| | --- |
| | license: mit |
| | tags: |
| | - neuroscience |
| | - connectomics |
| | --- |
| | |
| | Random forest models trained on heat kernel signature (HKS) features to predict the compartment target of a synapse on neuronal meshes. |
| | Method is described in more detail in [Pedigo et al. bioRxiv 2026](https://www.biorxiv.org/content/10.64898/2026.02.19.706834v1). |
| |
|
| | The feature generation pipeline used here can be found in the [meshmash repo](https://github.com/bdpedigo/meshmash/blob/7395e0d8d3b0e3fb23069761961eb7055b7b9700/src/meshmash/pipeline.py#L389). |
| | The parameters for that function are in `parameters.toml`. |
| |
|
| | `synapse_hks_model.joblib` takes in the 32 HKS features generated by that pipeline, as well as a column of the distance from each point to the postsynaptic nucleus. |
| | The model outputs predictions for {"spine", "shaft", "soma"} targets. This is the model used in the paper. |
| |
|
| | `simple_hks_model.joblib` uses only the HKS features and outputs {"spine", "not_spine"} labels. |
| | |
| | Please see the paper above for much more detail on the performance of the classifier and its application. Questions and comments are more than welcome: ben.pedigo@alleninstitute.org. |
