def filter_probes(pacall, probes, threshold=0.5):
"""
Performs intensity based filtering (IBF) of expression probes
Uses binary indicator for expression levels in `pacall` to determine which
probes have expression levels above background noise in `threshold` of
samples across donors.
Parameters
----------
pacall : list
List of PACall files from Allen Brain Institute. Optimally obtained by
calling `abagen.fetch_microarray()` and accessing the `pacall`
attribute on the resulting object
probes : pandas.DataFrame
Dataframe containing information on genetic probes that should be
considered in filtering
threshold : (0, 1) float, optional
Threshold for filtering probes. Specifies the proportion of samples for
which a given probe must have expression levels above background noise.
Default: 0.5
Returns
-------
filtered : pandas.DataFrame
Dataframe containing information on probes that should be retained
according to intensity-based filtering
"""
probes = io.read_probes(probes)
signal, samples = [], 0
for fname in pacall:
data = io.read_pacall(fname).loc[probes.index]
samples += data.shape[-1]
# sum binary expression indicator across samples for current subject
signal.append(data.sum(axis=1).values)
# calculate proportion of signal to noise for given probe across samples
keep = (np.sum(signal, axis=0) / samples) > threshold
# drop "bad" probes
filtered = probes[keep]
return filtered
def reannotate_probes(probes):
"""
Replaces gene symbols in `probes` with reannotated data
Uses annotations from [1]_ to replace probe annotations shipped with AHBA
data. Any probes that were unable to be matched to a gene in reannotation
procedure are not retained.
Parameters
----------
probes : str or pandas.DataFrame
Probe file or loaded probe dataframe from Allen Brain Institute.
Optimally obtained by calling `abagen.fetch_microarray()` and accessing
the `probes` attribute on the resulting object
Returns
-------
reannotated : pandas.DataFrame
References
----------
.. [1] Arnatkeviciute, A., Fulcher, B. D., & Fornito, A. (2018). A
practical guide to linking brain-wide gene expression and neuroimaging
data. bioRxiv, 380089.
"""
# load in reannotated probes
reannot = resource_filename('abagen', 'data/reannotated.csv.gz')
with gzip.open(reannot, 'r') as src:
reannot = pd.read_csv(StringIO(src.read().decode('utf-8'))).iloc[:-1]
reannot = reannot[['probe_name', 'gene_symbol', 'entrez_id']]
# merge reannotated with original, keeping only reannotated
probes = io.read_probes(probes).reset_index()[['probe_name', 'probe_id']]
merged = pd.merge(reannot, probes, on='probe_name', how='left')
# reset index as probe_id and sort
reannotated = merged.set_index('probe_id').sort_index()
return reannotated
def get_expression_data(atlas,
atlas_info=None,
*,
exact=True,
tolerance=2,
metric='mean',
ibf_threshold=0.5,
corrected_mni=True,
reannotated=True,
return_counts=False,
return_donors=False,
donors='all',
data_dir=None):
"""
Assigns microarray expression data to ROIs defined in `atlas`
This function aims to provide a workflow for generating pre-processed,
microarray expression data for abitrary `atlas` designations. First, some
basic filtering of genetic probes is performed, including:
1. Intensity-based filtering of microarray probes to remove probes that
do not exceed a certain level of background noise (specified via the
`ibf_threshold` parameter), and
2. Selection of a single, representative probe for each gene via a
differential stability metric, wherein the probe that has the most
consistent regional variation across donors is retained.
Tissue samples are then matched to parcels in the defined `atlas` for each
donor. If `atlas_info` is provided then this matching is constrained by
both hemisphere and tissue class designation (e.g., cortical samples from
the left hemisphere are only matched to ROIs in the left cortex,
subcortical samples from the right hemisphere are only matched to ROIs in
the left subcortex); see the `atlas_info` parameter description for more
information.
