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

The tool, called scFormer, is a heterogeneous graph transformer framework designed specifically for the identification of rare cell states. To appreciate why it matters, it helps to understand where conventional approaches break down. Most existing methods for analyzing single-cell or spatial transcriptomic data begin by constructing a neighborhood graph based on intercellular similarity, then propagate information among similar cells to smooth out technical noise and learn robust representations. That strategy works well for abundant populations, which are surrounded by legions of near-identical neighbors. But rare cells often lack similar neighbors altogether. When information is propagated across such a graph, their distinctive expression signatures are prone to over-smoothing, and their representations are gradually absorbed into the dominant populations around them. The rare signal simply dissolves. In multi-sample analyses the problem compounds, because batch correction procedures, which are designed to align datasets collected in different experiments, can further weaken the already fragile signals of low-abundance cell states.

ScFormer’s central contribution lies in redefining how graph connectivity is established. Instead of building a global cell neighborhood first and then hunting for rare populations within it, the method prioritizes gene specificity. In the primary analysis, the model selects twenty highly specific candidate genes for each cell based on Z-scores, and then constructs bidirectional connections between cells and genes accordingly. This is a subtle but powerful shift. Even when a rare cell has no similar cellular neighbors, it can still establish indirect associations with other cells through shared gene nodes. Its specific expression pattern is therefore preserved at the graph construction stage rather than being eroded before analysis even begins. In effect, the genome itself becomes the bridge that keeps unusual cells tethered to the rest of the data.

On top of this specificity-driven graph, scFormer layers a heterogeneous graph transformer that employs attention mechanisms designed for distinct node types and relation types, learning low-dimensional representations of both cells and genes. This architecture builds on a growing family of graph-based approaches for single-cell biology, including heterogeneous graph transformers used for biological network inference, heterophily-aware graph neural networks for cell type annotation and novel cell detection, and single-cell graph transformers for rare population inference in multi-omics settings. ScFormer integrates representation learning, clustering, and optional batch correction into a single unified optimization framework. The clustering component applies discriminative constraints derived from pseudo-labels generated by an initial clustering pass, while simultaneously incorporating objectives for expression reconstruction and intra-cluster compactness, so that the learned representation is faithful to the underlying expression data and pulls cells of the same state tightly together.

For multi-batch datasets, the model can further incorporate losses for global distribution alignment and local structure preservation. This dual objective is designed to mitigate batch effects, the systematic technical differences that arise when cells are processed in separate experiments, while deliberately seeking to retain rare biological signals rather than averaging them away. The distinction is crucial. Traditional integration methods measure success by how well cells from different batches intermix, but aggressive alignment can erase exactly the low-abundance states that biologists most want to find. By optimizing for both alignment and the preservation of local structure, scFormer attempts to walk the narrow line between removing technical artifacts and destroying biological reality.

The benchmarking results are striking. Across 125 Splatter-simulated datasets generated with an extreme 99:1 imbalance ratio, a regime in which the target population constitutes just one percent of all cells, scFormer achieved a mean F1 score of 0.988 with an F1 variance of only 0.008, indicating both high accuracy and remarkable stability. On 18 real datasets, where reference rare populations were defined as cell types with an abundance of no more than five percent according to original study annotations, scFormer achieved the highest mean F1 score among the compared methods at 0.595, reaching 0.955 on the mouse retina dataset. These comparisons included dedicated rare-cell detection tools such as scCAD, a cluster decomposition-based anomaly detection method, putting scFormer’s performance in direct context with the state of the art.

The multi-batch tests may be the most consequential for everyday practice. On a peripheral blood mononuclear cell dataset comprising 30,669 cells across four batches, scFormer achieved an integrated least-squares… more precisely, an iLISI score of 2.695, exceeding the 2.191 achieved by Harmony, one of the most widely adopted batch integration tools in the field. At the same time, the framework delivered an F1 score of 0.675 for rare-cell identification, with precision of 0.977 and recall of 0.516. That precision figure deserves emphasis: when scFormer flags a cell as rare, it is almost always right, a property that matters enormously for experimentalists who must decide which computationally identified populations to spend limited resources validating at the bench. Collectively, these results demonstrate that scFormer delivers stable performance in rare-cell identification across complex, imbalanced datasets and multi-batch data, a combination that has historically forced researchers to choose between integration quality and rare-state sensitivity.

