A Brain-wide Neuronal Spiking and Behavior Dataset for Working Memory-Specific Activation and Reactivation in Mice

This dataset provides comprehensive brain-wide single-unit spiking and behavioral data, capturing neural activity from 40 head-fixed mice performing an olfactory delayed paired-association task. The dataset is specifically designed to facilitate the study of working memory-specific activation and spontaneous memory reactivation. A comprehensive data descriptor article is uploaded alongside these data files; viewers are highly encouraged to refer to this accompanying document for further in-depth methodological details, extended analysis parameters, and visual references.Scientific ValueThis dataset provides extensive brain-wide coverage, encompassing 33,028 single units across 62 brain regions, with 34 regions each containing more than 100 neurons. Its distinctive contribution lies in systematically documenting both task-related activity motifs and their spontaneous hierarchical replay during inter-trial intervals — a dimension that has remained largely inaccessible in prior working memory research. The dataset provides novel resources to: Characterize and constrain computational models of multi-second spiking activity organization underlying working memory, such as cell assemblies, attractor networks, synfire chains, and trajectory dynamics; investigate the role of spike-coupled motifs and their replay in working memory maintenance, decision making and long-term circuit-level plasticity; and inform the design of brain-inspired artificial intelligence architectures grounded in hierarchical neural dynamics.Subjects and Task ParadigmThe experimental subjects utilized were 40 male adult C57BL/6J mice, aged 8 to 12 weeks at the beginning of experiments. Mice performed a delayed paired-association task using a custom olfactometry apparatus where a 1-second sample odor of either ethyl propionate (S1) or propyl acetate (S2) was followed by a variable delay period of 3 or 6 seconds. Following the delay, a 1-second test odor consisting of butyl formate (T1) or methyl butyrate (T2) was presented, and mice were trained to lick a water port exclusively during a 1-second response window on correctly paired trials (S1-T1, or S2-T2) to receive a reward.Electrophysiological RecordingRecordings were conducted using the Neuropixels 1.0 silicon probes, which enabled highly dense extracellular sampling across 62 anatomically defined brain regions at a temporal resolution of 30,000 samples per second. Data acquisition was managed using SpikeGLX software alongside a custom embedded controller that provided a temporally encoded synchronization pattern, ensuring sub-millisecond alignment between behavioral states and neural events across multiple simultaneous probes.Spike Sorting RegistrationThe raw neural data underwent fully automated spike sorting using the robust Kilosort 2 algorithm. This sorting algorithm performed continuous drift correction and template tracking to reliably assign individual neural spikes to 33,028 distinct neuronal clusters. Histological images of the brain slices were then carefully registered to the Allen Institute Common Coordinate Framework version 3 using a custom open-source pipeline, mapping each isolated single unit to precise three-dimensional coordinates within the brain.File Formats and VolumeThe published dataset consists of 116 individual data files, presenting a cumulative total data volume of 17,323 Megabytes. Each file corresponds to a single continuous recording session from one subject and is conveniently stored in the standard Neurodata Without Borders (.nwb) format. These comprehensive files encompass sorted spike times, unit quality metrics, hierarchical anatomical labels, behavioral event markers, and pre-computed peri-stimulus time histograms.Data Sample DescriptionEach file in this dataset is hierarchically structured into distinct groups—functioning much like a directory system—that store specific subsets of experimental data using pre-defined naming conventions. This standardized architecture within the NWB format ensures viewers can systematically evaluate the dataset and quickly access necessary content. The data is fundamentally organized into four primary groups: General, Processing, Intervals, and Units.General and Processing GroupsThe General group houses essential metadata regarding the experimental subject, including species, strain, sex, and age, alongside detailed specifications of the Neuropixels recording devices. The Processing group contains pre-computed analysis products, specifically storing trial-aligned neuronal firing rate time series cleanly binned at 1000-millisecond and 250-millisecond temporal resolutions. These pre-computed firing rates allow researchers to rapidly