Google unveils ZAPBench to forecast whole-brain neural activity

ZAPBench is an open-source benchmark and dataset launched by Google Research, HHMI Janelia, and Harvard at ICLR 2025, designed to test how accurately AI models can forecast full-brain neural firing patterns.

Built on whole-brain recordings from larval zebrafish, the benchmark sets up a clean challenge: just as large language models predict the next word, an AI model takes a window of recorded neural activity and predicts what roughly 70,000 individual neurons will do over the subsequent 30 seconds.

How the whole-brain data was captured

Most functional neuroscience benchmarks focus on tiny slices of the brain. The mouse visual cortex benchmark Dynamic Sensorium, for example, covers less than 0.1% of the mouse brain. But real behaviors like navigating river currents or chasing prey require sensation, memory, and motor decisions working across the entire brain at once.

The team chose six-day-old larval zebrafish because they are transparent and already capable of complex sensory and motor behaviors. Using a specialized light-sheet microscope, researchers scanned the fish brain slice by slice, tracking the flashes of genetically encoded GCaMP calcium indicators whenever neurons fired.

Over a continuous two-hour recording session, the fish was placed in a virtual reality rig projecting nine different environmental stimuli, like moving stripes that mimic water currents. Electrodes on the tail recorded muscle contractions as the fish tried to swim against the virtual flow.

The resulting dataset offers two complementary formats. Researchers can work with the raw 3D volumetric light-sheet video, or use a pre-segmented time-series version where individual voxel traces are separated into single-cell activity lines over the full two hours.

Forecasting neural patterns in 3D

Models submitted to ZAPBench are evaluated using mean absolute error (MAE) against ground-truth neural firings over the 30-second target window.

Baseline tests revealed several practical takeaways. Context length matters significantly: models fed longer historical video clips made far fewer forecasting mistakes than those given short context windows.

Interestingly, models trained directly on the raw 3D volumetric video outperformed time-series models in several scenarios. Having direct access to the 3D spatial layout between cells appears to give spatial architectures an edge, even though volumetric files require more compute. The team also observed that prediction errors clustered in specific brain regions, revealing that certain functional circuits are far harder to model than others.

The authors are now working to reconstruct the complete physical connectome for the exact specimen used in the recordings, which will soon add synaptic wiring data to the functional traces.

Why it matters

If you work on spatial-temporal modeling, brain-inspired machine learning, or computational biology, ZAPBench is well worth a look. It removes the painful barrier of processing raw microscopy imagery and hands ML teams a plug-and-play benchmark at an unprecedented vertebrate scale.

What makes this project clever is its LLM-like framing: treating dynamic neural states as predictable tokens across an entire organism. While a six-day-old fish brain is far simpler than mammalian cortex, this dataset finally gives the research community a standardized proving ground to test whether our neural network architectures can truly grasp the dynamics of biological intelligence.