DRIADA: an open-source Python toolkit for linking single-neuron selectivity and population activity to behavior

arXiv:2607.00851 · q-bio.NC, q-bio.QM · Submitted 2026-07-01 · Read on arXiv

Listen

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: "DRIADA: an open-source Python toolkit for linking single-neuron selectivity and population activity to behavior".

Marcus: Here's the gist of the paper. The gist The DRIADA framework presents an open-source Python toolkit that unifies neural signals and time-aligned behavior in a shared data model to enable cross-scale analysis of single-neuron selectivity, population dynamics, and network organization.

Ines: Unified Data Model DRIADA addresses the fragmentation problem by establishing a shared data model where all analysis modules are interoperable The Experiment object serves as the central container holding neural activity matrix (raw calcium fluorescence or spike counts), time-aligned behavioral variables, and per-neuron metadata in a single structure from which any module can draw without format conversion Each Neuron wraps its calcium and spike traces as TimeSeries objects, and the full population is jointly accessible as a MultiTimeSeries that inherits dimensionality-reduction capabilities from its parent class MVData Behavioral variables undergo automatic type detection—continuous, discrete, circular, or multivariate—so that appropriate estimators and preprocessing are selected automatically This automatic handling avoids a common source of errors by ensuring the correct MI estimator is applied to the correctly classified variable Cross-Scale Analysis Modules DRIADA integrates six modules organized around complementary perspectives: single-neuron characterization, population-level geometry, and network structure 1.

Marcus: Single-neuron analysis characterizes individual neurons from two complementary angles: selectivity testing via INTENSE identifies how each neuron’s activity covaries with behavioral features, and network analysis of time series characterizes how each neuron’s dynamics unfold through recurrence networks, visibility graphs, and ordinal partition networks 2.

Ines: Population activity analysis compresses high-dimensional neural recordings into interpretable low-dimensional representations using 15 methods spanning linear, spectral graph-based (Laplacian eigenmaps, diffusion maps), neighborhood-based (t-SNE, UMAP), and neural network-based approaches 3.

Paper summary: Ines: So, we're talking about this paper called "DRIADA: an open-source Python toolkit for linking single-neuron selectivity and population activity to behavior." Basically, the authors are saying that understanding brain function requires looking at neurons, groups of neurons, and networks all together.

Marcus: Right. And what they claim is that current tools are all separate, using different data formats and conventions, so linking them up is really hard. This DRIADA framework creates a shared data model where everything can talk to each other without needing messy format conversions between modules Ines.

Yuki: It seems like the core idea here is bridging the gap between these different scales of analysis, moving from single neurons to population dynamics and network structure all in one system. That’s what makes it potentially useful for understanding how neural activity translates into actual behavior.

Ines: Exactly. They describe this unified data model with an Experiment object that holds the raw neural activity, time-aligned behavioral variables, and per-neuron metadata in one single structure Ines. It allows any module to draw data from this structure without having to convert formats first.

Marcus: And they handle the behavioral variables automatically—continuous, discrete, circular, or multivariate—so you get the right estimators applied automatically for things like mutual information Marcus. That’s a big win because it removes a common source of error when setting up these analyses.

Yuki: From a population genetic perspective, it’s interesting how they are focusing on linking selectivity to behavior. They test this on synthetic populations with known ground truth, calcium imaging from thirteen mice across thirty-nine sessions, and even a simulated continuous attractor neural network Yuki. That gives them a good range of testing grounds.

Ines: And the results on the synthetic data were pretty strong. Across most hippocampal sessions, the embeddings of spatially selective neurons predicted held-out positions better than equally sized random subsets or all recorded neurons Ines. It shows that selectivity actually helps you predict where things are located in the population structure.

Marcus: That’s interesting because it directly connects individual neuron tuning to population geometry. They also show how they related population dimensions back to individual neurons, even including cells that didn't show selectivity for the measured behavioral variables Marcus. It makes sure you account for everything in the system.

Paper summary: Yuki: And in that simulated network, the complementary analyses captured features consistent with its known organization Yuki. This suggests that if you have a known structure like a toroidal organization, these methods can actually recover those features.

Ines: So, what this means for us is that DRIADA provides a common framework for relating neuronal selectivity and population structure and network activity to actual behavioral dynamics Ines. It’s about making these cross-level connections systematic rather than ad hoc.

