If they behave anything like the introverts I know, we won't have to worry about robo spam callers anymore because the LLMs will have to psyche themselves up for hours before they can make the call.

Hopefully this'll make our models a bit more accessible/usable for all the PyTorch folks out there (including me now 😂).

(1) small tweaks to bpnetlite's BPNet class to reduce memory footprint & enable training on large, multi-individual datasets and (2) a full reimplementation of CLIPNET in PT using BPNet as its backbone and w/ 2114 bp context length (instead of the original's 1 kb).

Read the prompt again. I guess some of those papers aren't directly related for gene expression specifically, but successes in epigenomic VEP should be part of the discussion of gene expression VEP IMO.

It's missing a bunch of important papers too. Fine-tuning work from Ioannidis & Pollard labs, DeepAllele, BigRNA, ChromBPNet, my stuff (obvious COI but it's relevant). Not a cutoff date issue since it's got the pretty recent review from Ioannidis lab.

Positively transformative *raises pinky*

I guess that's true, but using language models to learn a cell state autoencoder feels a bit overkill to me.

It's also not clear to me why we should expect MLM to work for gene expression vectors? Like at least for PLMs and DNALMs I can see the language analogy (and PLMs actually work) and how we might learn something from MLM training. What's being learned from ordered gene expression vectors?

🎉🎉🎉

My next pull request will make wingdings the default font for sequence logos.

I think you mean this review? The Stark paper is a cool one, but definitely an example of conflating design method w/ oracle

www.nature.com/articles/s44...