Matching of microarray samples to parcels in `atlas` is done via a multi-
step process:
1. Determine if the sample falls directly within a parcel,
2. Check to see if there are nearby parcels by slowly expanding the
search space to include nearby voxels, up to a specified distance
(specified via the `tolerance` parameter),
3. If there are multiple nearby parcels, the sample is assigned to the
closest parcel, as determined by the parcel centroid.
If at any step a sample can be assigned to a parcel the matching process is
terminated. If multiple sample are assigned to the same parcel they are
aggregated with the metric specified via the `metric` parameter. More
control over the sample matching can be obtained by setting the `exact`
parameter; see the parameter description for more information.
Once all samples have been matched to parcels for all supplied donors, the
microarray expression data are normalized within-donor via a scaled robust
sigmoid (SRS) procedure before being combined across donors via the
supplied `metric`.
Parameters
----------
atlas : niimg-like object
A parcellation image in MNI space, where each parcel is identified by a
unique integer ID
atlas_info : str or :class:`pandas.DataFrame`, optional
Filepath to or pre-loaded dataframe containing information about
`atlas`. Must have at least columns 'id', 'hemisphere', and 'structure'
containing information mapping atlas IDs to hemisphere (i.e, "L", "R")
and broad structural class (i.e., "cortex", "subcortex", "cerebellum").
Default: None
exact : bool, optional
Whether to use exact matching of donor tissue samples to parcels in
`atlas`. If True, this function will match tissue samples to parcels
within `threshold` mm of the sample; any samples that are beyond
`threshold` mm of a parcel will be discarded. This may result in some
parcels having no assigned sample / expression data. If False, the
default matching procedure will be performed and followed by a check
for parcels with no assigned samples; any such parcels will be matched
to the nearest sample (nearest defined as the sample with the closest
Euclidean distance to the parcel centroid). Default: True
tolerance : int, optional
Distance (in mm) that a sample must be from a parcel for it to be
matched to that parcel. This is only considered if the sample is not
directly within a parcel. Default: 2
metric : str or func, optional
Mechanism by which to collapse across donors, if input `files` provides
multiple donor datasets. If a str, should be in ['mean', 'median']; if
a function, should be able to accept an `N`-dimensional input and the
`axis` keyword argument and return an `N-1`-dimensional output.
Default: 'mean'
ibf_threshold : [0, 1] float, optional
Threshold for intensity-based filtering specifying. This number should
specify the ratio of samples, across all supplied donors, for which a
probe must have signal above background noise in order to be retained.
Default: 0.5
corrected_mni : bool, optional
Whether to use the "corrected" MNI coordinates shipped with the
`alleninf` package instead of the coordinates provided with the AHBA
data when matching tissue samples to anatomical regions. Default: True
reannotated : bool, optional
Whether to use reannotated probe information provided by [1]_ instead
of the default probe information from the AHBA dataset. Using
reannotated information will discard probes that could not be reliably
matched to genes. Default: True
return_counts : bool, optional
Whether to return how many samples were assigned to each parcel in
`atlas` for each donor. Default: False
return_donors : bool, optional
Whether to return donor-level expression arrays instead of aggregating
expression across donors with provided `metric`. Default: False
donors : list, optional
List of donors to use as sources of expression data. Can be either
donor numbers or UID. If not specified will use all available donors.
Default: 'all'
data_dir : str, optional
Directory where expression data should be downloaded (if it does not
already exist) / loaded. If not specified will use the current
directory. Default: None
Returns
-------
expression : (R, G) :class:`pandas.DataFrame`
Microarray expression for `R` regions in `atlas` for `G` genes,
aggregated across donors, where the index corresponds to the unique
integer IDs of `atlas` and the columns are gene names.
counts : (R, D) :class:`pandas.DataFrame`
Number of samples assigned to each of `R` regions in `atlas` for each
of `D` donors (if multiple donors were specified); only returned if
`return_counts=True`.
