Decoding sequence determinants of gene expression in diverse cellular and disease states

Code and data for the following publication:

Decoding sequence determinants of gene expression in diverse cellular and disease states

Avantika Lal*,1, Alexander Karollus*,1,2,3, Laura Gunsalus1, David Garfield4, Surag Nair1, Alex M Tseng1, M Grace Gordon5, John Blischak6, Bryce van de Geijn6, Tushar Bhangale6, Jenna L Collier1, Nathaniel Diamant1, Tommaso Biancalani1, Hector Corrada Bravo1, Gabriele Scalia1, Gokcen Eraslan1

*Equal contributions

1Biology Research | AI Development, gRED Computational Sciences, Genentech, South San Francisco, CA 94080, USA

2School of Computation, Information and Technology, Technical University of Munich, Germany

3Munich Center for Machine Learning  

4OMNI Bioinformatics and Department of Regenerative Medicine, Genentech, South San Francisco, CA 94080, USA

5 Department of Cellular and Tissue Genomics, Genentech Research and Early Development, Genentech, South San Francisco, CA 94080, USA

6 Department of Human Genetics, Genentech, South San Francisco, CA 94080, USA

Correspondence: Avantika Lal (lal.avantika@gene.com), Gokcen Eraslan (eraslan.gokcen@gene.com)


This package contains:
1. 4 replicate Decima models (rep0.ckpt, rep1.ckpt, rep2.ckpt, rep3.ckpt)
2. Decima software v0.1 (decima-v0.1.tar.gz). The latest version of this software is available at https://github.com/Genentech/decima.
3. Decima analysis code (decima-applications.tar.gz)
4. Data File 1: An .h5ad file containing Decima’s predictions for all 18,457 genes in the training, validation and test sets, along with metadata for the 18,457 genes and 8,856 pseudobulks in the dataset, and the Pearson correlation between measured and predicted expression values for each pseudobulk and each gene.
5. Data File 2: A .h5ad file containing Decima’s predicted effect sizes (log fold change in expression between alternate and reference alleles) for all 573 variants that are identified as high-confidence sc-eQTLs in the fine-mapped OneK1K dataset, as well as negative control variants, in all 8,856 pseudobulks.
6. Data File 3: A .h5ad file containing Decima’s predicted effect sizes (log fold change in expression between alternate and reference alleles) for all 837 variants that are identified as high-confidence fine-mapped GWAS causal variants, in all 201 cell types.