
Super happy to share that the first paper from my postdoc is now out on bioRxiv (biorxiv.org/content/10.64898…). We trained ProCapNet models on all ENCODE PRO-cap experiments (>200 datasets), representing all major human organs/tissues.

Machine learning, RNA transcription, genomics, statistical genetics. Postdoc @StanfordMed with @anshulkundaje. Previously PhD student with @charlesdanko

Super happy to share that the first paper from my postdoc is now out on bioRxiv (biorxiv.org/content/10.64898…). We trained ProCapNet models on all ENCODE PRO-cap experiments (>200 datasets), representing all major human organs/tissues.

Genomic tessellationsは、DNA配列を構成単位と、その長さ・位相・例外に分けて表現する手法。反復配列と非反復配列を統一的に扱い、反復長の違いで見え難い相同性を捉える。構成単位を照合するtBWTにより反復配列を拡張し、既存マッパーと同等の精度を保ちながら大幅な高速化を実現する
Genomic tessellations biorxiv.org/content/10.64898…

I'm super happy to have gotten paid to *not* sit through journal peer review! And to see funders reward maintenance/optimization of old research code. Old tweet thread on my project that got funded:
We funded 10 grad students to publish their hidden science. The work fell into three groups: results with no home in a paper, datasets too valuable to sit on a server, and tools that help other scientists work. See the results ➡ asterainstitute.substack.com…

We funded 10 grad students to publish their hidden science. The work fell into three groups: results with no home in a paper, datasets too valuable to sit on a server, and tools that help other scientists work. See the results ➡ asterainstitute.substack.com…

📣 MLCB + MLSB are joining forces under MLBio.org ! 🧬🤖 The first workshop will be held 2027, Join the MLBio mailing list to stay updated on calls for papers, deadlines and more info on the first event! groups.google.com/g/mlbio

Delighted to announce our brand-new PhD Program in Biomedical Informatics, Modeling, and AI (#BIMA). Our mission is to train computational scientists for biomedicine's #AI era. Applications due Dec 1 cuanschutz.edu/graduate-prog…

Introducing Manimble: a free, open-source slide editor in which every slide is a Manim program (so underlying the deck is Python). It alows for editing like in Keynote or PowerPoint while seamlessly incorporating AI edits / animations. pachterlab.github.io/Manimbl… 1/🧵

Distinct extraembryonic transcription factors converge on a shared DNA methylation-Polycomb state dlvr.it/TWL0T3

Can a person's T-cell repertoire from a blood draw diagnose disease? Many methods say yes, but each was developed on different cohorts, splits, and metrics. Excited to present BenchRep-T, a benchmark comparing methods head-to-head, at #NeurIPS2026 Evaluations & Datasets 1/🧵

A benefit of our Genomics×AI blog is posts are living, versioned, updatable documents. So I've updated our enhancer benchmark with NTv3's revised release, Cerberus and AlphaGenome 2. Cerberus ties Borzoi overall but is worse on proximal and (maybe) better on distal enhancers 1/2
New post w/@pkoo562: we test Enformer, Borzoi, NTv3 & AlphaGenome on K562 CRISPRi enhancers, extending Karollus 2023. Performance has improved with newer seq2func models, but all still underpredict distal CREs. Distal perturbations remain an open problem! genomicsxai.github.io/blogs/…

Online Now: A sequence-encoded promoter-proximal super pause stabilizes an offline RNA polymerase II state dlvr.it/TWHrfz

On an A100, DLS runs ~1.5x faster in FP32 and 2.4x faster in TF32 compared to tangermeme’s implementation. Both scoring and DLS optimizations have been verified to be within numerical precision of the naive/general implementations.

promoterai-torch’s DLS also caches each sequence’s forward, backprops only the k seq–ref pairs, recomputes FFN activations per block & auto-sizes chunks to free memory. As a result, DLS now fits on the 3080 10 GB I have in my gaming PC.

promoterai-torch now ships a custom DLS engine with significant speed and VRAM improvements. Notably, it incorporates some of the “fused DeepLIFT” ideas implemented in the most recent version of Cherimoya (github.com/jmschrei/cherimoy…).

Second, while I built the promoterai-torch port to be compatible with tangermeme’s DeepLIFT/SHAP implementation, it ended up being extremely expensive to run (>60GB VRAM on an A100).

The torch port is still slightly slower than the original TF at non-SNV scoring due to some unavoidable overhead, but with the new optimizations, SNV scoring is now >3x faster than the original package.