scHDeepInsight

Immune cells resist flat classification. A regulatory T cell is a kind of CD4 T cell is a kind of T cell, and the transcriptional distance between neighboring subtypes is often far smaller than the distance between lineages. Annotation methods that treat all labels as mutually exclusive alternatives throw that structure away, which shows up as confident but anatomically nonsensical calls and as poor recall on rare subtypes.

scHDeepInsight, from Tatsuhiko Tsunoda's Laboratory for Medical Science Mathematics at the University of Tokyo with collaborators at the RIKEN Center for Integrative Medical Sciences, Griffith University, and Korea University, builds the hierarchy into both the architecture and the loss. It extends the group's earlier scDeepInsight model, which converts a cell's expression vector into a two-dimensional image so that convolutional networks can exploit spatial locality among co-expressed genes. The new version adds a two-stage classifier — base immune type first, then subtype within that lineage — and an adaptive hierarchical focal loss that rebalances the two levels during training according to which is currently harder.

The result is a pretrained reference classifier rather than a per-dataset fit: a query dataset is batch-corrected onto a fixed immune atlas and passed through the frozen network, resolving more than 50 subtypes. Across seven benchmark datasets it averages 93.2% subtype accuracy, 5.1 percentage points above the next-best method. The work was posted to bioRxiv in June 2025 and published in Briefings in Bioinformatics.

#Key Features

• Expression rendered as images: Genes are assigned to pixel coordinates via t-SNE with a linear sum assignment step guaranteeing a unique location per gene, turning each cell into a 224x224 image that a CNN can process.
• Two-stage hierarchical classification: A base-type head narrows the lineage and probability masking zeroes out subtypes outside it, so predictions cannot violate the immune hierarchy.
• Adaptive hierarchical focal loss: The weighting between base and subtype objectives is adjusted dynamically according to relative difficulty, with a focusing parameter of 3.0 to counter class imbalance across rare subtypes.
• Reference-mapped inference: Query data is aligned to the reference atlas with STACAS batch correction, and random gene masking during training makes the model tolerant of genes missing from a query panel.
• Interpretability and rare-cell flags: SHAP attributions are reported separately at base and subtype level, and low prediction confidence is surfaced as a candidate rare population rather than forced into a label.

#Technical Details

The backbone is EfficientNet-B5 operating on 224x224x3 images built from the top 5,000 highly variable genes. The reference atlas integrates ten published studies covering blood, lung, intestine, and other tissues — over 460,000 cells from healthy donors — organized into 15 base immune types and more than 50 subtypes. Benchmarking across seven tissue datasets gives 93.2% average accuracy, 91.1% precision, 90.5% F1, and 89.7% AUPRC at subtype level, improvements of 5.1, 3.3, 3.1, and 3.6 points respectively over scDeepInsight, the strongest baseline. Comparisons also cover SingleR, Azimuth, CellTypist, Garnett, scType, and GPTCellType. The full pipeline, including batch correction and image transformation, annotates 45,000 cells in roughly ten minutes on a GPU and also runs on CPU.

#Applications

The model targets immunology groups profiling patient or model-organism samples who need consistent subtype labels rather than de novo clusters — immune infiltration in tumors, vaccine and infection response, autoimmune tissue profiling. Distribution as a pip-installable package with an R-side preprocessing step makes it straightforward to slot into an existing Seurat or Scanpy workflow, and the hierarchical output lets an analyst fall back to a confident base-type call when the subtype call is uncertain.

#Impact

scHDeepInsight is a useful counterpoint to the transformer-based single-cell foundation models such as scGPT and Geneformer: a comparatively small convolutional model, specialized to one well-curated domain, can beat general-purpose approaches on the task it was built for. Its limits follow from that specialization. Accuracy is bounded by the coverage of the reference atlas, so cell states absent from it cannot be recovered; its scope is the human immune compartment; and the image transformation adds preprocessing overhead relative to vector-based classifiers. The trained weights are pulled at runtime from Hugging Face and the 288 MB reference atlas ships in the repository via Git LFS, but the GitHub repository carries no license file — the PyPI package declares MIT through a classifier alone — so reuse terms for the code and weights are undefined.