RareCapsNet: An explainable capsule network enables robust discovery of rare cell populations from large-scale single-cell transcriptomics

Explainable capsule network: primary capsules expose the marker genes that signature a rare cell population, with knowledge learned on one batch transferred to another without retraining.

Higher sequencing throughput now allows tens of thousands of cells to be profiled per experiment, increasing the chance of capturing rare cell types but complicating their identification. RareCapsNet applies a capsule network to the detection of rare, poorly represented or putatively rare populations in large single-cell RNA sequencing datasets. Its claimed contribution is explainability: primary capsules yield the marker genes associated with each identified population, giving a readable transcriptomic signature rather than a bare cluster label. Evaluation covers simulated and real data, with better specificity and selectivity than state-of-the-art methods according to the authors, plus a cross-batch demonstration in which a model trained on one batch recovers rare cells in another without retraining. The code is publicly available.

Synthesis written by Geno'X. For the full original abstract, please refer to the source publication.

Marker-gene explainability is what rare-cell detectors most lack: a rare cluster without a readable signature is not actionable, whereas a marker list can be checked against known biology. The abstract, however, gives no quantitative metrics and no thresholded definition of "rare", and validation stays largely methodological — cross-batch transfer is shown on study batches, not on the variability of a routine laboratory. Useful in translational research, premature as a component of a diagnostic pipeline.

Analysis by Dr Thibaut Benquey