Skip to main content
Back to timeline
bioRxivSource publication:

GatorAnchor scores each cell-cell link by cross-modal disagreement, topping mean clustering metrics on ten paired multi-omics datasets

Synopsis

In single-cell multi-omics integration, molecular layers often disagree about whether two cells are neighbors, yet most methods treat those relations as equally reliable. GatorAnchor folds each modality's uncertainty and the disagreement between modality means into a fused representation, then scores candidate cell relations from RNA similarity, second-modality similarity and shared edge-decoder support, weakening or removing unreliable edges. Across ten paired RNA+ADT and RNA+ATAC datasets it beat nine representative integration methods on mean AMI, NMI, ARI and ACC; keeping the 90% least-unreliable candidate edges raised same-type edge purity in all ten datasets, though graph structure changed differently by dataset.

Source-provided article image: GatorAnchor: Cross-Modal Uncertainty-Aware Learning of Reliable Cell Neighborhoods in Single-Cell Multi-Omics
Figure 1 · bioRxiv · Page 5

Interpretation

Cell neighborhood relations now carry a decomposable reliability score instead of being treated as equally trustworthy graph edges. Earlier multi-omics integration methods mainly learned shared or aligned low-dimensional representations, and modality-specific uncertainty and cross-modal disagreement were not typically turned into relation-specific judgments about which cell connections should carry information. Across ten paired datasets, GatorAnchor's mean AMI, NMI, ARI and ACC were highest among nine representative integration methods; on Brain SNARE its AMI, NMI, ARI and ACC were 0.7947, 0.7956, 0.7764 and 0.8752, exceeding scMAGCA, the strongest baseline on all four metrics there, by 0.1126, 0.1121, 0.2586 and 0.2814.

Each candidate cell relation is fused from three evidence sources, disagreement itself enters the unreliability score, and high-scoring edges are attenuated or removed from the propagation graph. Fusion uses a precision-weighted product-of-experts rule that combines modality-specific Gaussian representations by learned variance, and adds the dispersion among modality means to the fused variance to cover mean disagreement that product-of-experts variance does not see. Candidates come from the observation graph, the fused representation and modality-specific neighborhoods, with RNA similarity, second-modality similarity and shared edge-decoder support merged under uncertainty-derived weights; retaining the 90% least-unreliable candidate edges raised same-type purity by a dataset-balanced mean of 0.0208 (95% CI 0.0139–0.0276) across ten datasets.

The unreliability score can be decomposed into specific sources of evidence conflict rather than serving only as an internal weight. Stratifying candidates by unreliability decile, the fraction of edges connecting different reference cell types rose from 0.146 in the lowest decile to 0.421 in the highest, a difference of 0.275 (bootstrap 95% CI 0.140–0.403). Second-modality-dominant conflicts showed the highest cross-type fractions in RNA+ADT and RNA+ATAC datasets, 0.60 and 0.59; on 10Xmalt, modality conflict was positively associated with RNA–protein detection discordance (β=0.0509, 95% CI 0.0200–0.0832) and conditional abundance discordance (β=0.0435, 95% CI 0.0083–0.0858).

Perspective

The demonstrated scope is clustering evaluation on ten public paired single-cell multi-omics datasets (five RNA+ADT, five RNA+ATAC, 713 to 30,672 cells) plus a three-modality analysis of 46,532 CD4-related cells from 12 donors in ALTRA TEA-seq; missingness and mosaic experiments use four datasets: 10Xmalt, GSE158013, PBMC unsorted 3k and GSE201402.

The unreliability score is explicitly framed as a decomposable graph-reliability diagnostic, not a calibrated edge-error probability; cross-type links are only a proxy for biological neighborhood inconsistency, since genuine cell-state transitions can also produce cross-type edges. Under missing and mosaic measurements, as paired cells fall from 100% to 10%, mean ARI drops from 0.613 to 0.396, and with zero paired anchors chance-adjusted balanced accuracy of cross-modal centroid transfer falls to −0.007, near chance.