iiinova — where your digital self keeps evolving

Reposted by Andreas Gschwind

Jesse Engreitz @jengreitz.bsky.social · 04/08/2026

Update: scE2G is now online at Nature Genetics! Since the preprint, we added a new benchmark on a fully held-out CRISPR dataset. scE2G continues to show state-of-the-art performance, reinforcing that it generalizes well nature.com/articles/s41588-026-02695-8 1/

Held-out CRISPR benchmark: n=189 positives, 4,175 tested pairs from 5 cell types. Weighted AUPRC bar chart with scE2G-Multiome and scE2G-ATAC highest, above Distance to TSS, ABC, and 8 other single-cell methods.

13514

Andreas Gschwind @argschwind.bsky.social · 15/07/2026

This was a wonderful community effort with countless people across ENCODE and beyond. Many thanks to everyone I had the joy of working with on this long, exciting journey! 18/18

030

Andreas Gschwind @argschwind.bsky.social · 15/07/2026

Summary: We present a new approach to map enhancer-gene regulatory interactions, link variants to genes and cell types, and interpret models to learn how enhancers work. Plus a framework combining perturbation screens with machine learning to build future, increasingly accurate models. 17/

110

Andreas Gschwind @argschwind.bsky.social · 15/07/2026

We believe the models, benchmarks and E–G maps will provide a lasting resource. Predictions for 1,458 biosamples: www.encodeproject.org Model: github.com/EngreitzLab/ENCODE_rE2G Benchmarking pipelines: github.com/EngreitzLab/... github.com/EngreitzLab/... github.com/Deylab999MSK... 16/

encodeproject.org

132

Andreas Gschwind @argschwind.bsky.social · 15/07/2026

What assays should we prioritize to build better E–G models in new cell types? Start with DNase-seq (already strong), then add H3K27ac ChIP-seq and/or cell-type-specific Hi-C to push performance higher. Overall, DNase-seq seems to beat ATAC as an enhancer activity feature. 15/

110

Andreas Gschwind @argschwind.bsky.social · 15/07/2026

On the point of enhancer-enhancer interactions: We followed up to experimentally examine whether enhancers act additively or non-additively on gene expression in CRISPR data. Enhancers in close 3D proximity can influence each other's activity, producing super-additive effects on expression. 14/

110

Andreas Gschwind @argschwind.bsky.social · 15/07/2026

Top E–G features include: 1. ABC score (combination of enhancer activity and 3D E-P contact) 2. Promoter class: Housekeeping genes are less responsive to enhancers 3. Enhancer-enhancer interactions: Nearby enhancers increase enhancer activity. 13/

110

Andreas Gschwind @argschwind.bsky.social · 15/07/2026

How does the model work? It’s a simple and easy to interpret logistic regression trained on CRISPR data with hold-one-chromosome-out CV. Using a best-subset feature selection, the final DNase-only model keeps just 8 features. Which E–G features matter most? 12/

130

Andreas Gschwind @argschwind.bsky.social · 15/07/2026

Example: we find a fine-mapped causal variant for mean corpuscular hemoglobin linked to the gene CPEB4 by ENCODE-rE2G+PoPs only in hematopoietic progenitors and not in other closely related blood cell types like B-cells, T-cells and myeloid progenitors. 11/

110

Andreas Gschwind @argschwind.bsky.social · 15/07/2026

How can we use ENCODE-rE2G to identify target genes of GWAS variants? By combining ENCODE-rE2G with PoPS, we achieve high precision at linking non-coding GWAS variants to genes. This not only prioritizes disease genes, but also provides information on their regulation. 10/

110

Andreas Gschwind @argschwind.bsky.social · 15/07/2026

Importantly, ENCODE-rE2G generalizes across cell types and prediction tasks: its favorable performance holds up on a separate held-out CRISPR test set across 5 cell types, and on fine-mapped eQTLs and GWAS variants in new cell types. 9/

110

Andreas Gschwind @argschwind.bsky.social · 15/07/2026

We tested two versions of ENCODE-rE2G: one using DNase-only features that can be applied across many cell types, and one with an extended feature set derived from many different assays. Result: ENCODE-rE2G achieves state-of-the-art performance across our benchmarks. 8/

