ndx-binned-spikes Extension for NWB
Installation
The extension is already available on PyPI and can be installed using pip. The following command installs the latest version of the extension: Python:
pip install -U ndx-binned-spikesIf you want to install the development version of the extension you can install it directly from the GitHub repository. The following command installs the development version of the extension:
Python:
pip install -U git+https://github.com/catalystneuro/ndx-binned-spikes.gitUsage
The BinnedAlignedSpikes object is designed to store counts of spikes around a set of events (e.g., stimuli or behavioral events such as licks). The events are characterized by their timestamps and a bin data structure is used to store the spike counts around each of the event timestamps. The BinnedAlignedSpikes object keeps a separate count for each of the units (i.e. neurons), in other words, the spikes of the units are counted separately but aligned to the same set of events.
Simple example
The following code illustrates a minimal use of this extension:
import numpy as np
from ndx_binned_spikes import BinnedAlignedSpikes
data = np.array(
[
[ # Data of unit with index 0
[5, 1, 3, 2], # Bin counts around the first timestamp
[6, 3, 4, 3], # Bin counts around the second timestamp
[4, 2, 1, 4], # Bin counts around the third timestamp
],
[ # Data of unit with index 1
[8, 4, 0, 2], # Bin counts around the first timestamp
[3, 3, 4, 2], # Bin counts around the second timestamp
[2, 7, 4, 1], # Bin counts around the third timestamp
],
],
dtype="uint64",
)
event_timestamps = np.array([0.25, 5.0, 12.25]) # The timestamps to which we align the counts
milliseconds_from_event_to_first_bin = -50.0 # The first bin is 50 ms before the event
bin_width_in_milliseconds = 100.0 # Each bin is 100 ms wide
binned_aligned_spikes = BinnedAlignedSpikes(
data=data,
event_timestamps=event_timestamps,
bin_width_in_milliseconds=bin_width_in_milliseconds,
milliseconds_from_event_to_first_bin=milliseconds_from_event_to_first_bin
)
The resulting object is usually added to a processing module in an NWB file. The following code illustrates how to add the BinnedAlignedSpikes object to an NWB file. We fist create a nwbfile, then add the BinnedAlignedSpikes object to a processing module and finally write the nwbfile to disk:
from datetime import datetime
from zoneinfo import ZoneInfo
from pynwb import NWBHDF5IO, NWBFile
session_description = "A session of data where a PSTH structure was produced"
session_start_time = datetime.now(ZoneInfo("Asia/Ulaanbaatar"))
identifier = "a_session_identifier"
nwbfile = NWBFile(
session_description=session_description,
session_start_time=session_start_time,
identifier=identifier,
)
ecephys_processing_module = nwbfile.create_processing_module(
name="ecephys", description="Intermediate data derived from extracellular electrophysiology recordings."
)
ecephys_processing_module.add(binned_aligned_spikes)
with NWBHDF5IO("binned_aligned_spikes.nwb", "w") as io:
io.write(nwbfile)Parameters and data structure
The structure of the bins are characterized with the following parameters:
milliseconds_from_event_to_first_bin: The time in milliseconds from the event to the beginning of the first bin. A negative value indicates that the first bin is before the event whereas a positive value indicates that the first bin is after the event.bin_width_in_milliseconds: The width of each bin in milliseconds.
Note that in the diagram above, the milliseconds_from_event_to_first_bin is negative.
The data argument passed to the BinnedAlignedSpikes stores counts across all the event timestamps for each of the units. The data is a 3D array where the first dimension indexes the units, the second dimension indexes the event timestamps, and the third dimension indexes the bins where the counts are stored. The shape of the data is (number_of_units, number_of_events, number_of_bins).
The event_timestamps argument is used to store the timestamps of the events and should have the same length as the second dimension of data. Note that the event_timestamps should not decrease or in other words the events are expected to be in ascending order in time.
The first dimension of data works almost like a dictionary. That is, you select a specific unit by indexing the first dimension. For example, data[0] would return the data of the first unit. For each of the units, the data is organized with the time on the first axis as this is the convention in the NWB format. As a consequence of this choice the data of each unit is contiguous in memory.
