Anshul Kundaje

Thanks. Happy to help if you run into any issues. If the goal is differential signal prediction, would need a slightly different fine tuning set up for optimal performance

Should have used this!

And just to be clear, I have no specific comments on the other primary models implemented in the CRESTED package including the ones they developed. They look good from the paper. I'm simply stating that the "Dilated CNN" implementation in the package is NOT ChromBPNet. 4/4

q.e.d Science Critical Thinking AI for constructive criticism and science evaluation

Here's the link to the system, try it! qedscience.com
@qedscience.bsky.social

BIG ANNOUNCEMENT📣: I haven’t been this excited to be part of something new in 15 years… Thrilled to reveal the passion project I’ve been working on for the past year and a half!🙀🥳 (thread 👇)

🧬Excited to be at #ASHG2025! My lab is presenting 5 studies this week. Join us for talks and posters on #SingleCell #CRISPR, #DeepLearning & more! Thread below with all presentation details 👇

ICE took a 13-year-old they said had a gun. Local cops say he didn’t. Now he's detained 500 miles away from his Massachusetts home.

NEW — I dug into the story of ICE abducting a 13-year-old boy in Massachusetts and moving him 500 miles across state lines without notifying his mother, local cop's complicity, and how the federal government's justification for such cruelty is already falling apart:

Thanks

IceQream: Quantitative chromosome accessibility analysis using physical TF models - Nature Communications Cis regulatory elements endow genomes with sequence-encoded logic to drive cellular differentiation. Here, the authors introduce a biophysically principled sequence model that characterises complex TF...

Came across this interesting paper www.nature.com/articles/s41... that compares their model to a "Dilated CNN" model from the CRESTED package that is supposed to "ChromBPNet inspired". Performance of the dilated CNN is miserable. This is categorically NOT a ChromBPNet model 3/3

Please don't refer to models from other packages as ChromBPNet. There is much more to ChromBPNet than the architecture. The training procedure, bias correction & data sampling strategy has a huge impact on performance. 2/

Friendly reminder that there is only one official release of ChromBPNet at our lab GitHub github.com/kundajelab/chr…

There are other versions floating around eg. in the CRESTED package that are NOT anywhere close to the official implementation. 1/

ATP-dependent remodeling of chromatin condensates reveals distinct mesoscale outcomes Adenosine triphosphate (ATP)–dependent chromatin remodeling enzymes mobilize nucleosomes, but how such mobilization affects chromatin condensation is unclear. We investigate effects of two major remod...

Today I am so pleased to present our work on how chromatin remodelers affect mesoscale chromatin organization.
www.science.org/doi/10.1126/...

DNA mutagenesis driven by transcription factor competition with mismatch repair Competition between transcription factors and mismatch repair machinery drives localized hypermutation at regulatory elements, with implications for cancer and genome evolution.

I'm very pleased to announce the official publication of our lab's paper "DNA mutagenesis driven by transcription factor competition with mismatch repair" in today's issue of Cell! www.cell.com/cell/abstrac...

Trump mandates all federal agencies send email blaming Dems for potential gov’t shutdown Workers call the partisan email "a vile slap in the face."

SCOOP — The White House instructed all federal agency leadership on Tuesday to send out a mandatory message to workers blaming the impending potential government shutdown on House Democrats, The Handbasket was first to learn. Absolutely no modifications to the email were permitted.

My story:

1/ Happy to share our preprint from the Greenleaf Lab: beCasKAS, our method to directly detect CRISPR base editor off-targets in primary cells. We additionally show how non-coding edits can be triaged for epigenetic dysregulation using deep learning.

www.biorxiv.org/content/10.1...

New paper from my lab and @jshendure.bsky.social lab! Led by the brilliant @zukailiu.bsky.social and @cxqiu.bsky.social. We tackled how anterior and posterior progenitor cells cooperate to self-organize into an embryonic structure (termed AP-gastruloid). (1/n) www.biorxiv.org/content/10.1...

The GRFP announcement from NSF cuts out an entire cohort of 2nd year students from consideration, without warning. This is so deeply unfair that it warrants a formal protest from the scientific community. If someone wants to work with me to craft an open letter and solicit signatures, LMK.

The projected rise in global cancer deaths
1990 —5.9 million
2023—10.4 million
2050—18.8 million
is yet another reason we need to get primary prevention into high gear, which is now possible but not getting done
www.thelancet.com/journals/lan...

And I think this is just the tip of the iceberg. There will be many ways to push this further. Really huge breakthrough IMO from my first reading. Can't wait to use this & build on it. Huge congrats to the team! 3/3

Also an absolute showcase for what self supervised learning can do when done right. IMO, this is the first paper that really shows the power of self supervised genomic language models. 2/

This is truly an incredible breakthrough IMO. Really exemplifies what you get when deep domain expertise (popgen/evolution/disease genetics in this case) fuses with cleverly crafted ML. What u get r sleek, well thought out architectures that absolutely destroy the behemoths. Wow!! 1/

We are excited to share GPN-Star, a cost-effective, biologically grounded genomic language modeling framework that achieves state-of-the-art performance across a wide range of variant effect prediction tasks relevant to human genetics.
www.biorxiv.org/content/10.1...
(1/n)

Excited to share our new study:
Single-cell multiomics reveals archetypal regulatory programs shared across CD4 and CD8 T cell subsets in viral infection
bioRxiv: doi.org/10.1101/2025...

🧵 Key findings below 👇

But lets be clear that glam journal publishing is a branding & marketing exercise. It is not in any way a step towards greater transparency, accountability & credibility. 12/12

... to not support the gate keeping system followed by glam journal publishing. But I also get that for them it is a very cheap (only 13K for gold open access) but effective branding tool. So I definitely do not blame them for taking advantage of it. 11/

So yes I would encourage them to release their models open weights, open source, with fabulous documentation & respond to community review. But also ... 10/

I would go further & claim that companies are especially are not bound by any of the constraints that masochistic academics impose on themselves by tying their futures to exploitative & wasteful gate keeping publishing systems. 9/

So kudos to Deepseek for being super transparent with their tech, models, documentation. But publishing this work in Nature or any other glam journal is entirely a branding and marketing exercise. Lets not pretend otherwise. 8/

Occasionally they do live upto the hype. But most of the time that's not the case. The whole gate keeping exercise of glam journals is based on vibes about what is "hot" at the moment. 7/

Lastly, I don't know how anyone can claim that publishing a paper in Nature is a way to dial back hype. It is literally a branding exercise. Papers published in glam journals are the most overhyped of all scientific literature. 6/