def make_surrogates(data, parcellation, scale, spatnull):
"""
Generates surrogates for `data` using `spatnull` method
Parameters
----------
data : (N,) pd.DataFrame
parcellation : {'atl-cammoun2012', 'atl-schaefer2018'}
scale : str
spatnull : {'burt2018', 'burt2020', 'moran'}
Returns
-------
surrogates : (N, `N_PERM`) np.ndarray
"""
if spatnull not in ('burt2018', 'burt2020', 'moran'):
raise ValueError(f'Cannot make surrogates for null method {spatnull}')
darr = np.asarray(data)
dmin = darr[np.logical_not(np.isnan(darr))].min()
surrogates = np.zeros((len(data), N_PERM))
for hdata, dist, idx in putils.yield_data_dist(
DISTDIR, parcellation, scale, data, inverse=(spatnull == 'moran')
):
# handle NaNs before generating surrogates; should only be relevant
# when using vertex-level data, but good nonetheless
mask = np.logical_not(np.isnan(hdata))
surrogates[idx[np.logical_not(mask)]] = np.nan
hdata, dist, idx = hdata[mask], dist[np.ix_(mask, mask)], idx[mask]
if spatnull == 'burt2018':
# Box-Cox transformation requires positive data :man_facepalming:
hdata += np.abs(dmin) + 0.1
surrogates[idx] = \
burt.batch_surrogates(dist, hdata, n_surr=N_PERM, seed=SEED)
elif spatnull == 'burt2020':
if parcellation == 'vertex': # memmap is required for this shit
index = np.argsort(dist, axis=-1)
dist = np.sort(dist, axis=-1)
knn = 1000 if USE_KNN else len(hdata)
surrogates[idx] = \
mapgen.Sampled(hdata, dist, index, knn=knn,
seed=SEED)(N_PERM).T
else:
surrogates[idx] = \
mapgen.Base(hdata, dist, seed=SEED)(N_PERM, 50).T
elif spatnull == 'moran':
mrs = moran.MoranRandomization(joint=True, n_rep=N_PERM,
tol=1e-6, random_state=SEED)
with threadpoolctl.threadpool_limits(limits=2):
surrogates[idx] = mrs.fit(dist).randomize(hdata).T
return surrogates
def make_surrogates(data, parcellation, scale, spatnull, fn=None):
if spatnull not in ('burt2018', 'burt2020', 'moran'):
raise ValueError(f'Cannot make surrogates for null method {spatnull}')
darr = np.asarray(data)
dmin = darr[np.logical_not(np.isnan(darr))].min()
surrogates = np.zeros((len(data), N_PERM))
for n, hemi in enumerate(('lh', 'rh')):
dist = get_distmat(hemi, parcellation, scale, fn=fn)
try:
idx = np.asarray([
n for n, f in enumerate(data.index)if f.startswith(hemi)
])
hdata = np.squeeze(np.asarray(data.iloc[idx]))
except AttributeError:
idx = np.arange(n * (len(data) // 2), (n + 1) * (len(data) // 2))
hdata = np.squeeze(data[idx])
# handle NaNs before generating surrogates; should only be relevant
# when using vertex-level data, but good nonetheless
mask = np.logical_not(np.isnan(hdata))
surrogates[idx[np.logical_not(mask)]] = np.nan
hdata, dist, idx = hdata[mask], dist[np.ix_(mask, mask)], idx[mask]
if spatnull == 'burt2018':
# Box-Cox transformation requires positive data
hdata += np.abs(dmin) + 0.1
surrogates[idx] = \
burt.batch_surrogates(dist, hdata, n_surr=N_PERM, seed=SEED)
elif spatnull == 'burt2020':
if parcellation == 'vertex':
index = np.argsort(dist, axis=-1)
dist = np.sort(dist, axis=-1)
surrogates[idx] = \
mapgen.Sampled(hdata, dist, index, seed=SEED)(N_PERM).T
else:
surrogates[idx] = \
mapgen.Base(hdata, dist, seed=SEED)(N_PERM, 50).T
elif spatnull == 'moran':
dist = dist.astype('float64') # required for some reason...
