Accurate and memory-efficient cell type annotation from multimodal single-cell RNA and protein data

AbstractAccurate cell annotation remains a central challenge in single-cell analysis, particularly when datasets contain rare populations, transitional states, and tissue-adapted phenotypes. We previously developed scODIN as an expert-guided framework for immune cell annotation in single-cell RNA-sequencing data. Here, we present pyODIN, a major extension of this framework for Python. pyODIN supports multimodal annotation using RNA, antibody-derived tag (ADT), or combined RNA-ADT information. It also substantially expands the annotation database from a CD4 T-cell-centred framework to a broader cell-type reference resource spanning major and minor peripheral blood cell populations as well as tissue-associated subsets. We benchmarked pyODIN against CellTypist, MMoCHi and HiCAT using independently curated immune reference datasets. pyODIN demonstrated the highest overall classification accuracy, and at the top-lineage level, pyODIN reduced cross-lineage errors relative to the benchmark methods. At finer resolution, pyODIN preserved substantially greater CD4 T-cell subtype structure than CellTypist and HiCAT, resolving regulatory, follicular, helper, memory, and cytotoxic states that were collapsed into broader categories by the comparator methods. In a controlled marker-ablation experiment, pyODIN outperformed MMoCHi when key lineage-defining RNA markers were removed from the expression feature set, and the addition of ADT information restored NK and CD8 T-cell annotation, demonstrating the value of multimodal annotation when transcript-level evidence is incomplete. Finally, application to liver fine-needle aspirate data showed that the expanded framework supports annotation beyond PBMC datasets. Together, pyODIN provides an expert-guided, adaptable cell annotation framework for Python, designed for multimodal cell phenotyping across blood and tissue single-cell datasets. pyODIN is freely available at https://github.com/SondergaardLab/pyODIN.