Using NWB Data

last updated: February 9, 2021
In this tutorial, we demonstrate the reading and usage of the NWB file produced in the File Conversion Tutorial. The output is a near-reproduction of Figure 1e from the Li et al publication, showing raster and peristimulus time histogram (PSTH) plots for neural recordings from anterior lateral motor cortex (ALM). This figure illustrates the main finding of the publication, showing the robustness of motor planning behavior and neural dynamics following short unilateral network silencing via optogenetic inhibition.

Reading NWB Files

NWB files can be read in using the nwbRead() function. This function returns a nwbfile object which is the in-memory representation of the NWB file structure.
nwb = nwbRead('out\ANM255201_20141124.nwb');

Constrained Sets

Analyzed data in NWB is placed under the analysis property, which is a Constrained Set. A constrained set consists of an arbitrary amount of key-value pairs similar to Map containers in MATLAB or a dictionary in Python. However, constrained sets also have the ability to validate their own properties closer to how a typed Object would.
You can get/set values in constrained sets using their respective .get()/.set() methods and retrieve all Set properties using the keys() method, like in a containers.Map.
unit_names = keys(nwb.analysis);

Dynamic Tables

nwb.intervals_trials returns a unique type of table called a Dynamic Table. Dynamic tables inherit from the NWB type types.hdmf_common.DynamicTable and allow for a table-like interface in NWB. In the case below, we grab the special column start_time. Dynamic Tables allow adding your own vectors using the vectordata property, which are Constrained Sets. All columns are represented by either a types.hdmf_common.VectorData or a types.hdmf_common.VectorIndex type.

Data Stubs

The data property of the column id in nwb.units is a types.untyped.DataStub. This object is a representation of a dataset that is not loaded in memory, and is what allows MatNWB to lazily load its file data. To load the data into memory, use the .load() method which extracts all data from the NWB file. Alternatively, you can index into the DataStub directly using conventional MATLAB syntax.

Jagged Arrays in Dynamic Tables

With the new addition of addRow and getRow to Dynamic Tables, the concept of jagged arrays can be worked around and no longer require full understanding outside of specific data format concerns or low-level nwb tool development. The below paragraph is retained in its entirety from its original form as purely informational.
All data in a Dynamic Table must be aligned by row and column, but not all data fits into this paradigm neatly. In order to represent variable amounts of data that is localised to each row and column, NWB uses a concept called Jagged Arrays. These arrays consist of two column types: the familiar types.core.VectorData, and the new types.core.VectorIndex. A Vector Index holds no data, instead holding a reference to another Vector Data and a vector of indices that align to the Dynamic Table dimensions. The indices represent the last index boundary in the Vector Data object for the Vector Index row. As an example, an index of three in the first row of the Vector Index column points to the first three values in the referenced Vector Data column. Subsequently, if the next index were a five, it would indicate the fourth and fifth elements in the referenced Vector Data column.
The jagged arrays serve to represent multiple trials and spike times associated to each unit by id. A convenient way to represent these in MATLAB is to use Map containers where each unit's data is indexed directly by its unit id. Below, we utilize getRow in order to build the same Map.
unit_ids =; % array of unit ids represented within this
% Initialize trials & times Map containers indexed by unit_ids
unit_trials = containers.Map('KeyType',class(unit_ids),'ValueType','any');
unit_times = containers.Map('KeyType',class(unit_ids),'ValueType','any');
last_idx = 0;
for i = 1:length(unit_ids)
unit_id = unit_ids(i);
row = nwb.units.getRow(unit_id, 'useId', true, 'columns', {'spike_times', 'trials'});
unit_trials(unit_id) = row.trials{1};
unit_times(unit_id) = row.spike_times{1};