The following diagram illustrates the structure of the data for a concrete example:
Linking to units table
One way to make the information stored in the BinnedAlignedSpikes object more useful for future users is to indicate exactly which units or neurons the first dimension of the data attribute corresponds to. This is optional but recommended as it makes the data more meaningful and easier to interpret. In NWB the units are usually stored in a Units table. To illustrate how to to create this link let's first create a toy Units table:
import numpy as np
from pynwb.misc import Units
num_units = 5
max_spikes_per_unit = 10
units_table = Units(name="units")
units_table.add_column(name="unit_name", description="name of the unit")
rng = np.random.default_rng(seed=0)
times = rng.random(size=(num_units, max_spikes_per_unit)).cumsum(axis=1)
spikes_per_unit = rng.integers(1, max_spikes_per_unit, size=num_units)
spike_times = []
for unit_index in range(num_units):
# Not all units have the same number of spikes
spike_times = times[unit_index, : spikes_per_unit[unit_index]]
unit_name = f"unit_{unit_index}"
units_table.add_unit(spike_times=spike_times, unit_name=unit_name)This will create a Units table with 5 units. We can then link the BinnedAlignedSpikes object to this table by creating a DynamicTableRegion object. This allows to be very specific about which units the data in the BinnedAlignedSpikes object corresponds to. In the following code, the units described on the BinnedAlignedSpikes object correspond to the unit with indices 1 and 3 on the Units table. The rest of the procedure is the same as before:
from ndx_binned_spikes import BinnedAlignedSpikes
from hdmf.common import DynamicTableRegion
# Now we create the BinnedAlignedSpikes object and link it to the units table
data = np.array(
[
[ # Data of the unit 1 in the units table
[5, 1, 3, 2], # Bin counts around the first timestamp
[6, 3, 4, 3], # Bin counts around the second timestamp
[4, 2, 1, 4], # Bin counts around the third timestamp
],
[ # Data of the unit 3 in the units table
[8, 4, 0, 2], # Bin counts around the first timestamp
[3, 3, 4, 2], # Bin counts around the second timestamp
[2, 7, 4, 1], # Bin counts around the third timestamp
],
],
)
region_indices = [1, 3]
units_region = DynamicTableRegion(
data=region_indices, table=units_table, description="region of units table", name="units_region"
)
event_timestamps = np.array([0.25, 5.0, 12.25])
milliseconds_from_event_to_first_bin = -50.0 # The first bin is 50 ms before the event
bin_width_in_milliseconds = 100.0
name = "BinnedAignedSpikesForMyPurpose"
description = "Spike counts that is binned and aligned to events."
binned_aligned_spikes = BinnedAlignedSpikes(
data=data,
event_timestamps=event_timestamps,
bin_width_in_milliseconds=bin_width_in_milliseconds,
milliseconds_from_event_to_first_bin=milliseconds_from_event_to_first_bin,
description=description,
name=name,
units_region=units_region,
)
As with the previous example this can be then added to a processing module in an NWB file and then written to disk using exactly the same code as before.
Storing data from multiple conditions (i.e. multiple stimuli)
BinnedAlignedSpikes can also be used to store data that is aggregated across multiple conditions while at the same time keeping track of which condition each set of counts corresponds to. This is useful when you want to store the spike counts around multiple conditions (e.g., different stimuli, behavioral events, etc.) in a single structure. Since each condition may not occur the same number of times (e.g. different stimuli do not appear in the same frequency), an homogeneous data structure is not possible. Therefore an extra variable, condition_indices, is used to indicate which condition each set of counts corresponds to.
from ndx_binned_spikes import BinnedAlignedSpikes
binned_aligned_spikes = BinnedAlignedSpikes(
bin_width_in_milliseconds=bin_width_in_milliseconds,
milliseconds_from_event_to_first_bin=milliseconds_from_event_to_first_bin,
data=data, # Shape (number_of_units, number_of_events, number_of_bins)
timestamps=timestamps, # Shape (number_of_events,)
condition_indices=condition_indices, # Shape (number_of_events,)
condition_labels=condition_labels, # Shape (number_of_conditions,) or np.unique(condition_indices).size
)Note that number_of_events here represents the total number of repetitions for all the conditions being aggregated. For example, if data is being aggregated from two stimuli where the first stimulus appeared twice and the second appeared three times, the number_of_events would be 5.
