INTENSE: Information-theoretic discovery of behavioral selectivity of neurons in calcium imaging data
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.
Ines: We're looking at the paper: "INTENSE: Information-theoretic discovery of behavioral selectivity of neurons in calcium imaging data".
Marcus: Here's the gist of the paper. Neurons encode information by selectively representing different stimuli, behaviors, or cognitive states through their activity, forming the basis of the neural code <ref:2603.04622#pg2> The paper 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 <ref:2603.04622#pg2>.
Ines: Core Methodology INTENSE employs mutual information as its core measure, which captures arbitrary statistical dependencies, substantially expanding the range of detectable selectivities <ref:2603.04622#pg2> Specifically, it uses the Gaussian Copula MI (GCMI) estimator
30, 34: , which provides efficient closed-form MI computation (see Methods: Section 3.2.1) <ref:2603.04622#pg2>.
Marcus: Statistical significance is assessed through circular-shift permutation testing with parametric p-value modeling (Fig.
Ines: 1E; see Methods: Section 3.2.2) <ref:2603.04622#pg2>.
Paper summary: Marcus: So to wrap up INTENSE: Information-theoretic discovery of behavioral selectivity of neurons in calcium imaging data, it’s essentially a framework that uses mutual information to detect neuron-behavior associations from raw calcium fluorescence data <ref:2603.04622#pg2>. The paper claims this method allows for high-throughput, information-theoretic selectivity mapping while providing principled control of temporal structure and behavioral covariance <ref:two thousand six hundred three point zero four six two two#pg3.
Ines: It’s about moving beyond just looking at simple correlations to actually understanding the statistical dependency between what a neuron is doing and what the animal is doing <ref:two thousand six hundred three point zero four six two two#pg1. The authors demonstrate that this information-theoretic approach can reveal robust selectivity to multiple variables like place, head direction, object interaction, and locomotion in mouse miniscope recordings <ref:two thousand six hundred three point zero four six two two#pg3.
Yuki: From a population perspective, if we can reliably map these selectivities across different behavioral features using this information-theoretic lens <ref:two thousand six hundred three point zero four six two two#pg1, it helps us build a more rigorous model of how the neural code is shaped by experience and environment in species like mice <ref:two thousand six hundred three point zero four six two two#pg1.
Marcus: The implication for genomics data scientists like myself is that when we look at large cohorts, this method gives us a principled way to test if the observed differences in neuronal activity are truly due to the variable we think is driving it, or just some hidden statistical artifact from the noise <ref:two thousand six hundred three point zero four six two two#pg3.
Ines: So ultimately, INTENSE enables high-throughput mapping with a principled control over temporal structure and behavioral covariance, bridging large-scale miniscope recordings to interpretable circuit-level hypotheses <ref:two thousand six hundred three point zero four six two two#pg1. It’s about making the connection between the calcium signal and the actual behavior much clearer for researchers studying neural circuits <ref:two thousand six hundred three point zero four six two two#pg1.
Conclusion: Ines: So, to wrap up this part, we're looking at INTENSE’s main point: they use mutual information to find out which specific neuron activity links to which animal behavior.
Marcus: Yeah, so it's like taking raw brain recordings and using information theory to figure out the statistical dependency between what a neuron is firing and what the mouse is actually doing in its environment.
Ines: Exactly, because we aren't just looking for simple correlations anymore; we’re quantifying how much information one variable carries about another. The authors use this mutual information to detect those neuron-behavior associations directly from the calcium fluorescence data.
Yuki: And from a population perspective, that means they can map out how different parts of the brain are tuned to specific things like where an object is or how fast it’s moving in a way that's statistically rigorous. That kind of mapping helps us understand how the neural code is shaped by experience and environment across the species.
Marcus: The authors show that this method can pull out robust selectivity even when you have a lot of noise in your data, which is pretty important when you're looking at large cohorts where batch effects are a real concern. It’s about finding true signals amidst all that variability.
Ines: But they also have to be careful with the details, because they use tools like interaction information to separate genuine connections from just statistical noise or redundant variables. It helps them untangle what’s actually driving the behavior versus what’s just there by chance.
Yuki: I wonder what this means for how we interpret those large-scale genetic datasets we're looking at. If we can reliably link specific neural activity to specific behaviors, it gives us a better way to connect those molecular changes back to observable actions in the animal.
Marcus: It opens up a new way for us to test hypotheses about how circuits are organized based on what the animals are actually *doing*, not just what we think they should be doing. That's a big shift for cohort analysis, because it grounds the statistical claims in actual behavioral data.
Ines: So INTENSE is essentially providing a more sophisticated mathematical tool to move from raw neural data to meaningful biological hypotheses about behavior encoding. It’s moving us closer to understanding the functional role of specific neurons in real-time exploration.
