Two organoids grown by the same protocol can have different network dynamics. Recording systems also differ between laboratories, as do the individual neurons that can be isolated from each preparation. That variability makes it hard to compare development, disease models and computational tasks. The PULSE project takes aim at the analysis problem.
Hunter Schweiger and colleagues posted A foundation model for brain organoid network dynamics uncovers intrinsic computation and convergent disease phenotypes on bioRxiv on 13 October 2026. It is a preprint. PULSE stands for Population Unified Latent Space Embedding.
What went into the model?
The authors assembled 3,200 high-density microelectrode-array recordings totalling 914.3 hours. The data included mouse-derived neural preparations and slices; the paper also evaluates transfer to human organoid lines held out from the initial training. Activity was represented through spike counts and stimulation features. The model uses a Mamba-2 state-space architecture to learn patterns in the population recordings.
A shared 512-dimensional latent state allows experiments with different numbers of recorded units to be compared in one representation. This is a digital analysis tool, running on conventional processors. It is not a new living-neuron computer or a replacement for the electrodes that acquire the data.
Development and genetic conditions
The researchers used the model to compare the development of human cortical organoid lines with different genetic backgrounds. They reported altered maturation in 16p11.2-deletion organoids and a related network-coordination dimension across experiments involving SHANK3 deficiency. The details require care: the individual SHANK3 comparison was not statistically significant on every metric, and different differentiation protocols complicate comparisons between distinct donor backgrounds.
The more general achievement is an analysis method that can read out some common features of network activity from preparations made and measured in different ways. Reproducing that transfer across independent laboratories would be a particularly useful next test.
The virtual inverted pendulum
In a closed-loop control experiment the organoid's recorded activity contributed to control of a simulated inverted pendulum, commonly known as a cart-pole task. The authors analysed where the task-related activity lay within patterns already produced by the neural network. Their results suggest that stimulation and control selected from existing spontaneous dynamics rather than creating an entirely new activity regime.
That interpretation is scientifically interesting because it bears on what a living substrate contributes to a computational task. It is narrower than claiming a biological network learned arbitrary new representations or solved the task independently of its encoding and decoding software.
Why this helps living-neural computing
More consistent description of neural state could help choose stimulation strategies, characterise differences between preparations and detect when the same device has drifted. Better software analysis may also improve the quality of benchmarks: one can distinguish substrate behaviour from changes in the decoder or experimental environment.
Those applications remain possibilities rather than demonstrated commercial features. There has been no independent replication of PULSE's results at the time of this update. For a second October example of a directly measured task, compare the peer-reviewed Braille-classification experiment; for the wider state of the field see the 2026 review.