Research map / Time series / Causal and generative time series
ZAPBench: A Benchmark for Whole-Brain Activity Prediction in Zebrafish
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces ZAPBench, a whole-brain larval zebrafish activity forecasting benchmark built on a 4D light-sheet dataset of ~71,721 neurons, evaluating time-series and volumetric video models against naive baselines.
Problem
Predicting future cellular-resolution neural activity across an entire vertebrate brain lacks a standardized benchmark; prior neural forecasting benchmarks covered only a small fraction of neurons.
Approach
A whole-brain larval zebrafish light-sheet dataset with motion alignment and voxel-level segmentation, paired with a forecasting benchmark that compares time-series models (Linear, TiDE, TSMixer, Time-Mix) and a volumetric U-Net against mean and stimulus baselines, measured by MAE over a 32-step horizon.
Key results
- Models largely outperform naive mean and stimulus baselines in grand average MAE across conditions.
- The U-Net volumetric video model ranks best for short context (C=4), with the largest advantage at single-step predictions.
- Stimulus baseline is competitive: fewer than half of model runs (53 of 120) beat it at 32 steps ahead for short context.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use forecasting models instead of naive baselines | ZAPBench forecasting models · ZAPBench test conditions (excluding taxis) | MAE | Helped |
| Use the stimulus baseline instead of other models | ZAPBench short-context forecasting · ZAPBench test conditions, short context, 32 steps ahead | MAE: 53 out of 120 runs | Mixed |
| Provide long context rather than short context | ZAPBench forecasting models · ZAPBench test conditions; predictions further ahead | MAE | Helped |
| Provide long context rather than short context | U-Net · ZAPBench test conditions; one step ahead | MAE | Hurt |
| Use Linear instead of TiDE, TSMixer, Time-Mix, or U-Net | ZAPBench long-context forecasting · ZAPBench test conditions, long context | MAE | Hurt |
| Use U-Net instead of the time-series models | ZAPBench forecasting · ZAPBench test conditions, short context | MAE | Helped |
| Add static covariates containing neuron spatial positions to TiDE (ablation) | TiDE · ZAPBench | MAE | Hurt |
| Add past stimulus covariates to TiDE (ablation) | TiDE · ZAPBench | MAE | Hurt |
| Add feature-mixing modules to Time-Mix (ablation) | Time-series forecasting models · ZAPBench | MAE | No clear effect |
| Add NLinear-style normalization to Linear (ablation) | Linear · ZAPBench | MAE | No clear effect |
Surprising findings
- The stimulus-evoked lookup baseline is competitive with many trained forecasting models in short-context settings.
- TiDE performed best without static spatial covariates, suggesting cross-neuron information is underutilized.
Limitations
- Single specimen and limited behavioral assays; models may not generalize to other fish.
- Low volume scan rate (~1 Hz) and GCaMP calcium indicator act as low-pass filter; structured stripe noise in signal.
- No proofreading of segmentation; split/merge errors expected and EM connectome not yet available.
Experimental setup
Datasets: ZAPBench light-sheet whole-brain activity dataset (71,721 neurons, 7,879 steps)
Benchmarks: ZAPBench
Metrics: mean absolute error (MAE)
Baselines: Mean baseline Stimulus baseline Linear TiDE TSMixer Time-Mix
Builds on: TiDE TSMixer U-Net Flood Filling Networks SOFIMA
Closest papers by meaning
- Discovering Mechanistic Models of Neural Activity: System Identification in an in Silico Zebrafish (NeurIPS 2026 · similarity 0.83)
- Retrospective for the Dynamic Sensorium Competition for predicting large-scale mouse primary visual cortex activity from videos (NeurIPS 2024 · similarity 0.82)
- POCO: Scalable Neural Forecasting through Population Conditioning (NeurIPS 2025 · similarity 0.82)
- Intrinsic Goals for Autonomous Agents: Model-Based Exploration in Virtual Zebrafish Predicts Ethological Behavior and Whole-Brain Dynamics (NeurIPS 2025 · similarity 0.81)
- Towards a "Universal Translator" for Neural Dynamics at Single-Cell, Single-Spike Resolution (NeurIPS 2024 · similarity 0.80)
- Discovering plasticity rules that organize and maintain neural circuits (NeurIPS 2024 · similarity 0.80)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.