2026-09-29
mc-ASTRA maps molecular, compositional and spatial views onto interpretable tissue-state maps. DSS colitis yields four remodeling programs; day-35 mucosa heals, molecules do not.
Single-cell and spatial assays resolve tissues down to genes, cell types and neighborhoods. The questions that matter in a cohort sit one level up: what state is this sample in, and how does it differ from the others. Current tools split that job. Distance-first methods such as GloScope, PILOT and QOT turn each sample into a cell distribution, embed pairwise distances, and never map the axes back onto genes or cell types. Deep sample embeddings and tensor methods usually ingest one tissue descriptor and rarely read the latent space as a coordinated program across cell types.
Benchmarks make the gap worse. They treat diagnosis or chronological age as ground truth, even though those labels often explain little of the variance in the features you actually measured. The Saez-Rodriguez group had already recast group factor analysis as multicellular factor analysis, sharing gene programs across cell types. That model was unsupervised. You could not pin a factor to time, biopsy fraction or a pathway gene set.
mc-ASTRA is the missing wrapper: build several tissue descriptors from single-cell and spatial data, fit a semi-supervised factor model, and read the factors as multicellular programs.
The statistical engine is not a new algorithm. It is MOFA-FLEX, a probabilistic group factor model. Each sample becomes several views: pseudobulk expression per cell type, centered-log-ratio composition, and spatial neighborhood enrichment from squidpy. Joint low-rank factorization returns factor scores (coordinates on the tissue-state map), feature loadings (the programs), and per-view R² from reconstruction error.
Two kinds of prior supply the semi-supervision. Sample-level covariates can absorb a known source of variation into one factor while the rest stay exploratory. Feature-level gene sets, for example PROGENy pathway genes, can anchor a factor to a known process without freezing the gene weights. Results land in anndata, so Scanpy clustering, ParTIpy archetypal analysis and patpy information-retention scores all apply.
A downstream step turns loadings into directed coordination networks. Program activity of one cell type is the response; the others are predictors in a multivariate or mixed-effects regression. Edge score equals coefficient times model R². Ligand-receptor pairs from the LIANA+ consensus database are kept only when their loading signs agree on those edges.
Crohn's ileal epithelium from scIBD. A model fit on one library chemistry pinned biopsy fraction to a single factor: mean R² across cell types 0.10, association adj. p = 1.65×10⁻⁹. Factor 2 associated with Crohn's status (adj. p = 0.002, mean R² 0.08) and carried JAK-STAT activity in enterocytes, goblet cells and cycling transit-amplifying cells. Projecting the second chemistry kept Factor 2 associated with disease (adj. p = 0.0004).
Lupus PBMCs before and after IFN-β. A two-factor non-negative model guided one factor with PROGENy JAK-STAT up-genes (top 500). The guided factor recovered 23% of model variance and associated with stimulation (adj. p = 7.88×10⁻¹³); the second factor recovered 75% and captured genes that dropped after treatment. For the ten genes with the largest PROGENy weights, Pearson correlation with the reweighted signature was 0.07. The prior anchored the process. It did not freeze gene contributions.
Myocardial infarction snRNA-seq, 27 samples. Two of five factors associated with tissue class (mean R² 0.24). ParTIpy found three archetypes: Arch 0 near control, remote and border zones; Arch 1 with MAPK, EGFR and VEGF; Arch 2 with inflammation and fibrosis. Samples sit between those extremes rather than in discrete bins.
DSS mouse distal colon MERFISH, 940-gene panel, days 0, 3, 9 and 35. Day 21 was dropped because mean Pearson correlation across samples and cell types was 0.04. Separate models used molecular, compositional or spatial views; a joint model added a Gaussian-process time prior in the MEFISTO style. Five-nearest-neighbor time-point prediction:
| sample map | time-point F1 |
| molecular only | 1.00 |
| composition only | 0.95 |
| spatial only | 0.93 |
| joint three views | 0.97 |
| GloScope | 0.97 |
GloScope parked day-35 recovery samples in an isolated region, away from healthier day 0 and day 3.
The joint model recovered four parallel trajectories. Factor 1 is peak inflammation, on at day 9 and off by day 35, mean R² across views 0.24, with TNFα, NFkB and JAK-STAT plus neutrophils and monocytes. Factor 2 is progressive injury (mean R² 0.05): day 3 already resembles peak, inflammation-associated epithelial cells appear early and plasma cells drop. Factor 3 is the early response: epithelial Fos, Fosb and Egr1 (MAPK/ERK) plus Mmp15; immune Ccl22 and Il17re. Factor 5 is residual remodeling. Mucosal architecture is nearly restored at day 35, molecular state is not: persistent Axin2 in fibroblasts and immune cells, epithelial Gdf15 and Nlrc3, immune Nod2 and Adamts12, and Per1 plus Irf8 still off baseline.
Coordination networks sharpen the contrast. Early epithelium couples more to immune and endothelial compartments (Fasl–Fas, Igf2–Igf1r, Mmp7–Sdc1). At peak it couples to fibroblasts and smooth muscle (Tnfsf12, Cxcl10, Csf2, Il23a, Thbs1, Bmp10). In single-view models, time explained 94% of compositional variance and 58% of spatial organization. In the molecular model, smooth muscle, fibroblasts and epithelium associated most with disease stage (84%, 75%, 78%).
For people running single-cell or spatial cohorts, this is a scverse-shaped end-to-end wrapper: descriptor generation, semi-supervised factors, pathway enrichment, ligand-receptor scoring and archetypal analysis on one anndata object. It does not claim to beat every sample embedding. GloScope already hits F1 0.97 on time-point recovery. The selling point is being able to say which view, which gene and which cell-type edge carry the variance.
Linear factors project new cohorts, as in the Crohn's cross-chemistry test, and can park a known technical effect in one factor without swallowing the disease signal. Labs already on MOFA, LIANA+ and decoupler will find the migration cheap. PyPI is not up yet. The stack targets Python 3.12/3.13 and the development branch of MOFA-FLEX.
This is incremental engineering, not a new representation-learning method. It is useful if you accept linear factors and cell-type labels as view boundaries.
The authors say so: segmentation, annotation, spatial metrics and gene filters all move the map; priors, guiding variables and factor count need diagnostics, not a template. Day 21 was dropped, so the time axis is incomplete.
A few claims stay soft. Coordination networks test whether program activity in A predicts B, not causation. Ligand-receptor calls mostly require matching loading signs, with spatial contact shown only for two interactions via LIANA+ inflow. Lupus has eight donors, infarction 27 samples; factor count partly follows sample size. DSS is a strong signal, so high time-point F1 is unsurprising. The supplement calls mc-ASTRA the most general factor or tensor method; that is a taxonomy, not a quantitative bake-off. The paper is a bioRxiv preprint.