Robustness of Spike Deconvolution for Neuronal Calcium Imaging

Area of Science:

• Neuroscience
• Computational Neuroscience
• Biophysics

Background:

• Calcium imaging enables large-scale neural population recordings but requires accurate inference of neuronal firing times.
• Inferring spike times from calcium signals is a significant computational challenge in neuroscience.

Purpose of the Study:

• To compare various algorithms for inferring spike times from calcium imaging data.
• To identify the most accurate and efficient method for spike inference, particularly in large datasets.

Main Methods:

• Comparative analysis of multiple spike inference algorithms using electrophysiological ground truth data.
• Development of a novel benchmark for evaluating algorithms without ground truth, using stimulus-repeat correlations.
• Application of algorithms to large-scale (∼10,000 cells) mouse visual cortex calcium imaging datasets.

Main Results:

• Simple non-negative deconvolution (NND) consistently outperformed other algorithms, including complex supervised methods, on out-of-sample data.
• NND demonstrated superior performance on large-scale datasets using the novel benchmark.
• NND-based methods matched the accuracy of convolutional neural network approaches but were faster and less prone to introducing artifacts.

Conclusions:

• Non-negative deconvolution (NND) is recommended for inferring spike times from calcium traces due to its simplicity, efficiency, and accuracy.
• NND avoids biases inherent in some supervised learning methods, reducing the likelihood of erroneous scientific conclusions.
• The developed benchmark provides a valuable tool for assessing spike inference algorithms in the absence of electrophysiological ground truth.