Functional Subtypes of Synaptic Dynamics Revealed by Model-based Classification

Synapses show preferential responses to particular temporal patterns of activity. Across synapses, however, there is a large degree of response heterogeneity that is routinely and tacitly separated into classes. Furthermore, feedforward neural networks incorporating simple models of these dynamic responses outperform similar recurrent neural network (RNN) models in several neuroscience related tasks. Here we combined a kernel-based model and machine learning techniques to infer biologically grounded, and functionally distinct classes of short-term synaptic dynamics in data from the Allen Institute Synaptic Physiology data set. To identify the presence of classes of synaptic dynamics, we independently performed unsupervised clustering of model parameters and supervised subclass prediction followed by inspection of functional classes used in prediction. In rodent data, we found a remarkable convergence onto six functional classes. Three of these groups corresponded to different types of increasing synaptic strength and were labelled “strongly facilitating”, “depressing then facilitating”, and “short duration facilitating”. The other three groups corresponded to different intensities of strength decrease and “no plasticity” synapses. Application of the same clustering methods in human data inferred the same classes. Next we examined to what extent our inferred groups corresponded to high level transcriptionally defined clusters in rodent data. Better prediction of transcriptomic subclass can be achieved by using model parameters than by using standard electrophysiological features, and suggests that our functionally defined clusters align with specific predicted pre- or post-synaptic transcriptomic subclasses. We propose that these functional types of synaptic dynamics shape the connectivity of the brain and that, paired with our model fits, will be useful as readily available, biologically plausible building blocks for computational simulations.