Process Units

We now do the following for each Unit:
sorted_ids = sort(unit_ids);
Photostim = struct(...
'ind', true,... % mask into xs and ys for this photostim
'period', 'none',...
'duration', 0,... % in seconds
'ramp_offset', 0); % in seconds
% Initialize Map container of plotting data for each unit, stored as structure
Unit = containers.Map('KeyType',class(unit_ids),'ValueType','any');
unit_struct = struct(...
'id', [],...
'xs', [],...
'ys', [],...
'xlim', [-Inf Inf],...
'sample', 0,...
'delay', 0,...
'response', 0,...
'left_scatter', false,...
'right_scatter', false,...
'photostim', Photostim); % can have multiple photostim
for unit_id = unit_ids'
We first extract trial IDs from the Unit IDs.
unit_trial_id = unit_trials(unit_id);
Then filter out outliers from the Sample, Delay, and Response time points with which we derive a "good enough" estimate.
trial = nwb.intervals_trials.getRow(unit_trial_id, 'useId', true,...
'columns', {'PoleInTime', 'PoleOutTime', 'CueTime', 'GoodTrials'});
unit_sample = trial.PoleInTime;
unit_delay = trial.PoleOutTime;
unit_response = trial.CueTime;
unit_good_trials = trial.GoodTrials;
% Subjective parameters
delay_threshold = 0.064;
response_threshold = 0.43;
expected_delay_offset = 1.3; % determined from figure 1a
expected_response_offset = 1.3;
expected_delay = unit_sample + expected_delay_offset;
expected_response = unit_delay + expected_response_offset;
good_delay = (unit_delay > expected_delay - delay_threshold) &...
(unit_delay < expected_delay + delay_threshold);
good_response = (unit_response > expected_response - response_threshold) &...
(unit_response < expected_response + response_threshold);
avg_sample = mean(unit_sample(good_delay & good_response));
avg_delay = mean(unit_delay(good_delay & good_response));
avg_response = mean(unit_response(good_delay & good_response));
Filter the rest of the data by "good" trials.
unit_good_trials = unit_good_trials & good_delay & good_response;
unit_trial_id = unit_trial_id(unit_good_trials);
unit_spike_time = unit_times(unit_id);
unit_spike_time = unit_spike_time(unit_good_trials);
Retrieve good trial data and organize by stimulation type.
trial = nwb.intervals_trials.getRow(unit_trial_id, 'useId', true,...
'columns', {'start_time', 'HitR', 'HitL', 'StimTrials', 'PhotostimulationType'});
unit_is_photostim = logical(trial.StimTrials);
unit_stim_type = trial.PhotostimulationType;
unit_no_stim = ~unit_is_photostim & 0 == unit_stim_type;
unit_sample_stim = unit_is_photostim & 1 == unit_stim_type;
unit_early_stim = unit_is_photostim & 2 == unit_stim_type;
unit_middle_stim = unit_is_photostim & 3 == unit_stim_type;
Compose Scatter Plots and the Peristimulus Time Histogram zeroed on the Response time.
xs = unit_spike_time - trial.start_time - avg_response;
ys = unit_trial_id;
curr_unit = unit_struct;
curr_unit.xs = xs;
curr_unit.ys = ys;
curr_unit.left_scatter = logical(trial.HitL);
curr_unit.right_scatter = logical(trial.HitR);
curr_unit.sample = avg_sample - avg_response;
curr_unit.delay = avg_delay - avg_response;
curr_unit.response = 0;
% Photostim periods
curr_unit.photostim.ind = unit_no_stim;
% Sample
if any(unit_sample_stim)
SampleStim = Photostim;
SampleStim.ind = unit_sample_stim;
SampleStim.period = 'Sample';
SampleStim.duration = 0.5;
SampleStim.ramp_offset = 0.1;
curr_unit.photostim(end+1) = SampleStim;
% Early Delay
if any(unit_early_stim)
early_stim_types = unique(unit_stim_type(unit_early_stim));
for i_early_types=1:length(early_stim_types)
early_type = early_stim_types(i_early_types);
EarlyStim = Photostim;
EarlyStim.period = 'Early Delay';
EarlyStim.ind = early_type == unit_stim_type & unit_early_stim;
if early_type == 2
EarlyStim.duration = 0.5;
EarlyStim.ramp_offset = 0.1;
EarlyStim.duration = 0.8;
EarlyStim.ramp_offset = 0.2;
curr_unit.photostim(end+1) = EarlyStim;
% Middle Delay
if any(unit_middle_stim)
MiddleStim = Photostim;
MiddleStim.ind = unit_middle_stim;
MiddleStim.period = 'Middle Delay';
MiddleStim.duration = 0.5;
MiddleStim.ramp_offset = 0.1;
curr_unit.photostim(end+1) = MiddleStim;
Unit(unit_id) = curr_unit;

