KQE-2 clusters cell atlases reproducibly at large scale
Reproducible quantum clustering for single-cell transcriptome atlases · arXiv:2611.11745
In plain words
Single-cell atlases group cells by how their measured genes differ. Common approaches keep only some genes, and trained methods can produce different groups when run with different random seeds.
KQE-2 is a way to cluster cells that calculates the same representation each time without training. Across ten complete atlases, its clusters matched annotated cell types best on eight; on the other two, its scores were within the best method’s run-to-run range. The comparisons used the annotated number of types.
The research
Clusters shape how researchers read cell types and markers from an atlas, so changes between runs can alter the biology they infer.
KQE-2’s repeatable assignments address that specific source of variation, while its atlas comparisons test whether repeatability comes with accuracy competitive with established methods.
KQE-2 uses an exactly calculated, training-free cell representation to cluster transcriptome atlases reproducibly, then compares its clusters with annotations and established methods.
Across ten complete atlases, the authors compared KQE-2 with scanpy, Seurat, scVI and scGPT. They measured agreement between clusters and annotated cell types using adjusted mutual information (AMI), with all methods scored at the annotated number of types.
KQE-2 with Q-PSC ranked highest on eight atlases and was within the best method’s run-to-run range on two. Its mean margin over the best other method was 0.050 AMI across the ten atlases, supporting competitive accuracy in this evaluation.
The atlas comparisons used the annotated number of cell types both to choose KQE-2’s gene input and to score the clusters. A separate test chose clustering resolution without labels only after that input configuration had been fixed.
What the evidence shows
- A fixed representation of cells
KQE-2 is a training-free hybrid quantum-classical framework for clustering single-cell RNA-sequencing atlases. It represents each cell from its gene measurements and clusters cells by similarities between those representations.
The motivation is that common workflows discard most measured genes before clustering, while learned representations may change across runs and affect the clusters used to interpret an atlas.
- Seeds can change clusters
A run-to-run comparison makes that reproducibility concern concrete. On the complete fetal atlas, 23.1% of cells changed clusters between runs of scVI, a trained model for analyzing single-cell data, using seeds 0 and 2; two KQE-2 runs on the same data changed no cell assignments.
The percentage counts cells in that atlas. This comparison is limited to the named data and runs, but it shows why a stable score alone may not guarantee stable clusters.
- Accuracy across ten atlases
The accuracy test asked how closely clusters matched annotated cell types. It measured this agreement with adjusted mutual information (AMI), a score that is higher when clusters align more closely with the annotations.
Across ten complete atlases, KQE-2 with Q-PSC had the highest AMI on eight and fell within the best method’s run-to-run range on two. Its mean margin over the best other method was 0.050 AMI across those atlases. The result supports competitive accuracy in this test, with the two-atlas qualification intact.
- What the comparisons support
A shared graph construction and Leiden clustering protocol helps separate representation from the final clustering step. Under that protocol, KQE-2 had the highest AMI on six atlases and came within 0.004 of the best result on three more.
A separate run clustered 11.4 million mouse embryo nuclei on one server. These findings concern the tested atlases and configurations, which used the annotated type count to select KQE-2’s input and score clusters.