Novae

A study that images two breast tumour slides — one on a MERSCOPE, one on a Xenium — ends up with two different gene panels, and the usual way to find spatial domains across both is to throw away every gene the panels do not share; in the benchmark reported for Novae, that intersection left 185 genes. What survives must then be batch-corrected by Harmony and clustered by mclust or Leiden, external steps that re-run every time the analyst wants a different number of domains, and methods built this way still tend to return domains specific to one slide — precisely what a multi-slide study cannot use.

Novae removes the intersection step by never conditioning on a fixed panel. Each gene gets a trainable embedding; a cell is represented as its expression vector weighted against the L2-normalised embeddings of whichever genes its panel happens to contain, so the same encoder reads a small panel and a large one alike. Cells are wired into a Delaunay graph from their coordinates, a graph attention network summarises the local subgraph around each cell into a spatial-context representation, and those representations are projected onto a set of learnable prototypes that stand for elementary spatial domains. Training follows SwAV: two nearby subgraphs are augmented, their prototype assignments are smoothed by optimal transport, and a swapped cross-entropy loss pulls neighbouring contexts together. Because domains are prototypes rather than clusters, the optimal transport step doubles as batch correction and the prototype hierarchy can be cut at any resolution without re-clustering.

The model comes from the MICS laboratory at CentraleSupélec, Université Paris-Saclay, with the IHU PRISM precision-oncology centre at Gustave Roussy and Institut Cochin. It was posted to bioRxiv in September 2024 and published in Nature Methods in 2025, and ships as a scverse-compatible Python package from the group behind the Sopa spatial-omics pipeline.

#Key Features

• Panel-invariant cell embedding: Trainable per-gene embeddings, optionally initialised from frozen scGPT gene embeddings behind a trainable linear layer, let one checkpoint operate on any panel from any vendor.
• Zero-shot domain assignment: A pretrained checkpoint labels spatial domains on an unseen slide directly; fine-tuning is offered for refinement, not required.
• Nested domain hierarchy: Fewer than a thousand prototypes are hierarchically clustered once, so switching from 7 to 15 domains is a mapping lookup that runs in constant time.
• Native batch correction: Domain assignments arrive already aligned across slides, and are reused to align the continuous representations — no Harmony pass.
• One framework for downstream work: The same representations drive spatially variable gene analysis, spatial pathway enrichment, and PAGA-based slide-architecture comparison.

#Technical Details

Pretraining used 78 slides — nearly 30 million cells across 18 tissues — from three subcellular-resolution platforms: Xenium, MERSCOPE and CosMx. Two augmentations do the heavy lifting: synthetic per-gene noise that mimics batch effect, and random subsetting of the panel at a fixed ratio, which simulates a panel being swapped or updated mid-study. Subgraphs are built lazily rather than materialising the full slide graph, which kept training within a single 40 GB GPU. The three checkpoints released with the paper — human, mouse and brain — each carry about 32 million parameters; a smaller multi-species checkpoint of about 3 million parameters was added to the Hugging Face organisation in June 2026. On the two-panel breast dataset and the multi-panel colon dataset, Novae improved both FIDE (domain spatial continuity) and Jensen–Shannon divergence (cross-slide domain homogeneity) over SpaceFlow, GraphST, SEDR and STAGATE in zero-shot and fine-tuned modes, and reached higher adjusted Rand index on a synthetic five-slide, seven-domain benchmark. Post-inference domain assignment plus batch correction takes seconds where the Harmony-plus-mclust route can take days at six million cells.

#Applications

Novae fits cohort-scale spatial transcriptomics, where the question is which niches recur across patients rather than what a single slide contains. The paper works two such cases: a reactive versus non-diseased lymph node pair, where trajectory analysis over the domain representations showed germinal-centre reorganisation and an expanding CXCL12-enriched stromal domain, and six mouse brain slides across ages in which TgCRND8 animals with Alzheimer's-like pathology showed elevated brain-ageing pathway activity in specific domains. It accepts spot-resolution data such as Visium alongside single-cell platforms, and the journal version extends the analysis to spatial proteomics and to slides paired with H&E.

#Impact

By folding clustering and batch correction into the pretrained model, Novae turns spatial domain identification from a multi-tool pipeline into a single call, and gives multi-slide studies a domain vocabulary stable across panels and instruments. The code is BSD-3-Clause on GitHub and installable from PyPI, and the checkpoints and training samples are on the Hugging Face Hub. Two caveats bound the results: the pretraining corpus was segmented with vendors' proprietary software, which the authors flag as a limit on data quality, and the training slides were image-based, so Visium, Visium HD and proteomic inputs are supported at inference without having been seen during pretraining.