Chemical connections were detected manually, curve fit, and then characterized for strength, kinetics, and short-term plasticity in a multi-stage process. Manual connectivity calls were used to train a machine classifier (see Seeman, Campagnola, et al. 2018 for more detail). After training, the machine classifier revealed a set of potential false positives or false negatives that were manually re-checked. Electrical connections were also detected and characterized.
Chemical connections were visually identified from average postsynaptic responses sorted by recording mode (voltage and current clamp) and membrane potential (see QC criteria for holding potential ranges for excitatory and inhibitory connections). Individual responses were time aligned to the presynaptic spike. Responses that failed QC were not included in the average.
|Users could visualize connections sorted by recording mode and membrane potential, then move the yellow line to set the onset of the response. The Fit Response button would then generate a fit constrained to ± 100 µs of the user-defined latency.|
The User Latency value was used to initialize automated curve fitting of the postsynaptic response. Parameters from the curve fits were shown for each quadrant. When the user was satisified with the curve fitting output, the goodness of fit was manually passed or failed. The user would take into consideration gap junctions, electrical artifacts, the shape of the fit or other factors when deciding whether the curve fit accurately reflected the response. Notes were also made by the user and used to update, and subsequently test, new fitting algorithms.
|In this example, the voltage clamp fits were failed because they could not accurately capture the shape of the response, while the current clamp response was a good fit and passed. No data was collected at depolarized potentials in current clamp and thus is ignored|
Once a chemical connection was identified, the strength and kinetics of that connection, as well as the dynamics (short-term plasticity (STP) and variability), were characterized utilizing the curve fits generated during connection detection.
The "strength" of a connection is an important characteristic particularly when we want to start comparing connections across cell class. However, we know that strength changes dynamically over time (as is highlighted in our analysis of STP) as well as stochastically from spike to spike. Given that our stimulus set utilizes trains of various frequencies we want to ensure that our metric of strength would be useful in comparisons across connections and not be stimulus dependent. Our resting-state strength is a metric that works well for most connections; however, some facilitating connections have a very small resting-state strength which may be misleading. Rise and decay kinetics are more faithfully preserved even as the amplitude of the connection changes and thus, was more straightforward to characterize.
Resting-state amplitude was determined from responses in which the presynaptic spike follows a period of quiescence. Usually this is the first pulse in each stimulus train. Individual responses were averaged and curve fit; the amplitude output of the fit served as our metric of strength.
|Average resting-state responses in current and voltage clamp. Each trace is the resting state average PSP/C from a single connection. Traces were fit and the fit amplitude is plotted above.|
Latency, rise and decay kinetics were also determined from curve fits to average synaptic responses. Kinetics were calculated for postynaptic currents (PSCs) and potentials (PSPs) held at either -70 or -55 mV depending on whether the synapse was excitatory or inhibitory (see QC requirement).
Short-term plasticity was analyzed from data recorded in current clamp utilizing different stimuli depedending on the analysis. We measured three main metrics of short-term plasticity: paired-pulse ratio (PPR), train-induced STP, and recovery from train-induced STP.
Short-term plasticity is often calculated as a ratio of response amplitudes for each pulse in the stimulus train. Many connections in our dataset have small responses that are close to the noise level in our recordings making the use of an amplitude ratio unstable. This situation is accentuated for connections that are strongly depressing or facilitating resulting in spurious ratio measurements. To avoid this instability, we use the 90th percentile PSP amplitude as an approximation of the "maximum" amplitude of a synapse and normalize our STP measurements by this value. Below are the calculations for each STP metric.
Paired-pulse STP was measured from the 50Hz stimulus as:
(Avg 2nd pulse - Avg 1st pulse) ÷ 90th percentile amplitude
Train-induced STP was measured from the 50 Hz stimulus as:
(Avg (6th, 7th, 8th pulses) - Avg 1st pulse) ÷ 90th percentile amplitude
Recovery from STP was measured from all stimulus frequencies with a 250 ms delay between the last induction pulse (8th) and first recovery pulse (9th):
(Avg (9th, 10th, 11th, 12th pulses) - Avg (1st, 2nd, 3rd, 4th pulses)) ÷ 90th percentile amplitude
The amplitude of PSPs for a synaptic connection vary randomly each time neurotransmitter is released. This is often reported using the coefficient of variation (CV). In the mouse cortex, however, typical PSP amplitudes can be much smaller than the background electrical noise in the cell. CV in this regime is thus dominated by noise and tells us little about the physiology of the synaptic connections. In our dataset, PSP amplitude variability is reported using a metric that is adjusted (aCV) to correct for the effect of background noise:
This metric has a value of 1.0 when the standard deviation of the PSP amplitudes (after noise correction) is the same as the median amplitude. The noise correction itself introduces a new source of variance, however, which can sometimes lead to this value being negative. We measure variability in the resting state for each connection as well as in various states of induced short-term plasticity.
|Variability in postsynaptic response amplitude from trial-to-trial (black traces) with average (blue) response|
We developed a new model of synaptic vesicle release that provides a more comprehensive description of each synaptic connection and also allows to predict the behavior of the synapse in response to arbitrary stimuli. The model includes basic quantal release parameters (release probability, number of release sites, and quantal size) as well as short term plasticity (vesicle depletion, depression, and facilitation with varying recovery time constants) and accounts for recording noise.
Our dataset includes best fit parameters (using a maximum likelihood estimation) of the model for many connections that can be used to simulate different types of connections, or as a basis for comparison between connection types.
Electrical connections, formed by gap junctions, were also detected in our dataset. We used the long-pulse stimuli to characterize the strength of electrical connections.