Happy New Year! I’m still catching up on papers from my late 2024 backlog.

Background Bluesky, atrrr, local LLMs I’ve written a few posts lately about Bluesky — first, Bluesky for Science, about Bluesky as a home for Science Twitter expats after the mass eXodus, another on using the atrrr package to expand your Bluesky network. I’ve also spent some time looking at R packages to provide an interface to Ollama.

I had good intentions to give NaNoWriMo a try this year but didn’t get very far. Instead I gave OpenAI’s Creative Writing Coach GPT a try for a (very) short story I had in mind, inspired by my frustration trying to access closed-access research articles for a review article I’m preparing.

The Baader–Meinhof phenomenon (aka the frequency illusion) is the name for that thing that happens when you buy a new car, and suddenly you notice that same model car everywhere you drive.

This week’s recap highlights the Evo model for sequence modeling and design, biomedical discovery with AI agents, improving bioinformatics software quality through teamwork, a new tool from Brent Pedersen and Aaron Quinlan (vcfexpress) for filtering and formatting VCFs with Lua expressions, a new paper about the NHGRI-EBI GWAS Catalog, and a review paper on designing and engineering synthetic genomes.

A few days ago I wrote about translating R package help documentation using a local LLM (e.g. llama3.x)… …when Mick Watson commented: I was already thinking of wiring up something like this using local AI models — something to summarize podcasts, conference recordings, etc. The relatively new (as of this writing) Gemini 2.0 Flash model will do this for you for YouTube videos. But what if you wanted to do this offline using a local LLM?

Last week I posted about a web app that turns a GitHub repo into a single text file for LLM-friendly input. This is great for capturing LLM-friendly text from a GitHub repo, but what about any other arbitrary website or PDF? I was catching up on Simon Willison’s newsletter reading about an app he made with Claude artifacts that uses the Jina Reader API to generate Markdown from a website. You don’t need to use the API to do this.

Using LLMs in R Most of the developer tooling for AI/LLM training and evaluation is Python-centric, but just over the past few months we’ve seen a surge of new tooling for AI/LLM applications for the R ecosystem. ollamar and rollama provide wrappers around the Ollama API allowing you to run LLMs locally on your machine.

This week’s recap highlights a new way to turn Nextflow pipelines into web apps, DRAGEN for fast and accurate variant calling, machine-guided design of cell-type-targeting cis-regulatory elements, a Nextflow pipeline for identifying and classifying protein kinases, a new language model for single cell perturbations that integrates knowledge from literature, GeneCards, etc., and a new method for scalable protein design in a relaxed sequence

If you use ChatGPT, Claude, or even some local model through Ollama or HuggingFace Assistants, you’ll know that the chat interface makes it challenging to feed in an entire repo like a Python or R package, because functions, tests, etc. can be scattered across many files throughout a repo.

This week’s recap highlights the WorkflowHub registry for computational workflows, building a virtual cell with AI, a review on bioinformatics methods for prioritizing causal genetic variants in candidate regions, a benchmarking study showing deep learning methods are best for variant calling in bacterial nanopore sequencing, and a new ML model from researchers at Genentech for predicting cell-type- and condition-specific gene expression across