Institute for Advanced Brain Studies of the Moscow State University · Center for Interdisciplinary Research in Biology, Collége de France, Paris, France · Semenov Institute of Chemical Physics, Russian Academy of Sciences
q-bio.NC, q-bio.QM
Submitted: 2026-03-04
Updated: 2026-11-07
Code: https://github.com/iabsneuro/bearmind
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 92/100
The gist: Neurons encode information by selectively representing different stimuli, behaviors, or cognitive states through their activity…
Key concepts
- Mutual Information (MI)
- A measure of how much knowing one variable (like a neuron's activity) reduces uncertainty about another variable (like a specific behavior). It captures complex statistical dependencies between neural signals and behaviors, allowing the framework to detect subtle encoding patterns.
- Gaussian Copula MI (GCMI)
- A specific mathematical estimator used within INTENSE for calculating mutual information efficiently. It provides a way to compute the statistical dependency measure between two variables using a flexible distribution model, which is crucial for accurately quantifying selectivity in this research.
- Interaction Information (II)
- A tool used to distinguish genuine multi-variable encoding from misleading associations. It helps researchers determine if two behaviors are truly independent or if they are statistically redundant (carrying the same information), helping to isolate the unique contribution of each feature.
Terminology
Summary
Neurons encode information by selectively representing different stimuli, behaviors, or cognitive states through their activity, forming the basis of the neural code <ref:2603. p. 1 Together, INTENSE enables high-throughput, information-theoretic
selectivity mapping with principled control of temporal structure and
behavioral covariance, bridging large-scale miniscope recordings to interpretable circuit-level hypotheses. 1 Main
Neurons encode information by selectively representing different stimuli, behaviors, or cognitive states through their activity, forming the basis of the neural
code [1]. 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.Where the paper says it
Where the paper says it
Core Methodology
INTENSE employs mutual information as its core measure, which captures arbitrary statistical dependencies, substantially expanding the range of detectable selectivities <ref:2603.04622#pg2> Specifically, it uses the Gaussian Copula MI (GCMI) estimator [30, 34], which provides efficient closed-form MI computation (see Methods: Section 3. p. 3 It uses mutual information — which captures any statistical dependency, linear or nonlinear, equaling zero if and only if variables are
independent [31, 32, 33] — as its core measure. Specifically, INTENSE employs
the Gaussian Copula MI (GCMI) estimator [30, 34], which provides efficient
closed-form MI computation (see Methods: Section 3. We validated INTENSE on synthetic datasets with known ground
truth, demonstrating robust detection across a wide range of signal-tonoise ratios and response reliability conditions, whereas methods lacking
temporal controls or relying on linear metrics show poor performance. 1
arXiv:2603. 1), making it feasible to
analyze thousands of neuron-behavior pairs across large populations. Statistical significance is assessed through circular-shift permutation testing
with parametric p-value modeling (Fig. It uses mutual information — which captures any statistical dependency, linear or nonlinear, equaling zero if and only if variables are
independent [31, 32, 33] — as its core measure. Specifically, INTENSE employs
the Gaussian Copula MI (GCMI) estimator [30, 34], which provides efficient
closed-form MI computation (see Methods: Section 3. We validated INTENSE on synthetic datasets with known ground
truth, demonstrating robust detection across a wide range of signal-tonoise ratios and response reliability conditions, whereas methods lacking
temporal controls or relying on linear metrics show poor performance. 1
arXiv:2603. Many hippocampal and cortical neurons show apparent selectivity
to multiple behavioral variables [35, 7], but distinguishing true mixed selectivity
from correlations induced by covarying behaviors is essential for understanding
neural coding principles and the true dimensionality of representations. The
framework’s disentanglement procedure uses conditional MI and interaction
information [36] to separate genuine multi-variable encoding from spurious
associations (Section 3.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Validation and Performance
The framework was validated on synthetic datasets with known ground truth, demonstrating robust detection across a wide range of signal-tonoise ratios and response reliability conditions <ref:2603.04622#pg2> INTENSE revealed robust selectivity to multiple variables (place, head direction, object interaction, locomotion) in CA1 miniscope recordings in mice freely exploring an open field <ref:2603. p. 1 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 miniscope recordings to interpretable circuit-level hypotheses. Mutual information detects associations invisible to linear correlation. Replacing mutual information with Pearson correlation in the otherwise identical
INTENSE pipeline detected 2. Replacing mutual information with Pearson correlation in the otherwise identical
INTENSE pipeline detected 2. 2 times fewer selective neurons and 2. 2 times fewer selective neurons and 2. 9 times fewer
significant neuron-feature pairs.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Disentangling Selectivity
