ZAPBench


The Zebrafish Activity Prediction Benchmark measures progress on the problem of predicting cellular-resolution neural activity throughout an entire vertebrate brain.


ICLR Paper GitHub

What is ZAPBench?

Predicting future behavior is a fundamental test towards understanding across natural sciences. How accurately can whole-brain neural activity be predicted from past activity? Larval zebrafish offer a unique opportunity to address this question, as they are the only vertebrate species in which whole-brain activity can be recorded at cellular resolution. For the Zebrafish Activity Prediction Benchmark (ZAPBench) we collected and extensively processed a novel 4d light-sheet microscopy recording of over 70,000 neurons, on which we propose a forecasting benchmark with the aim of catalyzing the development of increasingly accurate models of brain activity.

All details are in our upcoming ICLR 2025 paper.

Browse Datasets

You can interactively explore datasets by clicking the cards below. Further information on datasets.

Data usage may be significant. For optimal viewing, we recommend using a desktop computer.

Activity

Volumetric Activity

Whole-brain calcium activity after elastic alignment, acquired by light-sheet imaging.
Segmentation

Segmentation

Segmentation masks for ~70,000 neurons obtained by training one-shot flood-filling networks.
Traces

Traces

Extracted activity traces per segmented neuron, sorted by similarity.
Rendering

Cell Activity

Rendering of activity traces to visualize spatial relationships.

Interactive Results ✨

Test set performance of initial baselines, see manuscript for details.

Loading charts ...

🔍Click any circle to open an interactive rendering of predictions in a new tab.
This condition is entirely held-out from training. TiDE relies on stimulus covariates, which are out-of-distribution for this condition.

Code & Tutorials

To make it easy to reproduce and extend ZAPBench, we provide full code on GitHub, as well as tutorial-style colabs:

Datasets

Overview of various datasets we released and how to access them.

Training and evaluation

How to train and evaluate forecasting methods on ZAPBench in a framework agnostic way.

Metrics

Explains how to load predictions made by the methods reported in the paper for additional analyses, e.g., to compute custom metrics.

Interactive time-series forecasting

Shows how to run a jax time-series forecasting model interactively.

FAQ

Click to open answers, or expand all.

For general questions, feature requests, or broader conversations about ZAPBench, you can use our discussion page. If you encounter specific problems with our code, please report them by opening an issue. To contact us directly, you can reach out via email.

We'll be sharing key updates, e.g., related to the forthcoming connectome, through our announcement list via email. To subscribe, simply send a blank message to zapbench-announce+subscribe@googlegroups.com, then reply to the join request message you receive to confirm.

This work was done in collaboration with colleagues across Google Research, HHMI Janelia, and Harvard University, including Jan-Matthis Lueckmann, Alexander Immer, Alex Bo-Yuan Chen, Peter H. Li, Mariela D. Petkova, Nirmala A. Iyer, Luuk Willem Hesselink, Aparna Dev, Gudrun Ihrke, Woohyun Park, Alyson Petruncio, Aubrey Weigel, Wyatt Korff, Florian Engert, Jeff W. Lichtman, Misha B. Ahrens, Michał Januszewski, and Viren Jain. We are grateful to the many individuals, listed in the manuscript, who provided valuable feedback and support.