CATNAP: Celltype Annotation by Taxonomy of Nested Adaptive Predictors

Background: Single-cell RNA-seq has transformed biological research by resolving cellular heterogeneity invisible to bulk approaches, revealing fine cell types, elucidating host responses to infection, vaccination, and disease. Objectives: Hierarchical annotation models mirror the hierarchical definition of cell types, but existing tools fix a single classifier family and often feature space across all nodes. We offer a flexible hierarchical cell type annotation framework with a unified interface, supporting several classification methods, user-defined cell-types, node-specific feature selection and classifiers, and pre-trained human PMBC models. Methods: Hereby we introduce CATNAP, a customizable hierarchical cell-type annotation framework. It is fully modular: catnap_core drives the cell type annotation, given a cell type hierarchy tree, specifying which classifier and parameters shall be used for each subtree. Cell type hierarchies, model parameter presets and other configuration files, tailored to human PMBC studies, are available in catnap_preset. New models can be trained with own data or curated bundles of publicly available datasets, using catnap_pretrain. Finally, catnap_evaluation defines standardized metrics and representations for model performance assessment and comparisons among sets of models. Any classifier can be integrated, registering new catnap_core backends. We provide backends for predictors spanning correlation-based (Spearman), distance-based (k-nearest neighbors, nearest mean centroid), classical machine learning (logistic regression, random forest, gradient tree boosting, support vector machine), probabilistic (linear discriminant analysis, scANVI) and foundation models (Geneformer). Results: CATNAP was used for training, and cell-types prediction on six human PBMC scRNA-seq datasets. A scANVI-based hierarchical model achieved 86.42% accuracy and 83.61% macro-F1 on these datasets, and 55.31% accuracy and 51.38% macro-F1 on two additional out-of-distribution datasets. Conclusions: CATNAP is a customizable hierarchical cell-type annotation framework, providing also a framework for benchmarking existing methods, and a prototyping environment for new classifiers. Future work will explore model combinations, and alternative configurations for exploiting the full potential of our customizable hierarchical framework. Poster presented at the Pasteur Network Annual Meeting (PNAM 2026), Laval, Canada, 3–5 November 2026.