The monosynaptic inference problem: linking statistics and dynamics in ground truth models Zach Saccomano, Rodrigo Pena, Sam Mckenzie, Horacio Rotstein, Asohan Amarasingham
Given the ability to record spike trains from populations of neurons, a natural aim in neuroscience is to infer properties of synapses, including connectivity maps, from such recordings. These inferences can derive from observations of strong millisecond-timescale correlations among spike train pairs, as typically reflected by a sharp, short-latency peak in the causal direction of cross-correlograms (CCG) between the reference and target neurons. However, such sharp peaks may also occur when two disconnected neurons systematically fire close together in time in the absence of a direct monosynaptic connection. A further confound is that a monosynapse likely influences the postsynaptic cell on broader timescales as well. These observations motivate a systematic analysis of how a monosynapse exerts influence on the intrinsic dynamics of its postsynaptic target and how this affects the properties of the CCG and the ability to infer the monosynaptic properties. In previous work [1], we adapted a statistical framework for monosynaptic inference based on a (statistical) separation-of-timescale principle, in which monosynaptic interactions are systematically assumed to drive spike-spike correlations at finer timescales than non-monosynaptic interactions. We examined this principle in a simplified ground truth neuron model with minimal intrinsic dynamics, such as the leaky integrate-and-fire (LIF) model with an adaptive threshold. In this work, we extend these ideas to more realistic models and multiple time scales. We use a generalized LIF model with two-dimensional subthreshold dynamics and multiple (dynamic) time scales. The model describes the nonlinear dynamics of the voltage and a slower adaptation gating variable. These subthreshold dynamics also describe the onset of spikes, but not the spiking dynamics. Spikes are added manually. Our previous work exploited our ability to study counterfactual causal inferences in simulations. For example, in simulations with two neurons and comodulated noise, how much would the peak of the CCG change if the monosynapse were deleted? Here we extend this approach to the more complex models where the properties of the CCG are affected by the model nonlinearities and time scales. In this scenario, the model’s slow time scale (captured by the adaptation time constant) affects the CCG time scales (an emergent property of the monosynaptic interaction). Finally, we assess how bias induced by the separation-of-timescale principle (in the statistical sense) depends on the intrinsic dynamics of postsynaptic cells, in particular on the separation of time scales in the dynamic modeling sense.
** **
Reference
[1] Platkiewicz, J., Saccomano, Z., McKenzie, S., English, D., & Amarasingham, A. (2019). Monosynaptic inference via finely-timed spikes. arXiv preprint arXiv:1909.08553.
