๐ฏ๐ฒ๐ฑ๐ฑ๐ฒ๐ฟ, the successor to one of our flagship tool, #๐ฏ๐ฒ๐ฑ๐๐ผ๐ผ๐น๐! Based on ideas we conceived of long ago (!), this was achieved thanks to the dedication of Brent Pedersen.
1/n
thank you. I'll have a look next week!