There are two mutually exclusive holy grails of calcium imaging: First, recording from the highest number of neurons simultaneously. Second, detecting spike patterns with single-spike precision. This blog post focuses on the latter.

Many studies have claimed to demonstrate single-spike detection, but often only under specific conditions or for a subset of neurons. At the same time, nearly as many other studies have demonstrated that such single-spike detection is not possible under their respective conditions.

In our recent preprint, we’ve added systematic analyses based on ground truth recordings as our contributions to this debate. Specifically, we analyzed how single-spike detection depends on calcium indicators (GCaMP8s, GCaMP8m, GCaMP8f; GCaMP6f, GCaMP6m; XCaMP-Gf) and on the noise levels of the recordings.

What I particularly like about our approach is that it does not rely on arbitrary thresholds for false-positive vs. false-negative detections of action potentials. Instead, we trained a deep network (CASCADE) to predict spiking activity in general – optimizing for mean squared error loss when compared to ground truth spike rates.We then applied this network to individual single spike-related calcium transients, allowing us to quantify single-spike detection across calcium indicators and noise levels.

Without giving away all the details, I’ll say that I was pleasantly surprised by the performance of GCaMP8s and GCaMP8m! For the full analyses and more context, check out our preprint: Spike inference from calcium imaging data acquired with GCaMP8 indicators.