2026-09-29
Standard PCA clustering treats abundance as importance. DR-GEM self-supervises on reconstruction-error tails and raises mast-cell recovery in ovarian CosMx data from 0% to 85%.
Single-cell and spatial pipelines usually chain the same four steps: PCA, SNN clustering, UMAP, then marker-based cluster labels. Yeh, Jerby and colleagues at Stanford show that this stack treats abundance as importance.
Across 800 synthetic datasets, the rarer a cell type, the worse its reconstruction error (RE), the mean squared gap between input and the low-dimensional reconstruction. In real data, rare cells are 2.85–9.18 times more likely to sit in the worst 10% of RE. They then fail to form a cluster and get scattered into other labels. Sequencing more cells, or enriching a rare type you already know how to sort, does not fix a failure that happens at analysis time.
Swapping PCA for KL-DRO or CVaR-DRO lowers worst-case RE and still leaves rare cells mixed into other clusters. Tilted FairPCA, even when handed ground-truth group labels, equalizes RE and still cannot separate the rare type. Average-loss fairness and cluster recovery are different jobs.
DR-GEM (Distributionally Robust and latent Group-AwarE consensus Machine learning) is a self-supervised wrapper around whatever dimensionality reduction and clustering the user already runs. The main figures use PCA plus SNN. The Nature revision also wraps Kernel PCA, LDA, and scVI.
Four phases, the last optional.
The supervisory signals are all produced by the model itself: RE, weak labels, and consensus scores. The implementation is an R package at Jerby-Lab/drgem.
On Synthetic Dataset 1 (50 rare cells out of 21,250), DR-GEM puts 98% of the rare type in its own cluster; the standard pipeline gets 0%. Rare-vs-rest RE is no longer different (p=0.78). With simulated ambient RNA (Dataset 2): 88% vs 0% for rare cells, 64.7% vs 0% for contaminated rare cells, and 99.0% accuracy among high-confidence cells. Dataset 3, added in revision, plants five rare types at once: 90–98% each vs 0%.
| Dataset | Rare type | Standard | DR-GEM |
| Ovarian CosMx, non-malignant (82,808 cells) | Mast cells 0.397% | No matching cluster | 85% |
| Full CosMx (491,792 cells) | Mast cells and others | Top 4 types only | All 7 types; mast cells 86% |
| Ovarian MERFISH | B cells and others | Top 4 types only | All 6 types; fibroblasts 97.3% vs 47% |
| Mouse hypothalamus MERFISH | Pericytes 0.866% | No matching cluster | 79% |
| Breast-cancer IMC | B cells | No matching cluster | 88% |
On K562 CRISPRi Perturb-seq (75,305 non-targeting controls; some perturbations as rare as 0.133%), PTPN1i scores are Accuracy 0.95 / Balanced Accuracy 0.92 / Sensitivity 0.89 for DR-GEM versus 0.93 / 0.85 / 0.77 for the standard pipeline. Among sgRNA-positive cells with a low perturbation signature, 59% are still called PTPN1i, a state differential expression would miss. Detection of MRPL24i and NRDE2i stays below 50%, and those false negatives also have weaker signatures, closer to failed knockdown than to a missed cluster.
Two revision experiments matter for users. Against GiniClust3, scCAD, and CIARA on 14 rare-cell tasks, DR-GEM wins 12. scCAD is better on PBMC platelets; CIARA is better on developing thymus; both often hit Sens=0 on spatial and Perturb-seq data. At patient level, 185,071 malignant cells from 43 patients (12 patients below 1% of the cohort) map to patient-specific clusters for 34/43 patients under the standard pipeline and 39/43 under DR-GEM, including 8 of 12 underrepresented patients. DR-GEM also finds a 543-cell (0.29%) subclone in patient 1 that a pathologist matched to a morphologically distinct region.
The failure is in the analysis stack, not sequencing depth. Changing the PCA objective with DRO or FairPCA is not enough, because fair RE does not imply a separable cluster. The practical cases are spatial data with ambient RNA, Perturb-seq with extreme imbalance, and atlas work that needs under-sampled patients or new subclones. Because DR-GEM is a wrapper, it can sit on PCA or scVI without replacing a Seurat workflow.
This is a clear, incremental move: no new foundation model, just the supervised habit of attending to where the model is still wrong, applied to unsupervised embeddings. Phase I plus SMOTE+PCA can hit 100% rare-cell accuracy on synthetic data, but it paints the rare type as spuriously heterogeneous and has no per-cell confidence. When a clean reference map is the point, DR-GEM is the more complete tool.
The authors flag the cost of repeated Phase II embeddings, the choice not to use spatial coordinates or multi-omic channels, and online learning as future work. In imaging mass cytometry, high RE also marks abundant stromal cells whose proteins are under-covered by the panel, so a right-tail cell is not automatically rare biology.
Ground-truth labels on real datasets come from the original papers, which themselves used clustering and integration, so "recovering rare types" can be partly circular. Perturb-seq sgRNA identity is a partial check, and sgRNA detection is not the same as a functional hit. Clustering resolution was chosen to maximize AMI against those labels, which can flatter both pipelines. Annotations-in-the-loop steers the model toward a predefined signature dictionary; reviewers warned that this can suppress other biology. Empty droplets should be filtered in QC. DR-GEM does not do that job.