Marcus: I think the real power of this is in how they bridge those scales through explicit pathways. They show that you can filter the population before doing dimensionality reduction, which lets you directly compare full-population versus selective-population manifolds Marcus.

Yuki: And they also say that embedding coordinates become features for reverse selectivity analysis, where you treat those coordinates as continuous features and test their significance using the same mutual information testing used for behavioral variables Yuki. That’s a clever way to check if the structure matters.

Ines: They even found something interesting in the hippocampus recordings. Scale-bridging analysis showed that selectivity-based filtering transforms a collapsed one-dimensional embedding into a two-dimensional spatial manifold, with the filtered embedding spanning both arena dimensions Ines. That’s a concrete visualization of what they found.

Marcus: And on that same data, reverse selectivity analysis indicated that about fifty-seven percent of neurons informative about leading manifold dimensions were not selective to any of the eleven measured behavioral features Marcus. That shows there's more complexity than just direct behavior-related tuning.

Yuki: It supports the idea that the underlying organization is much broader than just what we measure directly, which aligns with how we see evolution working across different scales in species Yuki. This work on DRIADA seems to give us a way to map out those organizational principles.

Ines: On the simulated toroidal continuous attractor network, four modules—selectivity testing, autoencoder embedding, representational similarity analysis, and recurrence networks—independently converged on the same nonlinear topology Ines. That convergence is what gives confidence in the model they built.

Marcus: The autoencoder using a decorrelation loss achieved a DCI score of zero point nine six for disentanglement in that toroidal network benchmark Marcus. A high score like that shows the tool can actually separate the underlying structure cleanly when it’s known.

Paper summary: Yuki: So, to put it simply, DRIADA demonstrates that selectivity predicts network membership, filtering by tuning improves manifold structure, and temporal dynamics recover the same topology as population geometry Yuki. That’s a really neat summary of their main findings.

Ines: It makes integrated cross-scale analysis accessible to people working with calcium imaging, spike trains, or simulated networks Ines. The toolkit itself is open source Python code, which means other labs can use this unified approach for their own problems.

Marcus: That’s the implication for the community. Instead of everyone building separate pipelines for single-cell analysis and population modeling, they have one framework to start with Marcus. It tackles that fragmentation issue head-on.

Yuki: For a broader perspective, this moves us closer to understanding how different levels of neural organization—individual tuning, population structure, network connectivity—all contribute to the same underlying behavioral system Yuki. It’s about seeing the whole picture instead of just pieces.

Ines: So, when we look at the title again, "DRIADA: an open-source Python toolkit for linking single-neuron selectivity and population activity to behavior," it really captures that goal of unification Ines. It’s not just one analysis method; it's a whole system for connecting these different types of data.

Marcus: And the authors, Pospelov, Plusnin, Rogozhnikova, Ivanova, Sotskov, Orobets, and Toropova are bringing together computational physics and neuroscience to tackle this problem Marcus. That mix of backgrounds seems crucial for building such a comprehensive toolkit.

Yuki: For us in population genetics and systems thinking, it’s valuable because it provides the mathematical machinery to test hypotheses about how selective pressures shape the structure of neural circuits across different scales Yuki. It gives us tools to see if those structures are evolutionarily stable or not.

Ines: The implication is that we can start asking questions that span scales more easily, because the tooling is built to handle the complexity of those connections Ines. It takes the fragmentation out of the process.

Marcus: It's a practical tool, though. If you want to do this kind of cross-scale work without wrestling with entirely separate codebases and manual translations between frameworks, this framework is designed for that Marcus. It’s about making it usable by experimentalists.

Yuki: So, the big picture is that we are getting a better way to map out the entire hierarchy of neural activity, showing how individual choices feed into population dynamics and network organization Yuki. That’s a fundamental piece of understanding for neuroscience.

Conclusion: Ines: So, we're wrapping up on DRIADA and what this whole thing actually means for how we study brain function across different scales.

Marcus: Right, it really comes down to that title, "DRIADA," linking single-neuron selectivity and population activity to behavior. It’s trying to make the connection explicit in a way that used to be super messy.

Yuki: From a population genetic standpoint, I think what's important is how they force you to consider the individual neuron's tuning when you look at the whole group dynamics.

Ines: Exactly, because it shows that selectivity isn't just some extra detail; it predicts where those neurons sit in the population structure.