The disentanglement procedure uses conditional MI to separate primary from secondary selectivities, and interaction information (II) to classify feature pairs as redundant or synergistic <ref:2603.04622#pg2>. For example, for the speed–locomotion pair, II was consistently negative (redundant), confirming that these variables carry overlapping information about neural activity <ref:2603. p. 9 Interaction information provided a complementary summary of behavioral
variable associations. For the speed–locomotion pair, II was consistently negative
(redundant), confirming that these variables carry overlapping information about
neural activity. The CMI-based disentanglement procedure resolved a substantial fraction of
these multi-variable associations. After controlling for behavioral correlations,
the number of multi-selective neurons decreased from 1,377 to 949 (15.Where the paper says it
Where the paper says it
Robustness and Limitations
INTENSE demonstrated graceful performance degradation as response reliability decreased, showing robustness to moderate response unreliability commonly observed in real neural data <ref:2603.04622#pg2>. The framework's ability to control false discoveries from temporal structure is critical because the slow dynamics of calcium indicators can inflate apparent selectivity and produce false discoveries without proper statistical controls <ref:2603.04622#pg2>. However, its disentanglement procedure can only control for measured covariates, and unobserved or poorly captured behavioral factors can still induce residual confounds <ref:2603. p. 2 [16], theta modulation confounds speed versus phase coding in the hippocampus
[17, 18]. These examples underscore a general requirement for selectivity analyses:
they must explicitly address correlated covariates and the temporal structure of
neural and behavioral signals.Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
1
The AI system can detect neuron-behavior associations from raw calcium fluorescence data using mutual information to capture arbitrary statistical dependencies, substantially expanding the range of detectable selectivities compared to linear measures.
2
The system can control false discoveries using circular-shift permutation testing that preserves temporal structure and optimizes temporal delays to account for indicator kinetics and prospective or retrospective encoding.
3
The system can apply conditional mutual information-based disentanglement to separate genuine mixed selectivity from associations driven by behavioral covariance, distinguishing true multi-variable encoding from spurious associations.
4
The AI system can perform high-throughput, information-theoretic selectivity mapping with principled control of temporal structure and behavioral covariance, bridging large-scale miniscope recordings to interpretable circuit-level hypotheses.
5
The system can recover diverse functional cell types including place cells, head direction cells, and neurons with genuine mixed selectivity from hippocampal calcium imaging data during free exploration.
6
The system can compare place cell populations identified by INTENSE and classical event-based methods using intersection over INTENSE, intersection over classical, and the Jaccard coefficient across a range of significance thresholds.
7
The system can quantify the similarity between calcium and event-based place fields using Structural Similarity Index (SSIM), mean squared error (MSE), and peak signal-to-noise ratio (PSNR) to assess spatial correspondence.
8
The system can perform two-stage testing to balance efficiency and power, identifying candidate pairs in Stage 1 and validating them with extensive shuffling in Stage 2 while controlling the family-wise error rate at alpha equals zero point one.
Abstract
Neurons in freely behaving animals exhibit selectivity for diverse variables. Calcium imaging makes this selectivity hard to read out: behavioral variables covary, and fluorescence is slow and autocorrelated. We introduce INTENSE (INformation-Theoretic Evaluation of Neuronal SElectivity), an open-source framework that screens every neuron against all measured variables on a common statistical scale. INTENSE applies mutual information directly to fluorescence to detect relationships missed by linear correlation, optimizes neural-behavioral delays, performs autocorrelation-preserving circular-shift tests and examines whether apparent mixed selectivity reflects behavioral covariance. In ground-truth simulations, it reduced false associations caused by temporal structure and recovered more true associations than correlation-based screening. In CA1 recordings from 16 freely exploring mice, INTENSE identified neurons selective for position, head direction, object interaction and locomotor state. Accounting for correlated variables reclassified a large fraction of apparently mixed-selective neuron-session observations. INTENSE turns neural and behavioral recordings into testable maps of neuronal selectivity.
Sources
- Interaction Information for Causal Inference: The Case of Directed Triangle
- Nonnegative Decomposition of Multivariate Information
Related papers
- Self-Supervised Learning of Local Predictive Directions as a Neural Computation
- Synchronized Hearts, Desynchronized Brains: A Nonlinear Dynamics Perspective for Function and Failure
- Message interaction dynamics covary with brain volume across mammalian connectomes
- Planning as Dynamics Relaxation: Hippocampal Recurrent Network Realizes Optimal Goal-Directed Navigation
- The Platonic brain bridge hypothesis: brain networks as architectural priors for multimodal large language models
- BrainPFN: Amortized Brain-Behaviour Prediction via Prior-Data Fitted Networks