Plot Example Neurons

neuron_labels = [2, 3]; % neuron labels from Figure 1e
neuron_ids = [11, 2]; % neuron unit IDs corresponding to the Fig 1e labels
num_conditions = 4; % photostim conditions: nostim, sample, early, middle if applicable
num_neurons = length(neuron_ids);
% Inititalize data structures for each summary plot of categorized neural spike data at specified stimulus condition
RasterPlot = struct(...
'xs', 0,...
'ys', 0);
ConditionPlot = struct(...
'label', '',...
'xlim', 0,...
'sample', 0,...
'delay', 0,...
'response', 0,...
'right_scatter', RasterPlot,...
'left_scatter', RasterPlot,...
'psth_bin_window', 0,...
'stim_type', '');
fig = figure;
% Plot neural spike data for each neuron and stimulus condition in a subplot array: num_neurons (rows) x num_conditions (columns)
for nn=1:num_neurons
Neuron = Unit(neuron_ids(nn));
% Initialize structure with neural + stimulus condition data
CurrPlot = ConditionPlot;
CurrPlot.xlim = [min(Neuron.xs) max(Neuron.xs)];
CurrPlot.sample = Neuron.sample;
CurrPlot.delay = Neuron.delay;
CurrPlot.response = Neuron.response;
% Plot each neuron/condition
plot_row = (nn - 1) * num_conditions;
for cc=1:num_conditions
ax = subplot(num_neurons, num_conditions, plot_row + cc, 'Parent', fig);
Stim = Neuron.photostim(cc);
CurrPlot.stim_type = Stim.period;
if strcmp(Stim.period, 'none')
CurrPlot.label = sprintf('Neuron %d', neuron_labels(nn));
CurrPlot.psth_bin_window = 9;
CurrPlot.label = Stim.period;
CurrPlot.psth_bin_window = 2;
stim_left_scatter_ind = Stim.ind & Neuron.left_scatter;
stim_left_scatter_trials = Neuron.ys(stim_left_scatter_ind);
CurrPlot.left_scatter.xs = Neuron.xs(stim_left_scatter_ind);
[~,CurrPlot.left_scatter.ys] = ismember(stim_left_scatter_trials,unique(stim_left_scatter_trials));
stim_right_scatter_ind = Stim.ind & Neuron.right_scatter;
stim_right_scatter_trials = Neuron.ys(stim_right_scatter_ind);
CurrPlot.right_scatter.xs = Neuron.xs(stim_right_scatter_ind);
[~,CurrPlot.right_scatter.ys] = ismember(stim_right_scatter_trials,unique(stim_right_scatter_trials));
plot_condition(ax, CurrPlot);