Marcus: And they’re not just doing this on clean computer models; they tested it on actual mice and simulated networks with known structures, which gives you some real validation on the methodology.

Yuki: I see what you mean about the validation—if you can show that a known toroidal organization recovers when you run their modules, that lends weight to the idea that their framework is capturing something real in those complex systems.

Ines: And what it changes for us is making integrated cross-scale analysis accessible. It takes the complexity out of trying to manually stitch together different ways of looking at neural data.

Marcus: So, basically, this toolkit lets you look at the individual neuron's choices, see how they build up into a population shape, and see how that shape relates to what the animal is actually doing.

Yuki: That’s a big step toward seeing how those different levels of organization—the single cell versus the network—all contribute to one overall system.

Ines: It shifts the focus from just describing what's happening at one scale to understanding how all these scales work together in real time during behavior.

Marcus: And it opens up a new way for experimentalists who use calcium imaging or spike trains to finally connect their raw data directly to behavioral predictions using this unified structure.

Yuki: So, the next question becomes, can we actually use this framework to test hypotheses about evolutionary constraints on how these neural structures form?

Institute for Advanced Brain Studies, Lomonosov Moscow State University · Moscow Physics and Technology Institute · Center for Interdisciplinary Research in Biology, Collége de France · A.P. Nelyubin Institute of Pharmacy, I.M. Sechenov First Moscow State Medical University · Semenov Institute of Chemical Physics, Russian Academy of Sciences

q-bio.NC, q-bio.QM

Submitted: 2026-07-01

Updated: 2026-10-20

Code: https://github.com/iabs-neuro/driada2https:

License: http://creativecommons.org/licenses/by-nc-sa/4.0/

Importance score: 91/100

The gist: The gist The DRIADA framework presents an open-source Python toolkit that unifies neural signals and time-aligned behavior in a shared data model to enable cross-scale analysis of single-neuron

Key concepts

Unified Data Model: DRIADA creates a single data structure called the Experiment object that holds neural signals (like calcium traces), time-aligned behavior, and neuron metadata. This allows all analysis modules to work together seamlessly without needing to convert data formats between different tools.
Single-neuron Analysis: This involves looking at individual neurons from two angles: first, testing their selectivity by seeing how their activity covaries with behavior. Second, it analyzes the neuron's internal dynamics using methods like recurrence networks to understand how its activity unfolds over time.
Scale-Bridging Insights: The framework explicitly links different levels of analysis. For instance, filtering the population based on single-neuron selectivity can transform a low-dimensional embedding into a richer two-dimensional spatial map, showing how individual properties influence the overall structure.

Terminology

Benchmark — 1× in this paper · explained in 2 episodes
Core — 1× in this paper · explained in 2 episodes

Summary

The gist The DRIADA framework presents an open-source Python toolkit that unifies neural signals and time-aligned behavior in a shared data model to enable cross-scale analysis of single-neuron selectivity, population dynamics, and network organization.

Unified Data Model

DRIADA addresses the fragmentation problem by establishing a shared data model where all analysis modules are interoperable The Experiment object serves as the central container holding neural activity matrix (raw calcium fluorescence or spike counts), time-aligned behavioral variables, and per-neuron metadata in a single structure from which any module can draw without format conversion Each Neuron wraps its calcium and spike traces as TimeSeries objects, and the full population is jointly accessible as a MultiTimeSeries that inherits dimensionality-reduction capabilities from its parent class MVData Behavioral variables undergo automatic type detection—continuous, discrete, circular, or multivariate—so that appropriate estimators and preprocessing are selected automatically This automatic handling avoids a common source of errors by ensuring the correct MI estimator is applied to the correctly classified variable

Cross-Scale Analysis Modules

DRIADA integrates six modules organized around complementary perspectives: single-neuron characterization, population-level geometry, and network structure

Single-neuron analysis characterizes individual neurons from two complementary angles: selectivity testing via INTENSE identifies how each neuron’s activity covaries with behavioral features, and network analysis of time series characterizes how each neuron’s dynamics unfold through recurrence networks, visibility graphs, and ordinal partition networks
Population activity analysis compresses high-dimensional neural recordings into interpretable low-dimensional representations using 15 methods spanning linear, spectral graph-based (Laplacian eigenmaps, diffusion maps), neighborhood-based (t-SNE, UMAP), and neural network-based approaches
Network analysis characterizes collective organization through functional connectivity, spectral decomposition, community detection, and thermodynamic entropy measures using a general-purpose graph-analysis framework

