Dimensionality Reduction for Integrated Activity Data Analysis — a Python framework for cross-scale analysis of neural population activity.
Neuroscience describes brain activity through three complementary paradigms: single-neuron selectivity (what does each cell encode?), population dynamics (what is the collective geometry?), and networks (how are units connected?). These are not competing theories — they are complementary lenses on the same data. The same neuron can be viewed simultaneously as a feature detector, a dimension of a population manifold, and a node in a functional circuit.
Yet in practice, each paradigm has its own tools, its own data formats, its own packages. If you want to ask a question that crosses paradigm boundaries — does selectivity predict network membership? does filtering by tuning improve manifold structure? how do individual neurons contribute to population coding? do temporal dynamics of single neurons reveal population-level organisation? — you have to build a custom pipeline from incompatible parts.
DRIADA provides a shared data model where all three paradigms operate on the same objects. Results from one domain flow directly into the others:
Three bidirectional bridges connect the paradigms:
- Neurons ↔ Populations. Selective neurons filter the population matrix for dimensionality reduction; embedding coordinates feed back into neuron-level importance estimation, revealing which cells drive each manifold dimension.
- Neurons ↔ Networks. Pairwise MI significance testing builds functional connectivity from neural activity; per-neuron recurrence graphs, visibility graphs, and ordinal partition networks yield functional networks grounded in temporal dynamics rather than correlations.
- Populations ↔ Networks. Graph-based dimensionality reduction (Isomap, diffusion maps) returns proximity graphs that inherit the full network toolkit — spectral decomposition, community detection, entropy — without format conversion. RSA provides a complementary route from population representations to network structure.
DRIADA's strength is integrating these scales — crossing them should be a function call, not a weekend of data wrangling.
- 🧠 INTENSE — detect neuron-feature selectivity via mutual information with rigorous two-stage statistical testing, delay optimization, and mixed selectivity disentanglement
- 📊 Dimensionality Reduction — linear (PCA), graph-based (Isomap, diffusion maps, LLE), neighbor-embedding (UMAP, t-SNE), and neural network-based (autoencoders) with manifold quality metrics
- 📐 Dimensionality Estimation — PCA-based, effective rank, k-NN, correlation, and geodesic dimension
- 🔗 Integration — map single-cell selectivity onto population manifolds and embedding components
- 🌐 Network Analysis — general-purpose graph analysis (spectral, entropy, communities) for connectomes, functional networks, or dimensionality reduction proximity graphs
- 🔄 Recurrence Analysis — delay embedding, recurrence plots, RQA, visibility graphs, ordinal partition networks for nonlinear dynamics and population module recovery
- 📏 RSA — representational dissimilarity matrices, cross-region and cross-session comparisons
- 🧪 Synthetic Data — generate populations with known ground truth for validation
- ⚡ Neuron — calcium kinetics optimization (rise/decay fitting), event detection, and signal quality metrics
Interactive notebooks — click a badge to open in Google Colab (no setup required):
| Notebook | Topics | |
|---|---|---|
 |
DRIADA overview | Experiment objects, feature types, quick tour of INTENSE, dimensionality reduction, networks |
|
Neuron analysis | Spike reconstruction, kinetics optimization, quality metrics, surrogates |
|
Selectivity detection (INTENSE) | Mutual information, two-stage testing, optimal delays, mixed selectivity |
|
Population geometry & DR | PCA, UMAP, Isomap, autoencoders, manifold quality metrics, dimensionality estimation |
|
Network analysis | Cell-cell significance, spectral analysis, communities, graph entropy |
|
Advanced capabilities | Embedding selectivity, leave-one-out importance, RSA, RNN analysis |
|
Recurrence analysis | Delay embedding, recurrence plots, RQA, visibility & ordinal graphs, population module recovery |
All notebooks generate synthetic data internally — no external files needed. See the examples reference for 25+ standalone scripts covering additional use cases.
DRIADA is designed for calcium imaging data but works with any neural activity represented as a (n_units, n_frames) array — RNN activations, firing rates, LFP channels, or anything else. Behavioral variables are 1D or multi-component arrays of the same length. Variable types (continuous, discrete, circular, multivariate) are auto-detected and preprocessed appropriately.
Input workflow: load your arrays into a Python dict and call load_exp_from_aligned_data:
from driada.experiment import load_exp_from_aligned_data
from driada.intense import (compute_cell_feat_significance,
compute_cell_cell_significance, compute_embedding_selectivity)
from driada.rsa import compute_experiment_rdm
from driada.network import Network
from driada.integration import get_functional_organization
import scipy.sparse as sp
data = {
'calcium': calcium_array, # (n_neurons, n_frames) — or 'activations', 'rates', etc.
'speed': speed_array, # (n_frames,) continuous variable
'position': position_array, # (2, n_frames) multi-component variable
'head_direction': hd_array, # auto-detected as circular
'trial_type': trial_type_array, # auto-detected as discrete
}
exp = load_exp_from_aligned_data('MyLab', {'name': 'session1'}, data,
static_features={'fps': 30.0})
feat_stats, feat_sig, *_ = compute_cell_feat_significance(exp) # neuron-feature selectivity
cell_sim, cell_sig, *_ = compute_cell_cell_significance(exp) # functional connectivity
net = Network(adj=sp.csr_matrix(cell_sig), preprocessing='giant_cc') # network analysis
embedding = exp.create_embedding('umap', n_components=2) # dimensionality reduction
rdm, labels = compute_experiment_rdm(exp, items='trial_type') # representational similarity
emb_res = compute_embedding_selectivity(exp, ['umap']) # which neurons encode which components
org = get_functional_organization(exp, 'umap',
intense_results=emb_res['umap']['intense_results']) # selectivity-to-embedding bridgeYou can also load directly from .npz files via load_experiment(). See the RNN analysis tutorial for a non-calcium example.
pip install driada
# Optional extras
pip install driada[gpu] # autoencoders, torch-based methods
pip install driada[mvu] # MVU dimensionality reduction (cvxpy)
pip install driada[all] # everything
# From source
git clone https://github.com/iabs-neuro/driada.git
cd driada
pip install -e ".[dev]"DRIADA pulls in ~30 dependencies (numpy, scipy, scikit-learn, numba, joblib, etc.). See pyproject.toml for the full list.
📖 driada.readthedocs.io — API reference, installation guide, and quickstart 📋 Changelog — release history and migration notes
We welcome contributions! Please see CONTRIBUTING.md for guidelines.
git clone https://github.com/iabs-neuro/driada.git
cd driada
pip install -e ".[dev]"
pytestIf you use DRIADA in your research, please cite:
@software{driada2026,
title = {DRIADA: Dimensionality Reduction for Integrated Activity Data Analysis},
author = {Pospelov, Nikita and contributors},
year = {2026},
url = {https://github.com/iabs-neuro/driada}
}DRIADA has been used in the following research:
- Sotskov et al. (2022) — Fast tuning dynamics of hippocampal place cells during free exploration
- Pospelov et al. (2024) — Effective dimensionality of hippocampal population activity correlates with behavior
- Bobyleva et al. (2025) — Multifractality of structural connectome eigenmodes
- Kononov et al. (2025) — Hybrid computational dynamics in RNNs through reinforcement learning
- Pospelov et al. (2021) — Laplacian Eigenmaps for fMRI resting-state analysis
See PUBLICATIONS.md for the complete list with abstracts and details.
- 📧 Email: pospelov.na14@physics.msu.ru
- 🐛 Issues: GitHub Issues
- 💬 Discussions: GitHub Discussions
This project is licensed under the MIT License — see the LICENSE file for details.