One of the assumptions of biology is that structure should predict function. Classifying neural cells is no exception: we classify cells that look alike (Purkinje cells, spiny neurons, etc.) on the assumption that they will function alike as well. Otherwise the classification would serve no purpose.
But reversing the causality, and attempting to classify neurons on the basis of just their structure (as opposed to, e.g., simpler staining methods), has proven to be a difficult endeavor. There is little consensus to the taxonomy. One problem is that there is no established set of geometrical measurements which should be used.
Zawadzki et al mined data from NeuroMorph, an online database with quantitative structural data from 5000+ neurons. First, here is a diagram of the structural data that NeuroMorph provides for each neuron:
When authors upload their neurons to NeuroMorph’s database, they often give the cell class that they have assigned to each neuron. The most common of these neuron classes are: 1) pyramidal cells from the hippocampus (Pyr-Hip), 2) medium spiny cells from the basal forebrain (Spi-Bas), 3) ganglion cells from retina (Gan-Ret), 4) uniglomerular projection neurons from olfactory bulb (Uni-Olf). Zawadzki et al then compared these classifications with the results of a naive statistical algorithm that categorizes cells solely based on the similarity and differences of their structures. The “temperature” result that their algorithm spits out seems to be basically the result of a principal component analysis.
As you can see, on the basis of this temperature, there is plenty of overlap between the classes:
The pyramidal neurons are scattered across the parameter space, suggesting that their morphological features overlap with the other categories. In contrast, the medium spiny neurons have the least overlap and thus have the most morphological homogeneity, indicating that it medium spiny neurons are the easiest class to segregate based on morphology. Overall, it seems fairly difficult to disambiguate neural classes based solely on the structure given some of the most popular current classifications.
One solution to this problem is presented by Bota and Swanson, who suggest an ontological approach to classifying neurons. Here is their hierarchical classification schema:
They also argue that lower level neuron classifications should be regarded as a hypothesis that certain cell taxa fulfill distinct functional roles. This hypothesis can be tested and potentially discarded. Given the infancy of the field of neuron classification, such low barriers to change appear expedient.
Zawadzki K, et al. 2010 Title: Investigating the Morphological Categories in the NeuroMorpho Database by Using Superparamagnetic Clustering. arXiv:1003.3036v1
Bota M, Swanson L. 2007 The neuron classification problem doi: 10.1016/j.brainresrev.2007.05.005. Available in PMC here.