Beyond the quantitative benchmarks, the authors examined whether the method could recover biologically known populations and plausible candidate states in real data, and the results read like a tour of recent single-cell biology. In a mouse airway epithelium dataset, scFormer detected Foxi1-positive pulmonary ionocytes and goblet cells, and identified two candidate states associated with basal cells and club cells respectively, both exhibiting proliferative features. Notably, these populations remained partially detectable even after the authors removed a set of cell-cycle-related genes, suggesting the identifications were not merely artifacts of proliferation signatures. In a mouse intestinal crypt regeneration dataset, the model identified Clu- and Anxa1-expressing revival stem cells, the damage-responsive stem cell population discovered in prior work on the regenerating intestine, as well as low-abundance immune subpopulations with B-cell and plasmacytoid dendritic cell characteristics that were detected almost exclusively in unirradiated samples. These subpopulations exhibited compositional differences associated with tissue injury status, hinting at how rare immune states shift as tissue damage unfolds.

Perhaps the most visually compelling demonstration came from spatial transcriptomics. In a representative section of the sci-Space mouse embryo dataset, an embryo-scale spatial transcriptomic resource, scFormer mapped a biliary-epithelial-like state at 0.92 percent abundance and a pancreatic-marker-enriched state at 0.52 percent abundance back to their original spatial coordinates. The two states exhibited spatial self-enrichment of 35.74-fold and 89.52-fold respectively relative to a permuted background, indicating pronounced spatial aggregation within the section. In other words, these vanishingly rare states were not scattered randomly across the embryo but concentrated in coherent anatomical locations, exactly the pattern one would expect of biologically meaningful, spatially organized cell states. The finding showcases the promise of applying specificity-driven graph learning to spatial data, where the combination of low abundance and spatial context carries rich developmental information.

The significance of scFormer extends beyond accurate identification of rare cell states. By allowing highly specific genes to participate in defining connectivity, the method reduces dependence on high-abundance cell neighborhoods and offers a promising route to rare-state discovery in multimodal atlases, perturbation screens, and larger-scale tissue datasets. As single-cell and spatial atlases continue to grow in size and complexity, ensuring that low-abundance yet biologically meaningful states remain detectable will become increasingly important, both for basic developmental biology and for translational questions ranging from tumor heterogeneity to regenerative medicine. ScFormer represents a viable starting point in this regard. Its subsequent validation across multimodal data, perturbation screens, and clinical samples is expected to further test and extend the value of the specificity-driven approach, and if those validations hold, the humble step of deciding which nodes in a graph should be connected may prove to be where rare biology was hiding all along.

The benchmarking design itself reflects the difficulty of this problem. Simulated data, generated with tools such as Splatter, allow ground truth to be known exactly, which is why the 99:1 imbalance experiments could yield precise F1 measurements; real datasets, by contrast, rely on annotations from prior studies as reference labels, meaning that apparent false positives may sometimes represent genuinely novel states rather than errors. This helps explain why the gap between simulated and real-data performance, from 0.988 to a mean of 0.595, is not necessarily a weakness of the method alone but also a reflection of imperfect reference annotations. The high precision paired with lower recall on the multi-batch PBMC data suggests the framework behaves conservatively, favoring confident calls over exhaustive recovery.

The choice of Z-scores for selecting candidate genes is also notable. A Z-score measures how strongly a gene’s expression in a given cell deviates from its typical behavior, making it a natural way to identify genes that distinguish an individual cell from the population at large. This per-cell, rather than per-cluster, selection is what enables the method to preserve signals from states too small to form their own cluster during initial analysis, breaking the common dependency between clustering quality and downstream rare-cell detection that constrains many existing pipelines.

For experimentalists, the practical implication is that tools of this kind could serve as a screening layer, nominating small, spatially coherent populations for targeted follow-up by markers such as Foxi1 or Clu before committing resources to sequencing or functional assays.

Subject of Research: A specificity-driven cell-gene graph transformer framework for identifying rare cell states in single-cell and spatial transcriptomic data

Article Title: Specificity-driven graph learning for the identification of rare cell states

Article References: Hou, L., & Hu, Z. (2026). Specificity-driven graph learning for the identification of rare cell states. Advanced Biotechnology, 4(3), Article 35. https://doi.org/10.1007/s44307-026-00132-9

Image Credits: AI Generated

DOI: 10.1007/s44307-026-00132-9

Keywords: scFormer, rare cell states, single-cell RNA sequencing, spatial transcriptomics, heterogeneous graph transformer, graph learning, batch correction, clustering, mouse embryo atlas, revival stem cells, pulmonary ionocytes, computational biology