2026-10-01
Free no-code web platform from the Theis lab: maps single-cell datasets onto reference atlases via scArches transfer learning, with annotation and uncertainty scores built in.
Single-cell analysis has shifted to a reference-based paradigm: new data are mapped onto large integrated atlases and inherit cell-type annotations, the way early genomics leaned on the reference genome. The mapping itself runs on transfer learning, which means Python skills, environment setup and, increasingly, GPU access. The people who stand to benefit most, wet-lab biologists, are exactly the ones locked out.
Existing tools each miss a piece. Azimuth has a GUI but only exposes Seurat v4 methods. FASTGenomics offered lung-atlas mapping and is no longer free. scGPT Hub takes the foundation-model route, yet zero-shot evaluations have shown single-cell foundation models underperform scVI on integration while costing more compute. CellTypist annotates but does not remove batch effects. A free, no-code, method-agnostic slot was open.
ArchMap (Architecture Mapping, archmap.bio), from Fabian Theis's group at Helmholtz Munich, is published as a Nature Genetics Brief Communication. Its backbone is scArches, a transfer-learning framework that freezes a reference model and fine-tunes new query data into the reference's latent space. The platform supports scVI, scANVI and scPoli integration methods, the first two being consistent top performers in atlas-level integration benchmarks.
Design choices worth calling out:
The paper's real novelty is in the metrics. Alongside the Euclidean uncertainty inherited from the HLCA (a weighted-KNN label probability, with p below 0.5 flagged as unknown), ArchMap adds a Mahalanobis uncertainty: a Gaussian mixture model fit on the reference embedding, with one component per cell type, scores query cells by Mahalanobis distance, which accounts for correlations between latent features. In the fibrosis use case it aligned better with disease status than the Euclidean version. A presence score counts how often each reference cell appears in the 15-nearest-neighborhoods of query cells, then diffuses those counts over the reference KNN graph with 100 rounds of random walk with restart, quantifying how well the reference's cell states are represented in the query.
The demonstration maps an idiopathic pulmonary fibrosis (IPF) patient dataset (GSE132771) onto the Human Lung Cell Atlas.
| Metric | Value | Reference point |
| Reference gene coverage | 96.85% | Platform recommends at least 80% |
| Leiden cluster preservation | 5/5 | Maximum is 5 |
| Query cells with anchors | 53.96% | Healthy data to HLCA: 63% |
| Cells labeled unknown (uncertainty > 0.5) | 8.99% | n/a |
Adventitial fibroblasts showed the widest spread in uncertainty. Splitting that cluster at 0.4 revealed a strong association between high uncertainty and diseased donors; CTHRC1, a known fibrotic marker, was elevated in the high-uncertainty group while SERPINF1 marked the low-uncertainty, homeostatic one. Everything ran in the browser, with no local code.
The anchor-rate direction checks out too: 63% for healthy samples mapped to the HLCA versus 54% for diseased ones, as expected when disease data meets a healthy reference.
For single-cell researchers the workflow collapses to four steps: pick an atlas, pick a model and classifier, upload an h5ad file, read the results. CELL×GENE is embedded for visualization, quality metrics are computed automatically, and the download is a joint labeled h5ad plus the fine-tuned model. Data are encrypted in transit and at rest, deleted projects leave cloud storage after 3 days, and a self-hosted route is documented with a cost estimate of about 200 euros a month on GCP, plus a local notebook replicating the exact pipeline.
To be clear about what this is: not a methods advance. Mapping is scArches, integration is the scVI family, and ArchMap's delta is engineering, metrics and collaboration. Two things make it interesting beyond single-cell. It is a template for productizing transfer-learning-to-a-reference: freeze a large pretrained model, adapt new data into it, the same pattern as adapting vision backbones. And the team explicitly rejects the foundation-model route, citing evaluations where scGPT-class models integrate worse than scVI at higher cost, a counter-trend judgment in the foundation-model era.
Stated by the authors: each mapping is capped at 200,000 cells by server capacity (batches can be mapped separately and concatenated with identical results, but cells sharing a batch label must stay together); high uncertainty does not by itself mean a novel cell state, since residual batch effects and low-quality cells also inflate the score and require manual validation; missing query genes are zero-padded, and coverage below 80% should be treated with caution; anchor percentage depends on cell counts on both sides and should only be compared across models, not across datasets.
Two more from reading it. The IPF case is a confirmatory demo on known-answer data, GSE132771 being a 2020 fibrosis dataset and CTHRC1 a known marker, closer to a demonstration of platform capability than new biology. And the classifier comparison's concrete F1 values appear only in the Extended Data figures; the main text gives conclusions without numbers.