Computational Approach for the Reliable Inference of Task States Encoded by Neurons Using Calcium Imaging

Background

Understanding the neural representations underlying distinct behavioral states is a central goal of systems and behavioral neuroscience. A key question is whether, and to what extent, individual neurons in a brain area preferentially encode one task-relevant state over another. Such investigations have been conducted across a wide range of behavioral contexts, including spatial navigation in safe versus aversive zones [1–4], working memory and decision making [5–7], social interactions [8–10], and fear learning [11]. In each case, the encoding preference of individual neurons is quantified by comparing their responses across task-relevant states and classifying neurons as selective or non-selective based on statistical testing.

Calcium imaging, or the imaging of intracellular neural calcium dynamics using genetically encoded calcium indicators (GECIs) such as GCaMP6f, has emerged as a popular method for such investigations, owing to its ability to access simultaneously large numbers of identified neurons, including targeted cell-type or projection-specific subpopulations, even in freely behaving animals [12–13]. However, unlike electrophysiological recordings that directly measure action potentials, calcium signals reflect a convolution of underlying spiking activity with the slow rise and decay kinetics of the calcium indicator [14–15]. This introduces inherent ambiguity in the choice of metric used to quantify neural activity and raises the question of whether continuous fluorescence traces [3,11,16], discrete detected calcium events [2], or convolved representations of those events provide the most reliable basis for inferring neuronal selectivity. This indirect relationship also causes calcium imaging to under-sample sparsely firing neurons [17–18], which may yield too few detectable events to support a reliable activity estimate; the number of events recovered depends further on signal quality and on the event-detection parameters chosen during preprocessing. These dependencies motivate the minimum event-count criterion applied in step B3, which should be re-examined whenever acquisition or preprocessing settings change.

Beyond the choice of neural data representation, studies have also varied in a number of other procedural and analytical choices: whether the animal's position is tracked using a head-centroid [9] or body-centroid [2,19–20] coordinate; whether neural and behavioral data are temporally binned beyond the acquisition frame rate [2,11]; whether the number of data samples from each task state is matched to control for unequal occupancy; and which statistical shuffling method is used to generate null distributions [2–3,5–9,20]. These choices span the full data processing and analysis pipeline and are made routinely across laboratories, often without systematic justification.

Our recent study [21] highlighted the absence of systematic investigation of the effects of diverse parameter choices and introduced an objective framework for evaluating the quality of inference of neural encoding preferences as a function of parameter choice. Using this approach, it revealed that there is substantial inconsistency in inferences across different choices of parameter values along this pipeline. Specifically, applying 120 different combinations of these parameter values to a single calcium imaging dataset produced markedly different selectivity labels for the same neurons, with 96.8% of all pairs of parameter combinations agreeing on fewer than 90% of neurons. Because these divergent outcomes arise from the same underlying recordings, they isolate the effect of the analysis choices themselves. The current protocol paper describes the step-by-step computational pipeline introduced in that study, implemented in MATLAB. It covers all stages, from preprocessed calcium imaging data to the characterization of individual neuronal selectivity. Critically, it implements a data-driven framework for systematically comparing the quality of different parameter combinations using quantitative metrics of accuracy and robustness and for identifying the combination that produces the most reliable inferences. The pipeline was developed and validated using calcium imaging of excitatory (CaMKII⁺) neurons in the medial prefrontal cortex (mPFC) of freely behaving mice navigating the elevated zero maze (EZM) [22], but is designed to be applicable to data obtained using any calcium indicator, in any brain area, with either 1-photon or multiphoton microscopy, in any animal species, and during any behavioral or stimulus presentation experiment. In other words, this protocol is applicable to any calcium imaging dataset obtained in which the goal is to infer the encoding of two discrete task states by individual neurons (as well as by neural populations).