ImmunoMap scANVI bundle (immunomap-scanvi-v1)

Everything needed to annotate human blood (PBMC) scRNA-seq with ImmunoMap: two scANVI reference models (coarse, fine), temperature calibration + Unknown thresholds, a distance-to-reference novelty detector, the 3,002-gene list, label tree and QC rules.

• Code and evaluation: https://github.com/sachitsatyal/Immunomap
• Live demo: https://huggingface.co/spaces/sawcheet/immunomap
• Training data: CELLxGENE Census 2025-11-08 (three reference studies); external test studies Hao 2021 and a lupus cohort's healthy controls were never trained on.

Macro-F1 (test = held-out donors; external = unseen studies):

{
  "coarse": {
    "test_aida": 0.9599,
    "val_sound_life": 0.9804,
    "val": 0.966,
    "external_ext_lupus": 0.9479,
    "val_onek1k": 0.9233,
    "test_sound_life": 0.9807,
    "external_ext_hao2021": 0.9074,
    "val_aida": 0.9649,
    "test_onek1k": 0.9176,
    "test": 0.9661
  },
  "fine": {
    "test_aida": 0.9616,
    "val_sound_life": 0.9566,
    "val": 0.9507,
    "external_ext_lupus": 0.9732,
    "val_onek1k": 0.9211,
    "test_sound_life": 0.9548,
    "external_ext_hao2021": 0.8553,
    "val_aida": 0.9623,
    "test_onek1k": 0.9179,
    "test": 0.9478
  }
}

Limitations: human blood only; new cell types close to a known type may be labelled as that type. Research use only.