A PA-only LDA rule gave weighted F1 of 0.57 for four mixed labels and 0.53 for three response labels in 73 zebrafish CLEM neurons

We tested a released zebrafish classifier that maps anatomy to neural response classes. We trained on 47 photoactivated (PA) neurons and tested on 73 correlative light and electron microscopy (CLEM) neurons from one fish. The source selected its published features with labels from both modalities. We confined recursive feature elimination and feature-count choice to PA cells. Linear discriminant analysis then gave weighted F1 of 0.566 for four mixed response-and-axon-side labels and 0.530 for three response labels. The frozen published feature set gave 0.764 and 0.653. These differences compare protocols; they do not give an exact estimate of historical selection bias. The three-label PA-only rule had balanced accuracy 0.525, above the majority rule value of 0.333. Reverse transfer gave weighted F1 of 0.693 and 0.609. A check with the frozen published features reproduced the released combined-training score of 82.1%. An added soma-position and axon-side rule gave three-label F1 of 0.666. Merging the four-label predictions gave PA-only response F1 of 0.582. These checks limit an interpretation that anatomy lacks functional information. The main limits are one CLEM fish, three PA slow-integrator cells, and a fixed model family. The package has both source inputs, code, source records, all predictions, figures, selection records, and repeat instructions.

zebrafish · structure and function · morphology · functional cell types · feature selection · cross validation · CLEM · evidence accumulation