An autoencoder outperforms two transformers on prediction of gene expression from chromatin accessibility

Jesse G. Meyer

Home/Archive/arXiv

arXiv · 2610.28853

An autoencoder outperforms two transformers on prediction of gene expression from chromatin accessibility

Jesse G. Meyer

View original source ↗Read source document ↗Export metadata ↓

Abstract

Models that predict RNA abundance from chromatin accessibility are rarely compared under the same features, targets and metrics. We built one benchmark on 126,604 paired single-nucleus profiles from 19 human studies with a shared vocabulary of 50,000 peaks and 3,185 genes, ten 5x2 cross-validation folds and one scoring target, and compared seven implemented pipelines and two reference predictors across 70 trained runs; each pipeline keeps its own input normalisation and native training target. The scButterfly-B pipeline, built on dual variational autoencoders, outperformed every transformer pipeline, including an 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. Transformer pipelines come within 0.02-0.04 on within-cell profile shape but lag by 0.10-0.24 on per-gene correlation. On two withheld cell types per-gene correlation is low; controls show the podocyte values are typical of within-type correlation for seen cell types of the same size, whereas Paneth values match a cell-agnostic shuffle for every Omic Translator arm. Within the transformer family, prespecified ablations show the Perceiver encoder was worth +0.0439 in within-donor per-gene correlation over full self-attention, a chromosome-aware genomic attention prior reduced it by 0.0096 (exact donor-level sign-flip test, Holm-adjusted p = 1.2e-4), and decoder output order had no detectable effect. A training-mean profile reaches 0.314 within-cell correlation with no input and a constant at the true expressed-gene positions reaches 0.996 per-gene correlation, so oracle-support protocols are reported only as diagnostics. Code, results and an audited release are provided.

Explore related subjects

q-bio.GN

Keep this discovery

Explore connections, maps & timelines

BibTeX RIS

Jesse G. Meyer. 2026-10-20. An autoencoder outperforms two transformers on prediction of gene expression from chromatin accessibility. https://arxiv.org/abs/2610.28853

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Benchmarking Recent Single-Cell Synthetic Data Generation Methods on the OneK1K Dataset

Synthetic single-cell RNA-seq data generation offers a promising route for privacy-aware biomedical data sharing, but donor-linked datasets require careful balancing of biological utility and privacy risk. In this work, we benchmark synthetic generation methods for the CAMDA 2026 ELSA Health Privacy Challenge using the donor-split OneK1K dataset, comparing scGFT, scLDM, and Cell2Sentence across statistical fidelity, downstream biological preservation, and membership inference vulnerability. Because full-transcriptome generation is computationally demanding, we focus generative modeling on highly variable genes and evaluate alternative strategies for reconstructing non-HVG expression. Our results show that non-HVG handling is a central determinant of privacy: directly copying non-HVG values leads to severe leakage, whereas Poisson-based reconstruction substantially reduces membership inference risk. We further find that scGFT and scLDM better preserve biological structure than an independent Poisson baseline, while maintaining near-random privacy profiles under appropriate configurations. These findings highlight the importance of evaluating synthetic single-cell data beyond marginal similarity alone, emphasizing robustness, interpretability, and privacy-conscious design choices throughout preprocessing, generation, and post-processing.

q-bio.GN↗

arXiv2026

PACE: Geometry-Aware Bridge Transport for Single-Cell Trajectory Inference

Single-cell trajectory inference from destructive time-course snapshots is fundamentally ill-posed: neither cross-time cell correspondences nor continuous trajectories are observed, so the snapshot distributions alone do not uniquely determine the underlying dynamics. Existing optimal transport and flow-based methods typically couple cells by Euclidean proximity at observed clock times, which can misalign trajectories when development is asynchronous and cells sampled at the same experimental time occupy different latent pseudotime stages. We propose PACE, a trajectory inference framework that recovers geometry-consistent continuous transport dynamics from destructive time-course snapshots through three coupled components. First, PACE constructs a state- and time-dependent anisotropic Riemannian metric that assigns low transport cost along locally supported tangent directions while penalizing normal velocity components. Second, it alternates between refining cross-time couplings under the induced path-action cost and fitting endpoint-preserving neural bridges between adjacent snapshots. Third, it distills the learned bridge dynamics into a global continuous-time velocity field over cellular states. Across seven controlled and biological datasets covering nine held-out reconstruction experiments, PACE achieves the strongest overall reconstruction performance, reducing MMD, Wasserstein-1 distance, and Wasserstein-2 distance by 23.7% on average relative to the strongest competing baseline. PACE also improves RNA-velocity alignment by 15.4% on an embryoid body differentiation benchmark, without requiring explicit cell pairing, lineage tracing, or RNA-velocity supervision during training. Code is available at https://github.com/AI4Science-WestlakeU/PACE.

q-bio.GN↗

arXiv2026

Reading Between the Reads: A Decade On, Revisiting Sequencing Protocol Annotation in the Sequence Read Archive

Background: Experimental metadata are essential for interpreting archived sequencing data. In 2015, Alnasir and Shanahan examined the annotation of fragmentation, adapter ligation and enrichment in the Sequence Read Archive (SRA). This study revisits those measurements using the original SQL keyword definitions on a September 2026 NCBI metadata snapshot reconstructed in the historical SRAdb schema. Results: The contemporary database contained 41,593,302 experiments, compared with 419,620 in 2013. In library_construction_protocol, the proportion matching all three historical keyword groups declined from 4.06% to 1.42%. The protocol field was NULL in 67.65% of contemporary experiments. Searching design_description alongside the protocol field increased combined three-step keyword detection to 1,197,367 experiments (2.88%), with most additional matches occurring where the protocol field was NULL. Coverage varied across sequencing library strategies, ranging from 0.26% for AMPLICON to 7.69% for RNA-Seq. Independent comparison of 150 sampled experiment records against source XML confirmed agreement in all historical keyword classifications. Conclusions: The original criteria identify a smaller proportion of experiments in the designated protocol field than a decade earlier. Additional preparation-related terminology is detectable in general experiment descriptions, exposing the limitations of searches restricted to one metadata field. Differences among sequencing strategies complicate historical comparisons as preparation methods evolve. The findings support more consistent protocol annotation and the need to distinguish keyword occurrence from the adequacy of experimental documentation.

q-bio.GN↗