visualize neural dynamics without recalculating metrics from raw spike times.Structured Tabular - IntervalsStructured tabular data forms a critical component of this dataset, prominently featured within the Intervals group which exhaustively documents trial-by-trial behavioral parameters. Allocating one distinct row for every executed trial, key columns record precise start and stop times, sample and test cue identifiers, and specific delay durations. Additionally, it tracks recorded lick responses and provides logical flags indicating correct trial outcomes and well-trained performance windows.Structured Tabular - UnitsCorrespondingly, the Units group features a dynamic tabular structure that catalogs the neuronal data for all isolated single units, utilizing one dedicated row per recorded neuron. This table integrates crucial cluster quality control metrics, mean firing rates, precise spike times, and physical probe channel depths. Furthermore, it embeds hierarchical Allen Common Coordinate Framework anatomical region annotations spanning granularity levels 3 through 8, facilitating rapid neural population filtering.Data Quality ControlStrict inclusion criteria were proactively applied to ensure high-quality behavioral performance data. Analysis of correctly executed tasks is strictly restricted to sessions where mice demonstrably achieved a well-trained performance threshold, defined as a minimum accuracy rate of 75 percent over a moving window of 40 consecutive trials. Furthermore, neural clusters were classified as valid single units only if they exhibited a statistically significant refractory period trough, a stable mean firing rate exceeding 1 Hertz, and an estimated spike contamination rate strictly below 10 percent.Required Software EnvironmentTo reproduce the data variables and process the analytical scripts, users are expected to set up a compatible computational environment. The provided demonstration scripts are written in MATLAB, requiring MATLAB R2023b or later, the MatNWB interface for file input/output (https://github.com/NeurodataWithoutBorders/matnwb), the FieldTrip Toolbox (https://www.fieldtriptoolbox.org), and the Buzcode toolbox (https://buzsakilab.com/wp/resources/buzcode). Alternatively, users working in Python can access the exact same underlying data structures utilizing Python 3.12 or later, combined with the standard PyNWB package. For robust custom downstream analyses, users should filter behavioral trials utilizing the embedded `intervals_trials` table. It is recommended to explicitly restrict correct task performance analyses to trials where both the `well_trained` and `is_correct` logical flags are marked true. The sample cue onset dynamically serves as the primary behavioral alignment reference point (time 0) for all peri-event analyses, and all primary spike times are natively pre-converted to seconds to comply with the NWB core schema.Reproducing Demonstrative PlotsResearchers can easily validate the data using the provided quick-start scripts, which dynamically generate visual representations matching those discussed in the accompanying scientific article, even without prior experience handling NWB formats. Users should start MATLAB, add the required toolboxes to their path, and run the scripts in the `jpsth` directory, which in inturn contains the `binary` directory for the downloaded `.nwb` data files. The repository (https://gitee.com/XiaoxingZhang/hierarchicalreplay_NSB2025) provides three demonstration scripts designed for rapid reproduction:Demo 1 (jpsth/+demo/selective_unit_raster.m): This script demonstrates how to successfully extract raw spike times directly from the data tables and align them to task events. Running this script reproduces the trial structure overlays and single-unit raster plots detailed in the uploaded scientific article, illustrating clear peri-stimulus time histograms.Demo 2 (jpsth/+demo/selective_ratio_per_region.m): This script details the statistical identification of memory-content-selective neurons exhibiting differential firing rates following specific sample odors in the working memory delay period. By accessing the embedded anatomical annotations, the code systematically generates a bar plot of neural distributions across brain regions, effectively recreating the spatial distribution metrics shown in the accompanying article.Demo 3 (jpsth/+demo/spike_coupling.m): This specialized script extracts simultaneous spike times to efficiently detect functionally coupled spike pairs within a 10-millisecond latency window. Utilizing the aforementioned Buzcode toolbox, it generates baseline-corrected cross-correlograms to visualize asymmetric spike coupling, directly reproducing the functional connection analyses referenced in the associated scientific article.