| name | emergent-topological-brain-organoids |
| description | Apply persistent homology to MEA recordings of spontaneous brain-organoid activity. Detects H1 loops and H2 voids in correlation networks, compares them to rate-preserving null models, and identifies non-redundant loop-carrying cores. |
| metadata | {"arxiv_id":"2607.16517","published":"2026-07-21","authors":"Eve Bodnia, Margaux Basart, Sofie Hai, Lenzie Ford, Nina Miolane, Kenneth S. Kosik, Dirk Bouwmeester, Lincoln D. Carr","tags":["neuroscience","brain-organoid","persistent-homology","topological-data-analysis","MEA","neural-dynamics","correlation-network"]} |
| license | Complete terms in LICENSE.txt |
Emergent Topological Structure in Brain-Organoid Activity
Apply persistent homology (Vietoris–Rips filtration) to microelectrode-array (MEA) recordings of spontaneous activity from human and mouse cortical organoids, to detect loops (H1) and voids (H2) that exceed a rate- and population-preserving null model.
When to Use
- Analyzing spontaneous multi-unit spike recordings from organoids, slices, or dense in-vivo arrays.
- You want to test whether correlation networks carry structured topology beyond firing rate and population bursting.
- You need to identify the non-redundant core of units that carry topological features.
- You want to know how many nodes are required for persistent homology to resolve H1 and H2.
Core Methodology
1. Build the Correlation Network
-
Spike-sort MEA data; use sorted single units as network nodes.
-
Gaussian-smooth spike trains with a 50 ms kernel.
-
For each pair of units (a, b), compute the maximum normalized cross-correlation over a short lag window (0, 10, or 20 ms):
C(a, b) = max_j ( Σ_i a_i b_{i+j} ) / ( sqrt(Σ_i a_i²) * sqrt(Σ_i b²) )
-
Convert to a dissimilarity: d_ij = 1 - C_ij.
2. Vietoris–Rips Filtration and Persistent Homology
- Build a Vietoris–Rips filtration on d_ij using Ripser (or similar).
- Report density-indexed Betti curves β0(ρ), β1(ρ), β2(ρ), where ρ is edge density.
- Integrate Betti curves (e.g., ∫β1(ρ)dρ) for a threshold-free summary.
3. Rate- and Population-Preserving Null Model
- Generate raster-marginals surrogates: preserve every unit’s total spike count and every time bin’s total population activity, while destroying higher-order co-firing.
- Use 10^5 binary swaps per surrogate; create ~100 surrogates.
- Recompute correlations and topology on each surrogate.
- Compare integrated β1 / β2 from data to the surrogate distribution; report empirical rank p-values.
4. Identify Non-Redundant Core
- Compute H1 cocycle representatives for the persistent loops.
- Randomly remove 10% of units and recompute topology (average over 100 trials).
- Targeted removal: remove the 10% of units most frequently appearing in H1 cocycle representatives.
- Compare retained integrated β1 and bottleneck distance between original and post-removal persistence diagrams.
5. Size-Scaling Analysis
- Repeat for networks with varying N (e.g., subsample units).
- H1 becomes resolvable above roughly N ≈ 100 in the organoid datasets.
- H2 emerges significantly only in larger networks (N ≥ 119 in the reported data).
Key Findings
- H1 loops exceeded the null in 14 of 18 organoid datasets (p ≤ 0.05).
- Loops peak at low edge density (ρ ≲ 0.15) and are carried by strongly co-active units, not weak correlations.
- Random 10% removal retained 92.5% of integrated β1; targeted removal retained 72.6% and produced larger bottleneck distances (median ratio 1.48 vs random).
- H2 voids appeared in 6 datasets, all with N ≥ 119, suggesting larger recordings are needed to resolve higher-order structure robustly.
- Loop structure reflects co-firing, not electrode layout; geometry is topologically trivial in the datasets tested.
Practical Workflow
- Load spike-sorted MEA data; create spike trains and quality-control units.
- Gaussian-smooth with 50 ms kernel; compute pairwise maximum cross-correlation at 0–20 ms lags.
- Build correlation distance matrix d = 1 - C.
- Run Ripser on d; obtain persistence diagrams and Betti curves indexed by density.
- Generate 100 raster-marginals surrogates; compute null distribution of integrated β1 and β2.
- Report p-values and z-scores for data vs null.
- (Optional) Compute cocycle representatives and test robustness to random vs targeted node removal.
- (Optional) Subsample units to estimate topology-resolvable node count.
Parameters and Defaults
| Parameter | Default | Notes |
|---|
| Smoothing kernel | 50 ms Gaussian | Matches tens-of-ms synaptic timescale |
| Lag window | 0 ms (also 10, 20 ms) | Zero lag is symmetric for undirected VR |
| Surrogate swaps | 10^5 per surrogate | Sufficient for destroying higher-order co-firing |
| Surrogate count | 100 | For empirical null distribution |
| Removal fraction | 10% | For robustness / core analysis |
| Density range | ρ ∈ [0, 1] | Threshold-free comparison |
Pitfalls
- Small N: networks below ~50 units may not resolve H1 above surrogate noise; report this as an insufficiency rather than absence of structure.
- Smoothing too wide or narrow: a kernel far from the synaptic timescale can suppress or distort co-firing structure.
- Using Erdős–Rényi as null: this only tests structure against random connectivity, not against rate/population effects. Use raster-marginals surrogates.
- Directed vs undirected: undirected VR is symmetric; for firing-order information, use directed flag complexes separately.
- Loop count alone is misleading: a hub deletion can create new loops. Use bottleneck distance to measure topological disruption.
Activation Keywords
- brain organoid topology
- persistent homology neural recordings
- MEA correlation network loops
- H1 H2 neural topology
- rate-preserving null model
- loop-carrying core
- topological data analysis neuroscience
References
- Giusti, Pastalkova, Curto, Itskov (2015). Clique topology reveals intrinsic geometric structure in neural correlations. PNAS.
- Schreiber et al. (2003). A new correlation-based measure of spike timing reliability. Neurocomputing.
- Okun et al. (2012). Diverse coupling of neurons to populations in sensory cortex. Nature.
- Ripser: https://ripser.scikit-tda.org
- Lancaster & Pasca organoid protocols (see paper references).