120

Andreas Gschwind @argschwind.bsky.social · 15/07/2026

To evaluate these models versus others, we needed a comprehensive set of benchmarks. We worked together to harmonize benchmarks based on: - Predicting E-G interactions from CRISPR experiments - Enrichment of eQTL variants - Enrichment of GWAS variants + linking them to genes 7/

110

Andreas Gschwind @argschwind.bsky.social · 15/07/2026

The challenge: many models, many features (ABC, correlation, chromatin assays…), all hard to compare. Which do you pick? Our solution: CRISPR data as ground truth to train a supervised model that combines the best features, then apply it across cell types to build genome-wide regulatory maps. 6/

110

Andreas Gschwind @argschwind.bsky.social · 15/07/2026

Can we use these data to build accurate enhancer-gene (E-G) maps across cell types as a resource for the community? We present: 1. A new predictive modeling framework (ENCODE-rE2G) 2. Benchmarking pipelines 3. >92 million E–G links in 1,458 biosamples (369 cell types & tissues) Details ahead… 5/

111

Andreas Gschwind @argschwind.bsky.social · 15/07/2026

For the last 6+ years, our team within the ENCODE Consortium has been considering how to use the data and insights from ENCODE — including deep epigenomic, 3D contact, and CRISPR perturbation datasets — to identify enhancers and their target genes 4/

110

Andreas Gschwind @argschwind.bsky.social · 15/07/2026

The central question: which enhancers regulate which genes and in which cell types? Enhancers act over long distances, can regulate multiple genes, and are cell type specific. We need genome-wide maps to look up any gene’s enhancers and interpret noncoding disease-linked variants. 3/

111

Andreas Gschwind @argschwind.bsky.social · 15/07/2026

A massive team effort with @kmualim.bsky.social , @karbalayghareh.bsky.social, @mayayayas.bsky.social , @kanishkadey.bsky.social, Evelyn Jagoda, Ramil Nurtdinov, Wang Xi, Lars Steinmetz, @anshulkundaje.bsky.social , @jengreitz.bsky.social and many others across ENCODE. 6+ years in the making! 2/

ualim.bsky.social

Bluesky

142

Reposted by Andreas Gschwind

Annique Claringbould @anniquec.bsky.social · 12/03/2026

🧬 How do immune disease-relevant variants affect gene regulatory networks in CD4+ T cells? 🧪🖥️ We coupled two large-scale CRISPRi screens (>4M cells) to map the downstream cascades of thousands of SNPs More details in the thread below 👇 or in the article 📖 on biorxiv tinyurl.com/CD4screens

12811

Reposted by Andreas Gschwind

Jesse Engreitz @jengreitz.bsky.social · 19/09/2025

New preprint from our lab! What can we learn about the properties of gene regulatory elements by CRISPR’ing a random set of accessible sites in human cells? Find out here: www.biorxiv.org/content/10.1... 👇 1/

bioRxiv - An unbiased survey of distal element-gene regulatory interactions with direct-capture targeted Perturb-seq

16018

Reposted by Andreas Gschwind

Jesse Engreitz @jengreitz.bsky.social · 18/09/2025

Excited for a major milestone in our efforts to map enhancers and interpret variants in the human genome: The E2G Portal! e2g.stanford.edu This collates our predictions of enhancer-gene regulatory interactions across >1,600 cell types and tissues. Uses cases 👇 1/

39242

Reposted by Andreas Gschwind

Wei-Lin Qiu @613weilin.bsky.social · 25/11/2024

Excited to share our latest preprint on scE2G – a new model to link enhancers to target genes using single-cell data – with state-of-the-art performance across multiple perturbation benchmarks. biorxiv.org/cgi/content/... Read more below! 1/12

biorxiv.org

Mapping enhancer-gene regulatory interactions from single-cell data

Mapping enhancers and their target genes in specific cell types is crucial for understanding gene regulation and human disease genetics. However, accurately predicting enhancer-gene regulatory interac...

14320