A new model predicts gene expression from chromatin accessibility

The scButterfly-B pipeline outperformed transformer models in predicting RNA abundance from chromatin accessibility.

The scButterfly-B pipeline outperforms two transformer models. It predicts gene expression from chromatin accessibility.

Researchers built a benchmark using 126,604 paired single-nucleus profiles. These profiles come from 19 human studies. The benchmark uses a shared vocabulary of 50,000 peaks and 3,185 genes.

Chromatin accessibility is used to predict RNA abundance. The scButterfly-B pipeline uses dual variational autoencoders. This pipeline had 3.8 times shorter training than other models.

The scButterfly-B pipeline outperformed every transformer pipeline. This includes an autoregressive Perceiver model called Omic Translator and a reformulation of Cisformer. It performed better in all sixty paired fold comparisons.

Transformer pipelines lag by 0.10-0.24 on per-gene correlation. On two withheld cell types, per-gene correlation is low. Paneth values match a cell-agnostic shuffle for every Omic Translator arm.

Key facts

  • The scButterfly-B pipeline uses dual variational autoencoders.
  • The benchmark includes 126,604 paired single-nucleus profiles from 19 human studies.
  • The scButterfly-B pipeline had 3.8 times shorter training than transformer pipelines.

All sections

Edition