The open-source pipeline setting the wetware benchmark

What the work claims

This is a software paper, and its claims should be weighted as such: the artifact, not any new experimental result, is the contribution. The authors, from the Paris Brain Institute, Sorbonne Université, Inria, the University of Naples Parthenope, and Aix-Marseille University, argue that neuronal avalanche research is held back by "fragmented, lab-specific scripts," and that results are unusually sensitive to arbitrary choices in event detection, thresholding, and binning, so findings are hard to compare across studies.¹

The package provides four things: avalanche detection from multi-channel electrophysiology (EEG, MEG, ECoG) with configurable thresholds and temporal binning; statistical characterization of size and duration distributions; computation of Avalanche Transition Matrices (ATMs), which the authors state is a first for any toolbox; and graph-theoretic network metrics derived from those matrices, wired to the Brain Connectivity Toolbox and NetworkX. The whole pipeline is scikit-learn compatible, so ATM features can be vectorized straight into machine-learning classifiers, and it ships with automated HTML report generation. It is distributed on PyPI.¹

The only empirical content is an illustrative group-level analysis on a public resting-state EEG dataset comparing Alzheimer's disease, frontotemporal dementia, and healthy controls, used to show the workflow end to end. The paper does not claim any new clinical finding; the example demonstrates plumbing, not biology.¹

How it works

Neuronal avalanches are cascading bursts of activity propagating through a neural network, a framework descending from Beggs and Plenz's 2003 recordings in neocortical circuits, which reported approximately scale-free size and duration distributions. That scale-freeness has been interpreted as evidence that brain dynamics sit near criticality, a regime argued to optimize information processing, dynamic range, and computational power.³ The detection step is simple to state and easy to get different answers from: z-score each channel, threshold it, bin time, and call a cascade an avalanche when at least one channel is active. Change the threshold multiplier, the bin width, or the normalization, and the detected avalanche population changes with it.

The paper's central methodological object is the ATM, introduced by Sorrentino and colleagues in eLife in 2021: a matrix whose entry in row i and column j is the probability that region j activates in the time step after region i activated during an avalanche. This turns a cascade description into a directed map of preferential propagation pathways, and it is where the paper's core contribution sits. pyAvalanches implements two variants: the original conditional-probability model, and a weighted-stochastic formulation that folds in the length of regional activation periods and avalanche sizes, taken from the group's own follow-up work.^{2 5}

Downstream, the package treats each ATM as a graph and extracts centrality and strength metrics, then packages everything into per-subject and per-group containers that export directly to pandas. The scikit-learn bridge matters more than it looks: an ATM is, mechanically, a feature vector, so the path from raw recording to classifier input no longer requires the user to write any analysis code at all.¹

Where a skeptic should push

The single most load-bearing assumption is that standardization removes the fragility, rather than merely relocating it into the defaults. The paper is candid that results are sensitive to "arbitrary choices in event detection, thresholding, binning." A shared package does not eliminate that sensitivity; it picks one answer and makes it frictionless. Every user who accepts the defaults inherits the maintainers' choices about what an avalanche is, and the maintainers are one academic group, funded by a French ANR grant and Italian infrastructure programs, with no governance mechanism attached.¹

Second, the demo is explicitly illustrative. It runs on a public dataset with three clinical groups, but the paper reports no inferential statistics, no effect sizes, and no claim that ATM features separate the conditions out of sample. Anyone tempted to read it as a biomarker result should not. Third, the criticality interpretation the framework inherits is itself contested: scale-free fits to finite neural data are notoriously forgiving, and "near criticality" is a hypothesis about function, not a measured quantity. Fourth, the ATM's transition probabilities are estimated from a finite, often sparse count of observed transitions per subject; on short recordings the matrix entries are noisy, and group-averaged heatmaps can look stable while individual-level estimates are not. Demonstrated: a clean, documented, genuinely useful pipeline. Asserted: that the pipeline's outputs are comparable across studies in a way the field should treat as settled.

Whoever sets the threshold sets the benchmark

For organoid intelligence, the non-obvious implication is that the evidence layer just consolidated. Avalanche statistics are the standard currency in which dish-grown neural networks are argued to be brain-like: a culture whose activity shows scale-free cascades is claimed to sit near criticality, hence to process information well. Until now, that claim was produced by lab-specific scripts no outsider could audit line by line. Once a single open-source package with default parameters becomes the obvious way to compute it, the package's defaults become the operational definition of the benchmark. That is genuinely good news for auditability: anyone can read the thresholding code, run it, and dispute it. But it quietly appoints the maintainers as the de facto arbiters of what counts as an avalanche, and the field has no mechanism for reviewing, versioning, or appealing that arbitration as the software evolves.¹

The opportunity is real and specific. Standardization makes it cheap for a small lab to put its organoid recordings through the same pipeline as everyone else, which is exactly what a credible comparative benchmark for wetware platforms needs. It also closes off a soft form of hype: a vendor can no longer publish criticality claims computed with a generously tuned detector, at least not without departing from a community-standard pipeline that reviewers will recognize. A pre-registered benchmark, where detection parameters are fixed before recordings are made and vendors cannot shop among analysis settings, becomes straightforward to run.¹

The threat is the mirror image. Parameter sensitivity means the pipeline can be tuned, and the same scikit-learn bridge that makes the tool accessible also makes it trivial to search parameter settings until the classifier performs. If ATM-derived features feed disease classifiers, BCI-training selectors, or vendor marketing claims, then software version drift becomes evidence drift: a result certified under one default set is not comparable to a result certified under the next point release, and no regulator currently tracks which version of a Python package a claim was computed with. Note also that the standardization is converging on both ends of the wetware stack: a companion pipeline called CASCADE, published in Bioinformatics in August 2026, already provides cross-manufacturer MEA avalanche analysis for in vitro recordings. One analytic yardstick is being laid across human EEG and dish-grown networks alike, which is precisely what makes cross-system claims possible, and precisely where governance should attach: at the analysis layer, not just the acquisition layer.⁴

None of this is an argument against the package. It is an argument that a measurement standard run by a research group is a governance object whether or not anyone calls it one, and that the field's current answer to "who decides what an avalanche is" is "whoever merges the pull request."¹

The bottom line

Established: pyAvalanches exists, is openly documented, and does integrate ATM computation into an end-to-end pipeline where no public tool did before, and the avalanche literature really has been fragmented across bespoke scripts with parameter choices that move results. Hypothesis, not demonstration: that its defaults will be adopted widely enough to function as a field standard, and that ATM-based features will carry predictive signal beyond what the illustrative demo shows. The grid-relevant consequence stands on either branch: the definition of "brain-like activity" for living neural tissue, in dishes and in heads, is migrating into versioned software maintained by a small academic team, and the credible response is not to slow that down but to treat the pipeline like what it is becoming, a piece of measurement infrastructure that deserves pinned versions, disclosed parameters, and pre-registered settings in any claim that leans on it. What would confirm the optimistic reading: independent groups replicating published avalanche findings with the shared pipeline at pinned versions. What would break it: a round-robin exercise showing that two reasonable parameter settings flip the headline conclusion, which the paper's own framing of threshold sensitivity suggests is entirely possible.¹