Benchmarking gene expression reconstruction from single-cell latent representations
Abstract
Single-cell transcriptomics is typically modeled in low-dimensional latent representations that improve the signal-to-noise ratio of the data. Such representations underpin data integration, cell state discovery, and perturbation prediction, with applications ranging from large-scale organ atlases to latent trajectory modeling. Recent virtual cell approaches further leverage these representations to predict cellular responses as distributional shifts in latent space. Each of these applications ultimately requires faithful gene expression reconstruction from latent spaces for biological interpretation, enabling gene-level analysis of predicted perturbed or batch-corrected cells. Yet representation choice is typically treated as an implementation detail rather than a primary modeling decision, with no systematic evaluation of how well latent representations support gene expression reconstruction. Here, we introduce ReconEval, a benchmark for evaluating gene expression reconstruction from single-cell latent spaces. We benchmark two classes of latent representations: end-to-end trained models such as PCA, autoencoders, and variational autoencoders, and pretrained single-cell foundation model embeddings coupled to newly trained decoders. Reconstruction is evaluated both directly and after latent-space perturbation prediction. Across perturbational and observational datasets totaling over 100 million cells, our metric suite quantifies statistical fidelity; biological signal preservation, including differential expression, coexpression, cell-cycle structure, cytokine response and pathway activity; and perturbation-specific effects. We find that autoencoders achieve the strongest stand-alone reconstruction at low dimensionality, while variational regularization does not improve generalization in reconstruction. Frozen foundation model embeddings retain recoverable gene-level information, with reconstruction quality depending strongly on decoder architecture and pretraining objective. In latent perturbation modeling, high-dimensional PCA matches foundation model embeddings, while low-dimensional AE embeddings are optimal for flow-based generative models. Overall, reconstruction depends critically on the interplay between representation and downstream model, and simpler representations can outperform complex alternatives given appropriate capacity. Our benchmark establishes reconstruction as a critical evaluation axis for single-cell foundation models. We envision it improving the biological interpretability of latent-space modeling, a prerequisite for future virtual cell models to be validated by domain experts and grounded in biology.
Full-text reasoning
From the deep-tier full-text analysis of this preprint.
Reasoning review — 4 key claims, 3 well-supported, 1 with gaps
- supported ReconEval provides a comprehensive benchmark for systematically assessing gene expression reconstruction quality from single-cell latent representations.
- partial Optimal latent-space design for gene expression reconstruction depends critically on the interplay between the representation and the downstream model. gap: The text states this as a finding but does not present the specific results or analysis that demonstrate this critical interplay between representation and downstream model.
- supported MLP decoders consistently outperform KNN and Transformer decoders across various score categories for gene expression reconstruction.
- supported The performance of KNN decoders for gene expression reconstruction is largely insensitive to the number of neighbors (k) within the tested range.
Claims and gaps are read from the full text by a language model, shown for transparency; they do not affect ranking or selection.
Lifecycle
- biorxiv v1 2026-06-18 source ↗
Discussion
Moderated digest of third-party discussion on Bluesky — substantive endorsement and critique. 3 post(s) filtered (self-promotion, off-topic, or low-substance).
-
AI x Bio Discovery @aixbiobot.bsky.social · 787 followers neutral
Benchmarking gene expression reconstruction from single-cell latent representations [new] evaluates how well latent representations enable gene-level analysis, vital for biological interpretability and virtual cell models.