np.fill_diagonal(dist, 1)
dist **= -1
mrs = moran.MoranRandomization(joint=True, n_rep=N_PERM,
tol=1e-6, random_state=SEED)
surrogates[idx] = mrs.fit(dist).randomize(hdata).T
return surrogates
def run_null(netclass, parc, scale, spintype):
"""
Runs spatial permutation null model for given combination of inputs
Parameters
----------
netclass : {'vek', 'yeo'}
Network partition to test
parc : str
Name of parcellation to be used
scale : str
Scale of `parcellation` to be used
spintype : str
Name of spin method to be used
Returns
-------
stats : pd.DataFrame
Generated statistics with columns ['parcellation', 'scale', 'spintype',
'netclass', 'network', 'zscore', 'pval']
"""
data = load_data(netclass, parc, scale)
# run the damn thing
print(f'Running {spintype:>9} spins for {scale}: ', end='', flush=True)
out = HCPDIR / parc / 'nulls' / netclass / spintype / f'{scale}_nulls.csv'
if out.exists():
permnets = np.loadtxt(out, delimiter=',')
elif spintype == 'cornblath':
# even though we're working with parcellated data we need to project
# that to the surface + spin the vertices, so let's load our
# pre-generated vertex-level spins
spins = SPDIR / 'vertex' / 'vazquez-rodriguez' / 'fsaverage5_spins.csv'
# get annotation files (we need these to project parcels to surface)
fetcher = getattr(nndata, f"fetch_{parc.replace('atl-', '')}")
annotations = fetcher('fsaverage5', data_dir=ROIDIR)[scale]
# pre-load the spins for this function (assumes `spins` is array)
print('Pre-loading spins...', end='\b' * 20, flush=True)
spins = np.loadtxt(spins, delimiter=',', dtype='int32')
# generate "spun" data; permdata will be an (R, T, n_rotate) array
# where `R` is regions and `T` is 1 (myelination)
permdata = nnsurf.spin_data(np.asarray(data['myelin']),
version='fsaverage5',
lhannot=annotations.lh,
rhannot=annotations.rh,
spins=spins,
n_rotate=spins.shape[-1],
verbose=True)
permnets = np.vstack([
_get_netmeans(permdata[..., n], data['networks'])
for n in range(spins.shape[-1])
])
putils.save_dir(out, permnets)
elif spintype in ['burt2018', 'burt2020']:
surrdir = SURRDIR / parc / spintype / 'hcp'
surrogates = get_surrogates(data['myelin'], surrdir, scale)
permnets = np.vstack([
_get_netmeans(surrogates[..., n], data['networks'])
for n in range(surrogates.shape[-1])
])
putils.save_dir(out, permnets)
elif spintype == 'moran':
surrogates = np.zeros((len(data['myelin']), 10000))
for hemi, dist, idx in putils.yield_data_dist(DISTDIR, parc, scale,
data['myelin']):
mrs = moran.MoranRandomization(joint=True,
n_rep=10000,
tol=1e-6,
random_state=1234)
mrs.fit(dist)
surrogates[idx] = np.squeeze(mrs.randomize(hemi)).T
permnets = np.vstack([
_get_netmeans(surrogates[..., n], data['networks'])
for n in range(surrogates.shape[-1])
])
putils.save_dir(out, permnets)
else:
spins = SPDIR / parc / spintype / f'{scale}_spins.csv'
permnets = gen_permnets(data['myelin'], data['networks'], spins, out)
# now get the real network averages and compare to the permuted values
real = _get_netmeans(data['myelin'], data['networks'])
zscores, pvals = get_fwe(real, permnets)
out = pd.DataFrame(
dict(parcellation=parc,
scale=scale,
spintype=spintype,
netclass=netclass,
network=list(NET_CODES[netclass].keys()),
zscore=zscores,
pval=pvals))
return out
lhdata, rhdata = data[:end - start], data[end - start:]
lhdist = np.loadtxt(DISTDIR / name / 'nomedial' / f'{scale}_lh_dist.csv',
delimiter=',')
rhdist = np.loadtxt(DISTDIR / name / 'nomedial' / f'{scale}_rh_dist.csv',
delimiter=',')
plot = np.hstack((burt.make_surrogate(lhdist, lhdata + 1, seed=1234),
burt.make_surrogate(rhdist, rhdata + 1, seed=1234)))
