(Rather than the models being a cross-fold ensemble, they were trained with identical train/test splits and only the random weight initialization and batched sequence order varied.) 8/

Finally, we apologize for the confusion and wasted time early users experienced due to a training script bug that caused us to misstate the train/valid/test splits of the model in the preprint. The published manuscript and github now accurately describe the split. 7/

- Flashzoi ( https://t.co/5FGuGawnPx) : Efficient borzoi in pytorch. - Decima ( https://t.co/0X0wBt417X) : Transfer-learning to single-cell atlas expression data. - gReLU ( https://t.co/raLnBFWTZl) : Software suite for training, interpretation, design. 6/

Second, a number of excellent follow-up tools & research by other groups has emerged, using the pre-trained Borzoi model weights as backbone. Here is a (possibly incomplete) list of them: - scooby ( https://t.co/OKN7vk6Ojv) : Transfer-learning to single-cell multiome data. 5/

This highlights the importance of all the publicly available resources we used for training, and we want to give a shoutout to those consortia. ENCODE ( https://t.co/WL3xjX8iY5), GTEx ( https://t.co/pV7xmfe6QI), FANTOM ( https://t.co/kXgZOrNE76), CatLAS ( https://t.co/b4aHenjNiY) 4/

In the revision, we performed ablation experiments where Borzoi was re-retrained to predict RNA-seq coverage alone, without any auxiliary data it was originally trained with (e.g. DNase). As it turns out, the model’s generalization performance for RNA-seq drops noticeably. 3/

Original tweet (preprint): https://t.co/LSrAq56QMX There are a couple of developments since the preprint worth highlighting. 2/

The Borzoi manuscript is now out in Nature Genetics: https://t.co/e6WcfztXx3 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/

Check our new paper “Predicting RNA-seq coverage from DNA sequence as a unifying model of gene regulation”.

New preprint w/@anshulkundaje introducing CPA-Perturb-seq! We systematically perturb regulators of cleavage and polyadenylation, and explore post-transcriptional changes at single-cell resolution. Led by @mh_kowalski @harm__w and @jjohlin (🧵) https://t.co/GQwKMQhlpX

Very excited for MP3-seq, a new high-throughput Y2H approach we use to screen de novo protein heterodimer interactions. Fantastic work by Alex Baryshev, Alyssa La Fleur, @benjaminbgroves,@CirstynMichel, David Baker @UWproteindesign, @AjasjaLjubetic https://t.co/5JPKw5Ly2m

Excited to highlight @Calico’s 2023 summer internship program, which my group will be participating in! If you’re interested in gaining experience with deep learning models in regulatory genomics, consider applying to join us here:

CaRPool-seq from @satijalab, @nevillesanjana and colleagues makes use of the RNA-targeting CRISPR-Cas13d system to perform combinatorial perturbations in single-cell screens. https://t.co/4AiuMSkR2H

@vagar112 & @drklly describe Saluki, which is capable of predicting the effects of mRNA sequences and genetic variants on mRNA stability 50% more accurately relative to existing models in mammals. https://t.co/gfvVB6YiT1

In this peer-reviewed version of the paper, @SamanthaKoplik experimentally assayed clinically relevant 3’ UTR variants in an MPRA in multiple cell lines and validated many of our predictions.

APARENT2, our latest model for scoring and interpreting the effects of variants on 3’ UTR polyadenylation, was published in Genome Biology: https://t.co/9GtahLK5mo. Great collaboration with @seeliglab, @anshulkundaje and @SamanthaKoplik.

Presenting this work at #bog22 Thursday May 12th 9am ET!

Figure 5: When performing in-silico mutagenesis of all human polyA signals, we find that loss-of-polyA is depleted among common variants. In contrast, we detect an enrichment of gain-of-polyA mutations in individuals with Autism in WGS cohort data. (4/4)

Figure 3: APARENT2’s variant predictions correlate strongly with 3’ aQTL effect sizes from GTEx. By learning residual models of tissue-specific regulation from endogenous data, we can better predict tissue-specific aQTLs. (3/4)

Figure 2: A deep residual NN called APARENT2 predicts polyA variant effect sizes measured in MPRAs more accurately than previous models. Mask-based attribution allows us to catalogue epistatic feature interactions disrupted by clinically relevant mutations. (2/4)