scShift

Comparing biological states across single-cell RNA-seq studies runs into a conceptual rather than an engineering problem: batch effect and biological difference are not separable from data alone. This is the classic non-identifiability of nonlinear latent variable models, and it is why integration methods that concatenate a cell representation with a batch encoding entangle disease signal with cell-type signal.

scShift, from Mingze Dong, Kriti Agrawal, Rong Fan, Esen Sefik, Richard Flavell and Yuval Kluger at Yale University, with Flavell also at HHMI, builds the model around the identifiability question. Instead of concatenating representations, scShift sums two sets of latent variables — one encoding intrinsic cellular properties shared across datasets, one encoding batch-dependent variation that carries both experimental artifact and genuine biological state. Under a sparsity constraint on the batch-dependent component and an independence constraint between the latents and the dataset label, the authors show that batch-dependent variation is linearly identifiable, which is what makes cross-dataset comparison of biological states well-posed.

Dataset labels are therefore needed only during training. Once the model is fit on a compendium of atlases, both embeddings can be extracted from a new dataset with no labels and no additional training. That is a different route to a single-cell foundation model than the masked-token pretraining of scGPT or Geneformer — the generality comes from a variational objective plus dataset-label supervision rather than sequence modelling.

#Key Features

• Identifiability by construction: Summing rather than concatenating latents, plus sparsity and independence regularization, guarantees that batch-dependent variation is recoverable up to a linear transform.
• Two separable embeddings: The unperturbed embedding retains cell-type structure with batch and perturbation signal removed; the biological embedding carries disease and perturbation state.
• No annotation required: Training uses only dataset labels, never cell-type or biological-state annotations, so it can consume public atlases as they are.
• Cross-system state alignment: Because biological states share one space, in vitro drug perturbation profiles can be projected alongside in vivo patient samples.
• Measured scaling behaviour: Over 200 models trained under varying conditions show zero-shot capability emerging past a transition threshold in donor and cell number rather than improving smoothly from the start.

#Technical Details

The architecture pairs a count encoder and a dataset-label encoder whose outputs are added to form the full centralized variation, which a decoder maps back to the gene expression distribution. Sparsity in the label encoding is imposed by a probabilistic l0 penalty using a stochastic gate; independence between the centralized latents and the label encoding is enforced by a kernel maximum mean discrepancy penalty. Training randomly permutes 25% of genes across cells within each mini-batch as a denoising scheme, and optimizes the evidence lower bound with both regularizers. Components tied to non-zero entries of the dataset-label encoding then form the biological embedding, the remainder the unperturbed embedding.

Two compendiums are used. The blood model trains on 1,000,000 cells subsampled from 30 CellxGene studies covering 2,538 donors, plus 240,090 cells from 144 drug perturbations. The lung model trains on 3,433,014 cells from 697 donors across 54 datasets. On held-out interferon-stimulation and pathogen-exposure datasets, scShift leads all compared methods on biological state inference — including perturbation-modelling approaches given control/non-control labels — and matches the best on cell-type representation. Donor-level fibrosis classifiers built on scShift embeddings and trained on pulmonary fibrosis data transfer to blood IPF and fibrotic hypersensitivity pneumonitis at AUC around 0.7, and to post-COVID fibrosis, a phenotype absent from training. Ablations attribute the zero-shot behaviour to the sparse encoding, denoising scheme and MMD term jointly.

#Applications

The immediate use is querying a new scRNA-seq dataset against everything already published: project the cells, read cell types from the unperturbed embedding and disease or treatment state from the biological embedding, then compare that state to any other study in the same space. The lung fibrosis analysis shows the pattern in full — universal myeloid-fibrosis signatures, candidate repurposing targets, and cell interaction networks assembled across datasets that no single study could have produced. It ships as a pip-installable scshift package with documentation, within reach of an ordinary scanpy workflow.

#Impact

scShift's contribution is as much statistical as computational: it identifies why cross-dataset biological state comparison fails and supplies a design under which it provably does not. The measured scaling law over 200-plus training runs is unusual in this literature, and the transition-threshold finding — that zero-shot ability appears only past a certain donor diversity — is a concrete guide for anyone assembling a training compendium. The work remains a bioRxiv preprint under a CC-BY-NC-ND license. The code is BSD-3-Clause licensed, but pretrained weights are distributed through a Google Drive folder linked from the README rather than a versioned artifact host, and only the tutorial checkpoint is available; the scaled models behind the scaling analysis are not released.