save_brainmap(plot, FIGDIR / 'burt2018_surf.png', lh, rh, **OPTS)
# burt 2020 (need to rescale to original data range)
plot = np.hstack((nnutils.rescale(
Base(lhdata, lhdist, seed=1234)(200, 50).T[:, 180], lhdata.min(),
lhdata.max()),
nnutils.rescale(
Base(rhdata, rhdist, seed=1234)(200, 50).T[:, 180],
rhdata.min(), rhdata.max())))
save_brainmap(plot, FIGDIR / 'burt2020_surf.png', lh, rh, **OPTS)
# moran spectral randomization
np.fill_diagonal(lhdist, 1)
np.fill_diagonal(rhdist, 1)
lhdist **= -1
rhdist **= -1
mrs = moran.MoranRandomization(joint=True,
n_rep=1000,
tol=1e-6,
random_state=1234)
plot = np.hstack((np.squeeze(mrs.fit(lhdist).randomize(lhdata)),
np.squeeze(mrs.fit(rhdist).randomize(rhdata))))[611]
save_brainmap(plot, FIGDIR / 'moran_surf.png', lh, rh, **OPTS)
def run_null(parcellation, scale, spintype):
"""
Runs spatial permutation null model for given combination of inputs
Parameters
----------
parcellation : str
Name of parcellation to be used
scale : str
Scale of `parcellation` to be used
spintype : str
Name of spin method to be used
Returns
-------
stats : pd.DataFrame
Generated statistics with columns ['parcellation', 'scale', 'spintype',
'n_sig']
"""
nsdata = load_data(parcellation, scale)
# run the damn thing
print(f'Running {spintype:>9} spins for {scale}: ', end='', flush=True)
out = NSDIR / parcellation / 'nulls' / spintype / f'{scale}_nulls.csv'
if out.exists():
permcorrs = np.loadtxt(out).reshape(-1, 1)
elif spintype == 'cornblath':
# even though we're working with parcellated data we need to project
# that to the surface + spin the vertices, so let's load our
# pre-generated vertex-level spins
spins = SPDIR / 'vertex' / 'vazquez-rodriguez' / 'fsaverage5_spins.csv'
# get annotation files
fetcher = getattr(nndata, f"fetch_{parcellation.replace('atl-', '')}")
annotations = fetcher('fsaverage5', data_dir=ROIDIR)[scale]
# pre-load the spins for this function (assumes `spins` is array)
# permdata will be an (R, T, n_rotate) array
print('Pre-loading spins...', end='\b' * 20, flush=True)
spins = np.loadtxt(spins, delimiter=',', dtype='int32')
permdata = nnsurf.spin_data(nsdata, version='fsaverage5',
lhannot=annotations.lh,
rhannot=annotations.rh,
spins=spins, n_rotate=spins.shape[-1],
verbose=True)
permcorrs = np.vstack([
_get_permcorr(nsdata, permdata[..., n])
for n in range(permdata.shape[-1])
])
putils.save_dir(out, permcorrs)
elif spintype in ['burt2018', 'burt2020']:
surrdir = SURRDIR / parcellation / spintype / 'neurosynth'
# generate the permuted data from the surrogate resampling arrays
print('Generating surrogates...', end='\b' * 24, flush=True)
permdata = get_surrogates(nsdata, surrdir, scale)
permcorrs = np.vstack([
_get_permcorr(nsdata, permdata[..., n])
for n in range(permdata.shape[-1])
])
putils.save_dir(out, permcorrs)
elif spintype == 'moran':
surrogates = np.zeros((*nsdata.shape, 10000))
for hemi, dist, idx in putils.yield_data_dist(DISTDIR, parcellation,
scale, nsdata):
mrs = moran.MoranRandomization(joint=True, n_rep=10000,
tol=1e-6, random_state=1234)
mrs.fit(dist)
surrogates[idx] = mrs.randomize(hemi).transpose(1, 2, 0)
permcorrs = np.vstack([
_get_permcorr(nsdata, surrogates[..., n])
for n in range(surrogates.shape[-1])
])
putils.save_dir(out, permcorrs)
else:
spins = SPDIR / parcellation / spintype / f'{scale}_spins.csv'
permcorrs = gen_permcorrs(nsdata, spins, out)
nsdata = nsdata.dropna(axis=0, how='all')
pvals = get_fwe(np.corrcoef(nsdata.T), permcorrs)
out = pd.DataFrame(dict(
parcellation=parcellation,
scale=scale,
spintype=spintype,
n_sig=np.sum(np.triu(pvals < ALPHA, k=1))
), index=[0])
return out
