Single-cell pairing gets a confidence score: a Bayesian posterior on the correspondence filters bad matches

Cross-modal single-cell matching can now carry a confidence on the pairing itself, not only a match. Earlier matching models return a point-estimate match score, which names the best pairs but not how close the runner-up was, whether the assignment would survive a different draw of training data, or whether an input lies where the encoders were never trained. On multi-omics benchmarks with known pairing, the cell-type-level predictive uncertainty separates wrong from correct pairs at AUROC 0.856±0.016, with median uncertainty 0.010 on correct pairs and 0.079 on errors.

Discarding pairs by uncertainty improves the matches that are kept. With every cell type seen in training, discarding the least certain fifth lowers the error rate from 6.8% to 2.5%, about 70% of the errors. In the same setting the deterministic network's confidence margin gives nearly the same curve, AUROC 0.858 against 0.856, so on seen types the posterior finds the errors the margin already finds.

At the cell level the posterior is better at flagging wrong matches that cross a cell-type boundary, the kind that carries the wrong label. The deterministic control reads the same assignment through the Sinkhorn transport weight of the mean score matrix, with its temperature chosen on the test set by AUROC for any wrong match. The transport-plan posterior reaches AUROC 0.679 on cross-type errors against 0.635 for the control; filtering on the most certain half removes 32.8% of cross-type errors for the posterior and 25.1% for the control.

The demonstrated scope is two public benchmarks of bone marrow mononuclear cells: 22 cell types in the multiome data and 38 in CITE-seq, with the cell-type-level findings replicating on CITE-seq at AUROC 0.970 and 0.963. Cell-level matching accuracy is read on the 20,009 cells of the NeurIPS 2021 competition test set, and should be placed beside the published CLUE and MatchCLOT scores.

On seen types the predictive standard deviation correlates with the deterministic confidence margin at Spearman 0.95 to 0.99 per pair, so abstaining on it is no gain over the deterministic network. Separation of unseen types is above chance in 19 of 22 types, is modest, and MC dropout and the cosine similarity of frozen embeddings also flag unseen types. Cell-level novelty detection does not remove the dependence on type size, and the authors limit that claim to the cell-type level.