Fiers, Tomas
(2024)
From Voltage to Wiring: Synaptic connectivity inference from neural voltage recordings.
PhD thesis, University of Nottingham.
Abstract
Systems neuroscience studies the links between (1) an animal’s behaviour, (2) the activity of its neurons, and (3) how these neurons are connected. Currently, only the first two can be observed simultaneously, using 'in vivo' recordings of neural activity. Observing the connections between neurons, on the other hand, requires brain slice imaging, which can only be done post-mortem. However, we might be able to *infer* the connectivity in-vivo, based on neural activity recordings. Historically, this has only been attempted using spike-timing data, with limited success.
In this thesis, we explore the possiblity of using neural *voltage* recordings for connectivity inference (and, specifically: postsynaptic potentials), instead of using only spikes. This is done in the context of recent improvements in in-vivo voltage imaging technology. We have tested this idea (network-inference from neural voltage signals) using simulated data, and we developed three new voltage-based connection detection methods.
We also extensively tested a simple existing algorithm -- namely measuring the height of the spike-triggered average (STA), which is a reflection of the postsynaptic potential -- under varying excitatory-inhibitory (EI) conditions, and both in a simplified 'N-to-1' setup, and in a fully recurrent network.
We find that voltage-based network inference seems feasible to a limited extent, under realistic voltage imaging conditions (6500 EI-balanced inputs to one neuron, 10-minute recording, voltage imaging SNR of 40): the simple algorithm performs considerably better than chance. Two of our newly developed inference methods perform better than the simple algorithm. The best-performing one correlates STAs with a 'template', which is obtained through a first pass of the simple algorithm. This detection method reaches an AUC value of 0.53 for 6500 inputs (chance level AUC: 0.25). At a false positive rate of 5%, we detect 33% of the neuron's inhibitory inputs, and 13% of its excitatory inputs. In absolute terms, this performance is not stellar. But we must note the high number of inputs that we used here (up to 6500). Studies of spikes-only connection-detection methods typically only test networks with about 100 inputs per neuron. Our methods have near perfect detection performance for up to 400 inputs per neuron.
We conclude that voltage-based network inference seems useful for (1) inferring simple networks (low number of inputs per neuron), and (2) finding the high-firing connections in more complex networks.
Actions (Archive Staff Only)
 |
Edit View |
Tools
Tools
