Limitations of principal component analysis as a method to detect neuronal assemblies

The anatomical and functional characterization of neuronal assemblies (NAs) is a major challenge in neuroscience. Principal component analysis (PCA) is a widely used method for feature detection, however, when dealing with neuronal data analysis, its limitations have not yet been fully understood. Our work complements previous PCA studies which, in general, characterise NAs based solely on excitatory neuronal interactions. We analysed the performance of PCA in two neglected scenarios: assemblies containing patterns of neural interactions (1) with inhibition and (2) with delays. The analyses considered two types of artificially generated data, one drawn from a traditional Poissonian model, and the other drawn from a latent multivariate Gaussian model; in both models, data from a behaving Wistar rat was used for parameter tuning. Our results highlight scenarios in which neglecting complex interactions between neurons can lead to false conclusions when using PCA to detect NAs. Also, we reinforce the importance of more realistic simulations in the evaluation of neuronal signal processing algorithms.