Fig. 6
From: Generalized leaky integrate-and-fire models classify multiple neuron types
We identify discrete putative clusters using an iterative binary clustering approach on 645 cells. The top six panels show the summary of clusters obtained by iterative binary clustering using electrophysiological features extracted from the traces and GLIF model parameters. In every panel, each row represents a cluster, and each column a transgenic line. The size of the circle indicates the fraction of cells from a given transgenic line falling into a specific cluster (such that the sum of fractions in a column add up to 1). The dendrogram on the y-axis shows the iterative binary splitting into clusters using the algorithm explained in the text. For each intermediate node, a support vector machine was trained on half the cells at that node and used to classify the remaining cells. The number at each node indicates the minimum percentage of test cells correctly classified over 100 iterations of randomly selected training and test cells. Clustering based on features and using the GLIF3 and GLIF4 model parameters shows separation among lines labeling inhibitory and excitatory cells. In addition, transgenic Cre lines marking Pvalb+, Ntsr1+, Nr5a1+, and Ctgf+cells tend to segregate into distinct clusters. The bottom two panels show two measures of overall clustering similarity: the adjusted Rand index (ARI) in red, and the adjusted variation of information (AVOI) in black. The bottom left panel shows similarity between each set of clusters and the transgenic lines. The bottom right panel shows similarity between each set of clusters and the clusters obtained using the features. An ARI of 1 indicates perfect agreement between partitions, whereas 0 or negative values indicate chance levels of agreement. A positive value of the AVOI indicates agreement between partitions that is better than chance (which is indicated by 0). The gray and pink traces in these two panels show the AVOI and ARI values, respectively, for random subsets of the features containing the same number of parameters as each of the four GLIF models
