A synthetic likelihood solution to the silent synapse estimation problem

Abstract

The proportions of AMPA-lacking silent synapses are believed to play a fundamental role in determining the plasticity potential of neural networks. It is, however, unclear whether current methods to quantify silent synapses possess adequate estimation properties. Here, by developing a biophysically realistic sampling model, we assess the performance of a common method, the failure-rate analysis (FRA), in estimating the fraction of silent synapses. We find that the FRA estimator is unexpectedly characterized by strong systematic biases, poor reliability, and the need for prohibitively high sample sizes for adequate statistical power. Key predictions from these in silico simulations were validated by whole-cell recordings from mouse hippocampal neurons. To address the inherent limitations of the FRA formalism, we propose an alternative statistical approach, an approximate maximum-likelihood estimator (MLE) based on Monte Carlo simulations of the experimental methodology. This estimator exhibits no bias and low variance and is thus suitable for precise, accurate and high-throughput estimation of silent synapse fractions. These results reveal that previous FRA-based estimates of silent synapse fraction were erroneously large and highlights the use of computational models of experimental methodologies to improve their accuracy and precision. Significance statementThe plasticity potential of networks of neurons is critically enhanced by the presence of silent (AMPAlacking) synapses. The global fraction of these silent synapses in a network is developmentally regulated as well as altered by drugs of abuse in adulthood, yet current methods can only estimate this global fraction by recording from a small and potentially biased subset of electrically stimulated synapses. We show that a widely used failure-rate calculation returns highly variable results and false positives, and develop an alternate estimator employing synthetic (simulated) likelihood functions exhibiting dramatically decreased bias and variance. This allows investigators to reliably interrogate fine-grained changes in plasticity potential, and simultaneously demonstrates how computational models of methodologies can be adapted into statistical inference tools.