References
----------
.. [1] Arnatkevic̆iūtė, A., Fulcher, B. D., & Fornito, A. (2019). A
practical guide to linking brain-wide gene expression and neuroimaging
data. NeuroImage, 189, 353-367.
.. [2] Hawrylycz, M.J. et al. (2012) An anatomically comprehensive atlas of
the adult human transcriptome. Nature, 489, 391-399.
"""
# fetch files
files = datasets.fetch_microarray(data_dir=data_dir, donors=donors)
for key in ['microarray', 'probes', 'annotation', 'pacall', 'ontology']:
if key not in files:
raise KeyError('Provided `files` dictionary is missing {}. '
'Please check inputs.'.format(key))
# load atlas_info, if provided
atlas = check_niimg_3d(atlas)
if atlas_info is not None:
atlas_info = utils.check_atlas_info(atlas, atlas_info)
# get combination functions
metric = utils.check_metric(metric)
# get some info on the number of subjects, labels in `atlas_img`
num_subj = len(files.microarray)
all_labels = utils.get_unique_labels(atlas)
if not exact:
centroids = utils.get_centroids(atlas, labels=all_labels)
# reannotate probes based on updates from Arnatkeviciute et al., 2018 then
# perform intensity-based filter of probes and select probe with highest
# differential stability for each gene amongst remaining probes
if reannotated:
probes = process.reannotate_probes(files.probes[0])
else:
probes = io.read_probes(files.probes[0])
probes = process.filter_probes(files.pacall,
probes,
threshold=ibf_threshold)
probes = process.get_stable_probes(files.microarray, files.annotation,
probes)
expression, missing = [], []
counts = pd.DataFrame(np.zeros((len(all_labels) + 1, num_subj)),
index=np.append([0], all_labels))
for subj in range(num_subj):
# get rid of samples whose coordinates don't match ontological profile
annotation = process.drop_mismatch_samples(files.annotation[subj],
files.ontology[subj],
corrected=corrected_mni)
# subset representative probes + samples from microarray data
microarray = io.read_microarray(files.microarray[subj])
samples = microarray.loc[probes.index, annotation.index].T
samples.columns = probes.gene_symbol
# assign samples to regions and aggregate samples w/i the same region
sample_labels = label_samples(annotation,
atlas,
atlas_info=atlas_info,
tolerance=tolerance)
expression += [
group_by_label(samples, sample_labels, all_labels, metric=metric)
]
# get counts of samples collapsed into each ROI
labs, num = np.unique(sample_labels, return_counts=True)
counts.loc[labs, subj] = num
# if we don't want to do exact matching then cache which parcels are
# missing data and the expression data for the closest sample to that
# parcel; we'll use this once we've iterated through all donors
if not exact:
coords = utils.xyz_to_ijk(annotation[['mni_x', 'mni_y', 'mni_z']],
atlas.affine)
empty = ~np.in1d(all_labels, labs)
closest, dist = utils.closest_centroid(coords,
centroids[empty],
return_dist=True)
closest = samples.loc[annotation.iloc[closest].index]
empty = all_labels[empty]
closest.index = pd.Series(empty, name='label')
missing += [(closest, dict(zip(empty, np.diag(dist))))]
# check for missing ROIs and fill in, as needed
if not exact:
# find labels that are missing across all donors
empty = reduce(set.intersection, [set(f.index) for f, d in missing])
for roi in empty:
# find donor with sample closest to centroid of empty parcel
ind = np.argmin([d.get(roi) for f, d in missing])
# assign expression data from that sample and add to count
expression[ind].loc[roi] = missing[ind][0].loc[roi]
counts.loc[roi, ind] += 1
# normalize data with SRS and aggregate across donors
expression = [process.normalize_expression(e) for e in expression]
if not return_donors:
expression = process.aggregate_donors(expression, metric)
if return_counts:
return expression, counts.iloc[1:]
return expression