SURE

Two labs sequence the same tissue and each ends up with a population called "CD4+ T cells" whose coordinates are defined by its own dataset's neighbours — meaningless outside that experiment, so comparing cell states across studies means re-integrating from scratch.

SURE (SUccinct REpresentation of cells), from Feng Zeng and Jiahuai Han at Xiamen University, fixes a set of landmarks first and projects everything else onto them. It treats metacell identification as density estimation, not clustering: a network emits a vector-quantization codebook of K landmarks, each a distribution over latent cell-state space, fitted by variational inference to cover the observed cells. Reuse comes from an asymmetry borrowed from batch-correcting autoencoders — generation is batch-aware, p(x|z,b), but inference is not, so the encoder q(z|x) can place cells from a dataset whose batch structure it never saw, making the codebook a coordinate system rather than a per-dataset embedding.

The authors instantiate that as HBMCA, the Human Blood MetaCell Atlas: 500 metacells distilled from 5,310,750 blood cells across 29 public datasets and 1,906 individuals. Cohorts from 13 diseases and bulk RNA-seq databases project onto it in one forward pass with no batch correction or fine-tuning, as SCimilarity also maps queries zero-shot. The method reaches users under several names: the SURE repository, the PyPI distribution SURE-tools, and the successor release SUREv2, which adds a SingleOmicsAtlas class for compression and assembly and a SUREMO class for paired multi-omics. Neither ships a pretrained checkpoint.

#Key Features

• Batch-invariant zero-shot mapping: Batch factors enter only the generative path, so a trained codebook embeds new datasets with no batch covariate or gradient updates.
• Out-of-reference scoring: Codebook-simulated reference cells are mapped into the query space, and a query cell scores as novel when its nearest-neighbour distance is improbable there.
• Distribution-preserving compression: An atlas travels as its codebook plus mapping and generation functions — a few hundred megabytes that regenerate counts matching the original distribution.
• Codebook-based annotation: Each metacell stores phenotype likelihoods aggregated from its member cells, so cell type is assigned by maximum likelihood, not a separate classifier.
• Paired multi-omics metacells: SUREMO reads two count matrices through one shared encoder and reconstructs each through its own decoder and likelihood, so one codebook indexes both modalities.
• Hierarchical atlas assembly: Per-dataset metacells are bootstrap-sampled into pseudo-cells and re-encoded into secondary metacells, so atlases integrate in parallel rather than jointly.

#Technical Details

SURE is implemented in Pyro and optimized by stochastic variational inference, with the codebook indexed by a OneHotDirichlet prior and a choice of four latent distribution families and three count likelihoods, optionally zero-inflated. On the NeurIPS 2021 BMMC benchmark (69,249 cells, 13 batches), metacell-derived cell type frequencies matched the single-cell composition at Pearson 0.990 against 0.705 for MetaQ and 0.434 for SEACells. SUREMO reuses that machinery with a second decoder head: its documented run fits a 500-entry codebook over the RNA and ATAC halves of the 10x Multiome BMMC data (GSE194122) for 200 epochs, each modality under its own zero-inflated negative binomial.

Compression fidelity, scored by a structural similarity index over gridded cell type distributions, exceeds 0.85 at 800 metacells for datasets of 69,249 and 195,632 cells alike — a limit set by distinct cell states rather than data volume. A donor classifier trained over 182 individuals fell from 0.893 accuracy on real held-out counts to 0.053 on counts regenerated from the compressed atlas.

#Applications

The immediate use is comparative analysis without re-integration: a cohort projected onto HBMCA lands in the same coordinate system as every other, so disease-versus-normal shifts in composition read off directly. Bulk RNA-seq profiles map onto that space too, resolving 29 sorted immune populations onto 43 metacells. Compression opens a second use: distributing a reference, or submitting a query, without donor-level count matrices. Monocyte subtypes withheld from a reference were correctly flagged as out-of-reference, where Symphony placed them among known cell types.

#Impact

Both preprints rest on benchmarks the authors ran themselves, and reach depends on what gets distributed: the source sits on GitHub in the SURE and SUREv2 repositories and is archived on Zenodo, but no pretrained codebook is downloadable, HBMCA included, so a user must rebuild a reference before querying it. The multi-omics path is narrower still — SUREMO.py ships only in SUREv2, not in the SURE repository the authors continue to develop, and it has no reference atlas or query-mapping route of its own, so its one documented workflow fits a model on the user's own paired data. Both availability statements call the code free for academic use while both repositories carry an MIT license, a discrepancy worth resolving before commercial use. What it contributes regardless is evidence that a small, fixed codebook, learned with batch information withheld from the encoder, is a workable substrate for a shared cell coordinate system.