AI and the World's Hard Problems · Life sciences · arXiv · 2026-10-22
An autoencoder outperforms two transformers on prediction of gene expression from chromatin accessibility
Models predicting RNA abundance from chromatin accessibility are rarely compared under the same features, targets and metrics, so the authors built a benchmark on 126,604 paired single-nucleus profiles from 19 human studies with a shared vocabulary of 50,000 peaks and 3,185 genes and compared seven pipelines. The scButterfly-B pipeline, built on dual variational autoencoders, outperformed every transformer pipeline — including the autoregressive Perceiver model Omic Translator and a support-free reformulation of Cisformer — on per-gene and within-cell Spearman correlation in all sixty paired fold comparisons, with 3.8 times shorter training. Transformers come within 0.02-0.04 on within-cell profile shape but lag by 0.10-0.24 on per-gene correlation, and prespecified ablations show the Perceiver encoder was worth +0.0439 in within-donor per-gene correlation over full self-attention.