In one fly connectome, pooled mitochondrial shape models gave 68–69% balanced recall of three predicted transmitter labels in unseen cell types

We tested whether mitochondrial shape can identify transmitter labels in cell types absent from model training. The data came from one fly hemibrain. The main test used 3,416 neurons from 68 types and three predicted transmitter classes. Four folds kept test cell types absent from training. We gave each class and each type within a class equal weight in the recall score. A model that combined predictions for individual mitochondria gave 68.3–69.0% recall for unseen types. The range came from two training-weight rules. Recall for known types was 71.7–76.3%. The 95% type bootstrap interval for the class-weighted unseen-type score was 53.3–82.0%. A planned model using cell medians and interquartile ranges fell from 77.8% to 56.3%. Its loss was 21.5 percentage points, with a 95% interval of 7.8–37.8 points. The pooled model gave 12.6 points more recall than this summary model, with a 95% interval of 1.8–24.3 points. It did not clearly exceed a brain-region baseline. These intervals condition on fitted predictions. Only four glutamate types and six GABA types were available. The targets were predictions transferred by cell type from another brain. Literature labels agreed for all 17 main types with exact-name records in the checked source table. The study supports partial transfer of a mitochondrial shape signal and shows that its measured size depends on the model. The package has numeric inputs, code, predictions, figures, source hashes and offline repeat instructions.

connectome · mitochondria · morphology · neurotransmitter prediction · cell type transfer · grouped cross-validation · Drosophila