How reliable is spike train reconstruction from Calcium imaging data?

The intracellular calcium concentration of a neuron is a correlate of its spiking activity. In-vivo calcium measurements of a whole local population of neurons with high temporal resolution have allowed the detection of calcium transients that are triggered by single spikes and spike bursts. As spikes are the basic unit of communication between neurons, it is an important goal to reconstruct the occurrence and timing of spikes from noisy calcium traces. The inferred spike trains can then be used to build statistical population models, such as Generalized Linear Models or maximum-entropy models.

Algorithms that aim to reconstruct spike trains from noisy calcium observations are abundant and range from template-matching algorithms that are inspired from signal-processing algorithms, clustering methods originating from the machine learning literature to ideas from information and coding theory. Additionally, many of these algorithms have a number of adjustable parameters. To our knowledge, the performances have never been systematically compared on the same data sets. We therefore create artificial calcium signals using a biophysically plausible forward model with varying noise levels. Different algorithms are evaluated on the same data set for which the ground truth (the biophysical model) is known.

One important aspect is to find an appropriate performance metric, i.e. how to choose a trade-off between false positives and missed spikes and which temporal precision is acceptable for the reconstruction. Furthermore, using these benchmarks, it can be explored which minimal signal-to-noise ratio is necessary to obtain a faithful spike train reconstruction. These results can aid experimentalists to optimize their set-ups and sampling schemes. Once the spike trains are reconstructed, subsequent analysis takes this data as input e.g. to reconstruct causal connectivity structures [1] or discover features in the stimulus that the neurons are tuned to.

References:
[1] Gerhard F, Pipa G, Lima B, Neuenschwander S and Gerstner W (2011) Extraction of network topology from multi-electrode recordings: is there a small-world effect? Front. Comput. Neurosci. 5:4. doi: 10.3389/fncom.2011.00004

Acknowledgments: This work is done in close collaboration with the group of Professor Fritjof Helmchen at the University of Zurich, Switzerland. Felipe Gerhard acknowledges support by the Swiss National Science Foundation (SNSF) under grant number 200020-117975.