INTENSE: Detecting and disentangling neuronal selectivity in calcium imaging data

TL;DR

This work presents INTENSE (INformation-Theoretic Evaluation of Neuronal SElectivity), an open-source framework that uses mutual information to detect neuron-behavior associations from raw calcium fluorescence data and reveals robust selectivity to multiple variables and refines mixed-selectivity estimates.

Abstract

Neurons encode information about the environment through their activity. As animals explore the environment, neurons rapidly acquire selectivity for distinct features of the external world; characterizing how these selectivity patterns emerge, reorganize, and overlap is key to linking neural activity to behavior and cognition. Calcium imaging in freely behaving animals can record large neuronal populations, but quantifying neuron-behavior selectivity directly from continuous fluorescence is challenging because both signals are temporally autocorrelated and calcium kinetics introduce time lags. Here we present INTENSE (INformation-Theoretic Evaluation of Neuronal SElectivity), an open-source framework that uses mutual information to detect neuron-behavior associations from raw calcium fluorescence data. INTENSE controls false discoveries using circular-shift permutation testing that preserves temporal structure and optimizes temporal delays to account for indicator kinetics and prospective/retrospective encoding. To separate genuine mixed selectivity from associations driven by behavioral covariance, INTENSE applies conditional mutual information-based disentanglement. We validated INTENSE on synthetic datasets, demonstrating robust detection across diverse signal-to-noise ratios and reliability conditions, whereas methods lacking temporal controls show poor performance. Applied to CA1 miniscope recordings in mice freely exploring an open field, INTENSE reveals robust selectivity to multiple variables (place, head direction, object interaction, locomotion) and refines mixed-selectivity estimates by distinguishing redundant from genuinely multi-variable encoding. Together, INTENSE enables high-throughput, information-theoretic selectivity mapping with principled control of temporal structure and behavioral covariance, bridging large-scale recordings to circuit-level hypotheses.

Figures (5)

• Figure 1: A: Examples of rescaled $dF/F$ calcium fluorescence traces (blue) with detected events (gray). B: Examples of behavioral/environmental variables used for correlation with neuronal activity: continuous (head direction, top), discrete (object interaction, middle), multidimensional (animal coordinates, bottom). C: Schematic overview of INTENSE pipeline for quantifying neuronal selectivity. D: Left -- example object-interaction cell: fluorescence trace (blue) and interaction periods (green). Right -- distributions of scaled $dF/F$ values inside (green) vs outside (gray) object interaction periods. E: Distribution of neuron-behavior pairs in the "significance-power" space with marginal distributions. Pairs considered relevant are shown in green.
• Figure 2: A: Upper -- Continuous Wavelet Transform map (CWT map) of a fluorescence signal. Cross-scale ridges correspond to detected events. Lower -- fluorescent trace (blue) and detected events (gray). B: Example animal trajectory, colors represent rescaled calcium activity of a single place cell. Stars represent detected events. C: Left -- intersection of identified PC populations from INTENSE (red) and classic analysis (blue) with $p < 0.05$. Right -- the same for all cells identified as PC by any method. D: Example of activity maps computed from raw calcium signals (left) and from detected events (right) with high SSIM score. E: Similarity metrics between activity maps computed from raw calcium signals and from detected events as functions of confidence (measured via $p$-value). Left -- mean squared error, center -- peak signal-to-noise ratio, right -- structural similarity score. Shaded regions indicate confidence intervals. Vertical lines show the $p=0.001$ threshold, as in Fig. \ref{['fig:pc']}F. F: Relative intersection between PC populations computed via INTENSE and classic analysis as a function of confidence (measured via $p$-value). Intersection is computed on cells for which $p_{\text{INTENSE}} \leq p$ and $p_{\text{classic}} \leq p$ simultaneously. Neurons considered as "non-PC" by both analyses are excluded. Shown are intersections as fractions of respective populations (red for INTENSE, blue for classic); Jaccard coefficient $J$=(INTENSE $\bigcap$ Classic) / (INTENSE $\bigcup$ Classic), orange; proportion of all identified PCs (by any method) left after thresholding by $p$, cyan dotted line.
• Figure 3: A: Example "object interaction neuron" and $dF/F$ distributions during interaction (green) and control periods (gray). B: Example "rear neuron" and $dF/F$ distributions during rearing (orange) and control periods (gray). C: Left -- distribution of neurons selective to 1--3 variables. Right -- distribution of optimal delays: negative ($-$), near-zero and positive ($+$). Statistical comparisons are detailed in Methods. D: Head direction tuning map of a multi-selective neuron. E: 2D spatial activity map of this neuron. F: Heatmap of MI between significantly correlated behavioral features (behavior-behavior analysis). Connection strength is represented by color intensity. G: Schematic of the interdependencies between information measures in neurons with entangled selectivity.
• Figure 4: A: Stages of generating pseudo-calcium signal, associated with a continuous variable. 1 -- continuous behavioral time series with selected range of interest; 2 -- randomized binary time series (with $p_{\text{skip}}$); 3 -- heterogeneous Poisson process with two different rates; 4 -- resulting pseudo-calcium signal after convolution with a calcium indicator kernel. For discrete variables, the procedure starts from stage 2. B: $F_1$-score heatmaps for the continuous variable test across the whole parameter range. Methods shown: correlation-based, MI-based, correlation-based with shuffles, INTENSE. C: Same as B for discrete variables (generated starting from stage 2 of A). Methods shown: average-based, MI-based, average-based with shuffles, INTENSE. D: Precision-Recall maps for continuous variable test. Methods shown: correlation-based, MI-based, correlation-based with shuffles, INTENSE. Left -- precision-recall coordinates of all methods for different SNR values (taken at $p_{skip} = 0$). Right -- precision-recall coordinates of all methods for different $p_{\text{skip}}$ values (taken at $SNR = 64$). Vertical lines represent random guesser precision ($PR_{\text{random}} =0.05$). E: Precision-Recall maps for discrete variable test. Methods shown: average-based, MI-based, average-based with shuffles, INTENSE. Left -- precision-recall coordinates of all methods for different SNR values (taken at $p_{skip} = 0$). Right -- precision-recall coordinates of all methods for different $p_{\text{skip}}$ values (taken at $SNR = 64$). Vertical lines represent random guesser precision ($PR_{\text{random}} =0.05$).
• Figure 5: Two possible acyclic interaction graphs between the three considered variables under the condition of weak connection between A and Y (taken from ghassami2017).