Helper Functions

PSTH helper function
function [psth_xs, psth_ys] = calculate_psth(xs, bin_window, bin_width)
[bin_counts, edges] = histcounts(xs, 'BinWidth', bin_width);
psth_xs = edges(1:end-1) + (bin_width / 2);
moving_avg_b = (1/bin_window) * ones(1,bin_window);
psth_ys = filter(moving_avg_b, 1, bin_counts);
Plotter function for each stimulus condition
function plot_condition(ax, ConditionPlot)
left_cdata = [1 0 0]; % red
right_cdata = [0 0 1]; % blue
hist_margin = 50;
scatter_margin = 10;
% Calculate PSTH values
% moving average over 200 ms as per figure 1e
hist_bin_width = 0.2 / ConditionPlot.psth_bin_window;
[left_psth_xs, left_psth_ys] =...
calculate_psth(ConditionPlot.left_scatter.xs, ConditionPlot.psth_bin_window, hist_bin_width);
[right_psth_xs, right_psth_ys] =...
calculate_psth(ConditionPlot.right_scatter.xs, ConditionPlot.psth_bin_window, hist_bin_width);
right_scatter_offset = min(ConditionPlot.right_scatter.ys);
right_scatter_height = max(ConditionPlot.right_scatter.ys) - right_scatter_offset;
left_scatter_offset = min(ConditionPlot.left_scatter.ys);
left_scatter_height = max(ConditionPlot.left_scatter.ys) - left_scatter_offset;
psth_height = max([left_psth_ys right_psth_ys]);
left_y_offset = hist_margin...
+ psth_height...
- left_scatter_offset;
right_y_offset = scatter_margin...
+ left_y_offset...
+ left_scatter_offset...
+ left_scatter_height...
- right_scatter_offset;
subplot_height = right_y_offset...
+ right_scatter_offset...
+ right_scatter_height;
hold(ax, 'on');
plot(ax, left_psth_xs, left_psth_ys, 'Color', left_cdata);
plot(ax, right_psth_xs, right_psth_ys, 'Color', right_cdata);
% Scatter Plot
left_y_offset + ConditionPlot.left_scatter.ys,...
'Marker', '.',...
'CData', left_cdata,...
'SizeData', 1);
right_y_offset + ConditionPlot.right_scatter.ys,...
'Marker', '.',...
'CData', right_cdata,...
'SizeData', 1);
% sample, delay, response lines
line(ax, repmat(ConditionPlot.sample, 1, 2), [0 subplot_height],...
'Color', 'k', 'LineStyle', '--');
line(ax, repmat(ConditionPlot.delay, 1, 2), [0 subplot_height],...
'Color', 'k', 'LineStyle', '--');
line(ax, repmat(ConditionPlot.response, 1, 2), [0 subplot_height],...
'Color', 'k', 'LineStyle', '--');
% blue bar for photoinhibition period
if ~strcmp(ConditionPlot.stim_type, 'none')
stim_height = subplot_height;
stim_width = 0.5; % seconds
% end time relative to 'go' cue as described in the paper.
switch ConditionPlot.stim_type
case 'Sample'
end_offset = 1.6;
case 'Early Delay'
end_offset = 0.8;
case 'Middle Delay'
end_offset = 0.3;
error('Invalid photostim period `%s`', ConditionPlot.stim_type);
stim_offset = ConditionPlot.response - stim_width - end_offset;
patch_vertices = [...
stim_offset, 0;...
stim_offset, stim_height;...
stim_offset+stim_width, stim_height;...
stim_offset+stim_width, 0];
'Faces', 1:4,...
'Vertices', patch_vertices,...
'FaceColor', '#B3D3EC',... % light blue shading
'EdgeColor', 'none',...
'FaceAlpha', 0.8);
title(ax, ConditionPlot.label);
xlabel(ax, 'Time (Seconds)');
ylabel(ax, 'Spikes s^{-1}')
xticks(ax, [-2 0 2]);
yticks(ax, [0 max(10, round(psth_height, -1))]);
% legend(ax, [scatter_left_plot, scatter_right_plot], {'Left Lick', 'Right Lick'},...
% 'location', 'northwestoutside');
ax.TickDir = 'out';
ax.XLim = ConditionPlot.xlim;
ax.YLim = [0 subplot_height];
hold(ax, 'off');