Scale-Bridging Insights

The shared data model enables explicit connections between scales through several pathways

** Selectivity testing can filter the population before dimensionality reduction, allowing direct comparison of full-population versus selective-population manifolds **

** Embedding coordinates become features for reverse selectivity analysis, where embedding coordinates are treated as continuous features and tested for significance using the same MI-based significance testing used for behavioral features **

** Pairwise MI testing between neurons yields functional connectivity graphs, which align with selectivity types in the resulting network structure **

Validation and Findings

The framework was validated on synthetic data with known ground truth, where it recovered planted selectivity structure, circular manifold geometry, and modular network organization On hippocampal calcium imaging from 13 mice, DRIADA detected widespread selectivity across all 11 behavioral features and characterized the broader selectivity landscape of CA1 Scale-bridging analysis showed that selectivity-based filtering transforms a collapsed one-dimensional embedding into a two-dimensional spatial manifold, with the filtered embedding spanning both arena dimensions Furthermore, reverse selectivity analysis indicated that ∼57% of neurons informative about leading manifold dimensions were not selective to any of the 11 measured behavioral features

Benchmark Performance

On a simulated toroidal continuous attractor network, four modules—selectivity testing, autoencoder embedding, representational similarity analysis, and recurrence networks—independently converged on the same nonlinear topology The autoencoder with a decorrelation loss achieved a DCI score of 0.96 for disentanglement in the toroidal network benchmark

Conclusion

DRIADA demonstrates that selectivity predicts network membership, filtering by tuning improves manifold structure, and temporal dynamics recover the same topology as population geometry This unified framework makes integrated cross-scale analysis accessible to experimentalists working with calcium imaging, spike trains, or simulated networks.

Improvements for AI systems

Bold header: Unified Data Model for Cross-Scale Analysis

DRIADA makes cross-scale analysis routine across calcium imaging, spike trains, and simulated networks by using a shared data model that makes all analysis modules interoperable. This allows for direct transitions between analyses, such as using selectivity to reshape manifold geometry and using embedding coordinates to interrogate single-neuron roles.

Bold header: Automated Variable Type Detection

The system uses an automatic type detection mechanism where "continuous features (e.g., position, speed), discrete features (e.g., object interaction, behavioral state), circular features (e.g., head direction), and multivariate features are classified and preprocessed appropriately. This avoids a common source of errors: applying an inappropriate MI estimator to a misclassified variable."

Bold header: Cross-Scale Structure Recovery

The AI can perform reverse analysis by using embedding coordinates as features, which the paper states is used to show that the spatial population manifold emerges from a mixture of overtly selective and nominally non-selective neurons. This allows researchers to ask, Is single-neuron selectivity reflected in network membership? directly.

Bold header: Multi-Scale Topological Convergence

The system can demonstrate convergence across scales, as shown when four modules—selectivity testing, autoencoder embedding, representational similarity analysis, and recurrence networks—independently converged on the same toroidal structure. This validates that multi-scale analysis provides convergent evidence that recovered structure reflects underlying organization rather than method-dependent artifacts.

Bold header: Network Structure from Temporal Dynamics

The system can construct networks directly from time series using methods like recurrence networks, visibility graphs, and ordinal partition networks, which are shown to recover the same topology as population geometry, demonstrating that temporal dynamics recover the same topology as population geometry.

Abstract

Understanding how individual neurons contribute to population activity during behavior is central to understanding brain functions across biological scales. We developed DRIADA, an open-source Python toolkit that connects analyses of individual neurons, population activity, and neuronal networks to the same behavioral parameters. We evaluated it on synthetic populations with known ground truth, calcium-imaging recordings from dorsal hippocampal CA1 in 13 mice across 39 sessions, and a simulated continuous attractor neural network (CANN) with known toroidal organization. Across most hippocampal sessions, embeddings of spatially selective neurons predicted held-out position better than equally sized random subsets and all recorded neurons. The same workflow related population dimensions back to individual neurons, including cells without detected selectivity for the measured behavioral variables. In the simulated network, complementary analyses captured features consistent with its known organization. Thus, DRIADA provides a common framework for relating neuronal selectivity, population structure, and network activity to behavioral dynamics.