The method, called scFormer, was developed by Jinjin Huang, Xuanzhe Xia, Feng Luo, Lianghu Qu, Xiao Feng, Lingling Zheng and colleagues at Sun Yat-Sen University. At its heart lies a deliberate break with convention. Instead of building a graph in which similar cells are linked to one another, scFormer constructs a heterogeneous graph in which cells are connected directly to genes. Each cell is linked to a compact set of roughly twenty genes whose expression is highly specific to it, quantified through gene-wise Z-scores after normalization. These highly specific marker genes act as informational bridges: even when a rare cell has no close cellular neighbors, it remains anchored in the graph through the genes that define it. Rare transcriptional programs are therefore embedded in the topology itself rather than inferred from global neighborhoods that may not exist.

Once the cell-gene graph is built, a heterogeneous graph transformer takes over. The transformer distinguishes explicitly between cell nodes and gene nodes, and between gene-to-cell and cell-to-gene edges, applying node-type-specific and relation-specific attention as information propagates through the network. This matters because cells and genes are fundamentally different biological entities, and a model that treats them identically risks blurring the very distinctions that make rare states detectable. The architecture is trained with a multi-component objective that combines a KL-divergence-based reconstruction regularizer, a clustering loss anchored to Leiden-generated pseudo-labels, an intra-cluster compactness term that pulls cells of the same group together in embedding space, and an optional batch-correction term.

That last component addresses one of the thorniest dilemmas in modern single-cell analysis. Multi-sample studies are essential for statistical power, but batch effects from differences in protocol, sequencing depth, and sample handling can swamp genuine biology. Standard integration tools such as Harmony, BBKNN, ComBat, and MNN Correct are built to align global distributions across samples, not to preserve weak rare-cell signals, and over-correction can silently merge true rare populations into transcriptionally similar abundant types. Conversely, running rare-cell detection without integration risks mistaking technical variation for biological rarity. scFormer tackles this tension by jointly optimizing batch correction and rare-cell discovery within a single framework, combining global distribution alignment with a local-structure term that preserves fine-grained neighborhood relationships in the embedding.

The benchmarking effort behind the study is substantial. Across 125 simulated datasets generated with Splatter, each containing 2,500 cells and 5,000 genes with a rare population at an extreme 99:1 imbalance, scFormer achieved the highest average F1 score of 0.988 with strikingly low variance. When the simulations were stratified by difficulty, separating cases where rare and abundant populations differed only subtly, scFormer remained stable while several competing methods, including RaceID3, FiRE, and GiniClust3, degraded sharply. Against seven established rare-cell detection methods spanning rarity-scoring, differential-expression, clustering-based, and feature-learning strategies, the new framework combined accuracy with a robustness that the authors attribute directly to its specificity-driven graph topology.

Real data told a similar story. On 18 publicly available scRNA-seq datasets spanning immune, epithelial, nervous system, and multiple organ and disease settings, with sizes ranging from 313 to nearly 50,000 cells, scFormer achieved the best overall concordance with published rare-cell annotations, posting a mean F1 score of 0.595. That represents a 22.1 percent improvement over the next-best method, scCAD, and a 93.8 percent improvement over GiniClust3. In a mouse retina dataset, scFormer reached an F1 score of 0.955, far exceeding its competitors, and it remained the top performer even in the notoriously difficult human renal cell carcinoma setting. Notably, scFormer completed successfully on all 18 real datasets, while every competing method failed on at least one, a quiet but meaningful indicator of practical reliability.

The biological payoffs are where the method becomes genuinely exciting. Applied to mouse airway epithelium, scFormer recovered the known rare populations, including Foxi1-positive ionocytes and goblet cells, but went further, resolving a rare club-cell-associated cluster comprising just 0.92 percent of cells that scCAD missed entirely. This cluster showed significantly elevated G2/M cell-cycle scores, and control analyses removing canonical S/G2M genes confirmed the population was not merely an artifact of proliferation-driven clustering. In regenerating mouse intestinal crypts, the framework recovered the elusive clusterin-expressing revival stem cells that transiently replace depleted LGR5-positive stem cells after injury, and it uncovered two previously uncharacterized rare lymphocyte subsets, a B-cell population and a plasmacytoid dendritic cell subset, that vanished after irradiation and were absent from the original study’s annotations.

The multi-sample PBMC benchmark demonstrated the batch-correction advantage concretely. On a dataset of 30,669 cells across four batches, scFormer achieved an iLISI score of 2.695 and a batch entropy of 1.085, both exceeding Harmony, while simultaneously identifying five biologically plausible rare populations: B1 B cells, proerythroblasts, transitional B cells, lymphoid progenitors, and plasma cells. The failure modes of the alternatives were instructive. scCAD, lacking batch correction, misclassified plasmacytoid dendritic cells as rare based on technical variation, while the Harmony-Leiden pipeline merged hematopoietic stem cells and transitional B cells with lymphoid progenitors into a single artifactual cluster. An ablation analysis showed that the batch-correction components improved batch mixing substantially without sacrificing rare-cell recovery.

Finally, the authors extended scFormer to spatial transcriptomics using the sci-Space dataset of embryonic mouse sections. Without using spatial coordinates as input, the method identified rare cell states whose positions could then be mapped back onto the tissue. In the hepatic region of one section, it distinguished a rare biliary epithelial-like state co-expressing Cftr, Krt7, and Sox9, showing 35.74-fold spatial self-enrichment, alongside a pancreatic marker-enriched rare state showing 89.52-fold enrichment. Rare neurodevelopmental populations, including Foxg1-positive forebrain lineage cells, Slc1a3-positive radial glia, and Sox2-positive neural progenitors, were likewise resolved and spatially localized. The team notes that scFormer is an RNA-first framework, and that extending the specificity-driven graph to chromatin accessibility or other modalities is a natural next step. With the code publicly available on GitHub, the framework offers researchers across developmental biology, immunology, and precision medicine a practical new lens for spotting the rare cells that matter most.

Subject of Research: A heterogeneous graph transformer framework, scFormer, for detecting rare cell states in single-cell and spatial transcriptomic data

Article Title: Specificity-driven cell-gene graph learning identifies rare cell states in single-cell and spatial transcriptomic data

Article References: Huang, J., Xia, X., Luo, F., Qu, L., Feng, X., & Zheng, L. (2026). Specificity-driven cell-gene graph learning identifies rare cell states in single-cell and spatial transcriptomic data. Advanced Biotechnology, 4(3), Article 26. https://doi.org/10.1007/s44307-026-00121-y

Image Credits: AI Generated

DOI: 10.1007/s44307-026-00121-y

Keywords: scFormer, rare cell identification, single-cell RNA sequencing, heterogeneous graph transformer, spatial transcriptomics, batch effect correction, cell-gene graph learning, revival stem cells, airway epithelium, PBMC integration, graph neural network, computational biology