The condition_indices is an indicator vector that should be constructed so that data[:, condition_indices == condition_index, :] corresponds to the binned spike counts for the condition with the specified condition_index. You can retrieve the same data using the convenience method binned_aligned_spikes.get_data_for_condition(condition_index).
The condition_labels argument is optional and can be used to store the labels of the conditions. This is meant to help to understand the nature of the conditions
It's important to note that the timestamps must be in ascending order and must correspond positionally to the condition indices and the second dimension of the data. If they are not, a ValueError will be raised. To help organize the data correctly, you can use the convenience method BinnedAlignedSpikes.sort_data_by_event_timestamps(data=data, event_timestamps=event_timestamps, condition_indices=condition_indices), which ensures the data is properly sorted. Here’s how it can be used:
sorted_data, sorted_event_timestamps, sorted_condition_indices = BinnedAlignedSpikes.sort_data_by_event_timestamps(data=data, event_timestamps=event_timestamps, condition_indices=condition_indices)
binned_aligned_spikes = BinnedAlignedSpikes(
bin_width_in_milliseconds=bin_width_in_milliseconds,
milliseconds_from_event_to_first_bin=milliseconds_from_event_to_first_bin,
data=sorted_data,
event_timestamps=sorted_event_timestamps,
condition_indices=sorted_condition_indices,
condition_labels=condition_labels
)The same can be achieved by using the following script:
sorted_indices = np.argsort(event_timestamps)
sorted_data = data[:, sorted_indices, :]
sorted_event_timestamps = event_timestamps[sorted_indices]
sorted_condition_indices = condition_indices[sorted_indices]Example of building an BinnedAlignedSpikes for two conditions
To better understand how this object works, let's consider a specific example. Suppose we have data for two different stimuli and their associated timestamps:
import numpy as np
# Two units and 4 bins
data_for_first_stimuli = np.array(
[
# Unit 1
[
[0, 1, 2, 3], # Bin counts around the first timestamp
[4, 5, 6, 7], # Bin counts around the second timestamp
],
# Unit 2
[
[8, 9, 10, 11], # Bin counts around the first timestamp
[12, 13, 14, 15], # Bin counts around the second timestamp
],
],
)
# Also two units and 4 bins but this condition occurred three times
data_for_second_stimuli = np.array(
[
# Unit 1
[
[0, 1, 2, 3], # Bin counts around the first timestamp
[4, 5, 6, 7], # Bin counts around the second timestamp
[8, 9, 10, 11], # Bin counts around the third timestamp
],
# Unit 2
[
[12, 13, 14, 15], # Bin counts around the first timestamp
[16, 17, 18, 19], # Bin counts around the second timestamp
[20, 21, 22, 23], # Bin counts around the third timestamp
],
]
)
timestamps_first_stimuli = [5.0, 15.0]
timestamps_second_stimuli = [1.0, 10.0, 20.0]The way that we would build the data for the BinnedAlignedSpikes object is as follows:
from ndx_binned_spikes import BinnedAlignedSpikes
bin_width_in_milliseconds = 100.0
milliseconds_from_event_to_first_bin = -50.0
data = np.concatenate([data_for_first_stimuli, data_for_second_stimuli], axis=1)
event_timestamps = np.concatenate([timestamps_first_stimuli, timestamps_second_stimuli])
condition_indices = np.concatenate([np.zeros(2), np.ones(3)])
condition_labels = ["a", "b"]
sorted_data, sorted_event_timestamps, sorted_condition_indices = BinnedAlignedSpikes.sort_data_by_event_timestamps(data=data, event_timestamps=event_timestamps, condition_indices=condition_indices)
binned_aligned_spikes = BinnedAlignedSpikes(
bin_width_in_milliseconds=bin_width_in_milliseconds,
milliseconds_from_event_to_first_bin=milliseconds_from_event_to_first_bin,
data=sorted_data,
event_timestamps=sorted_event_timestamps,
condition_indices=sorted_condition_indices,
)Then we can recover the original data by calling the get_data_for_condition method:
retrieved_data_for_first_stimuli = binned_aligned_spikes.get_data_for_condition(condition_index=0)
np.testing.assert_array_equal(retrieved_data_for_first_stimuli, data_for_first_stimuli)This extension was created using ndx-template.