Johannes Hingerl

AlphaGenome is out in @nature today along with model weights! 🧬 📄 Paper: nature.com/articles/s4158… 💻 Weights: github.com/google-deepmin… Getting here wasn’t a straight path. We sat down @googledeepmind to discuss the story behind the model, paper & API: youtu.be/V8lhUqKqzUc

Excited to share Nona: a unifying multimodal masking framework for functional genomics. Models for DNA have evolved along separate paths: sequence-to-function (AlphaGenome), language models (Evo2), and generative models (DDSM). Can these be unified under a single paradigm? 1/15

Many thanks to co-first author @lauradmartens for this fun collaboration, and to @AlexKarollus, @trevmanz, @JD_Buenrostro, @fabian_theis and @gagneurlab for their contributions and guidance!

Paper link: nature.com/articles/s4159… Scooby github: github.com/gagneurlab/sco…

I'm excited that our paper, "scooby: modeling multimodal genomic profiles from DNA sequence at single-cell resolution," has been published in Nature Methods! scooby is a new deep-learning framework to understand how DNA sequence shapes gene expression in individual, single cells.

Big thanks to co-authors @AlexKarollus & @gagneurlab , and to Borzoi authors @jjohlin @drklly et al! Flashzoi is available on github and on pip: github.com/johahi/borzoi-….

The speedup does not come at the cost of accuracy: Flashzoi matches or improves upon Borzoi’s performance across benchmarks, including RNA-seq coverage prediction, variant effect prediction (GTEx eQTLs), and enhancer-promoter linking.

Happy to share that Flashzoi is now published! We enhanced Borzoi with RoPE & FlashAttention for >3x faster training/inference & 2.4x reduction in memory usage. This brings large-scale genomic analysis and fine-tuning within reach of academic budgets. 📄: doi.org/10.1093/bioinf…

Recent performance prediction models have made great advances in accurately predicting the execution time of database queries. See how we achieved the same accuracy with a 10,000x faster model, T3. We are happy to present our work on T3 at SIGMOD'25. db.in.tum.de/~rieger/papers…

@i000 @missing_a_rib @anshulkundaje @jmschreiber91 We only load the weights for the convs; we completely re-initialize all transformer weights (see Fig 1b). While we managed to train a Borzoi from scratch w similar performance on held out seqs/ VEP, a Flashzoi from scratch w same training hparams was subpar. Rotary is the diff

@missing_a_rib @anshulkundaje @jmschreiber91 So you can have different positional encoding yet still be weight-compatible (per authors flashzoi can load borzoi weights, I haven't tested)?

Just want to re-iterate here that bpnet-lite does not replicate the ChromBPNet repo. It is an excellent package for development but it doesn't implement several important functionality eg. upsampling peaks or bias correction checkpoint controls etc. 1/

1/ DNA sequence models like Borzoi predict gene expression and variant effects across 1000s of tissues — but what if your data comes from a custom experiment? @drklly @jjohlin and I propose a lightweight solution: parameter-efficient fine-tuning (PEFT). biorxiv.org/content/10.110…

Can DNA sequence models predict mutations affecting human traits? We introduce TraitGym, a curated benchmark of causal regulatory variants for 113 Mendelian & 83 complex traits, and evaluate functional genomics and DNA language models. Joint work w/ @gokcen and @yun_s_song 🧵👇

The Borzoi manuscript is now out in Nature Genetics: doi.org/10.1038/s41588… Borzoi predicts RNA-seq profiles in many tissues & cell types from DNA sequence as its only input. With it, we can score the impact of genetic variants on a number of gene-regulatory functions. 1/

Thanks to @AlexKarollus and @gagneurlab for their input, and to @jjohlin, @drklly et al for sharing their model!

On GTEx eQTLs, Flashzoi showed significantly improved correlation with observed variant effects and maintained Borzoi's performance in eQTL prioritization. Explore more results in the paper, find the code at github.com/johahi/borzoi-…, and models at huggingface.co/johahi/.

Flashzoi offers a slight but consistent improvement in the accuracy of genomic profile predictions compared to Borzoi, while delivering speed gains of up to 3x for both prediction (inference) and training or fine-tuning.

Introducing Flashzoi⚡! We’ve upgraded the Borzoi model with rotary pos. encodings and FlashAttention, resulting in a 3x speedup with similar or better accuracy for faster variant effect prediction or model development, and more efficient genomic analysis biorxiv.org/content/10.110…

From computational models of the splicing code to regulatory mechanisms and therapeutic implications go.nature.com/4gPju63 #Review by @tiny_captain, @OscarWilkins16, @nils_f_wagner, Julien @gagneurlab & Jernej @ule_lab @HelmholtzMunich @TU_Muenchen @TheCrick @KingsCollegeLon