def full_dPCA_dprimes(site, contexts, probes, meta):
raster, goodcells = _load_site_formated_raste(site, contexts, probes, meta)
# trialR shape: Trial x Cell x Context x Probe x Time; R shape: Cell x Context x Probe x Time
trialR, R, _ = cdPCA.format_raster(raster)
# calculates full dPCA. i.e. considering all 4 categories
_, trialZ, dpca = cdPCA._cpp_dPCA(R, trialR)
dPCA_projection = trialZ['ct'][:, 0, ...]
dprime = cDP.pairwise_dprimes(dPCA_projection,
observation_axis=0,
condition_axis=1,
flip=meta['dprime_absolute'])
# calculates the variance explained. special case for full dpca, not present in other dprime approaches
var_capt = cdPCA.variance_captured(dpca, R)
# Shuffles the rasters n times and organizes in an array with the same shape the raster plus one dimension
# with size n containing each shuffle
# calculates the pairwise dprime
rep, ctx, prb, tme = dPCA_projection.shape
transition_pairs = list(itt.combinations(contexts, 2))
print(f"\nshuffling {meta['montecarlo']} times")
rng = np.random.default_rng(42)
shuf_projections = np.empty([meta['montecarlo'], rep, ctx, prb, tme])
shuf_trialR = trialR.copy()
for rr in range(meta['montecarlo']):
shuf_trialR = shuffle(shuf_trialR,
shuffle_axis=2,
indie_axis=0,
rng=rng)
shuf_R = np.mean(shuf_trialR, axis=0)
# saves the first regularizer to speed things up.
if rr == 0:
_, shuf_trialZ, shuf_dpca = cdPCA._cpp_dPCA(shuf_R, shuf_trialR)
regularizer = shuf_dpca.regularizer
else:
_, shuf_trialZ, dpca = cdPCA._cpp_dPCA(
shuf_R, shuf_trialR, {'regularizer': regularizer})
shuf_projections[rr, ...] = shuf_trialZ['ct'][:, 0, ...]
shuffled_dprime = cDP.pairwise_dprimes(shuf_projections,
observation_axis=1,
condition_axis=2,
flip=meta['dprime_absolute'])
# add dimension for single PC, for compatibility with single cell arrays
dprime = np.expand_dims(dprime, axis=0)
shuffled_dprime = np.expand_dims(shuffled_dprime, axis=1)
return dprime, shuffled_dprime, goodcells, var_capt
def fit_transformt(R, trialR):
_, dPCA_projection, _, dpca = cdPCA._cpp_dPCA(R,
trialR,
significance=False,
dPCA_parms={})
dPCA_projection = dPCA_projection['ct'][:, 0, ]
dPCA_transformation = np.tile(dpca.D['ct'][:, 0][:, None, None],
[1, 1, T])
return dPCA_projection, dPCA_transformation
def fit_transform(site, probe, meta, part):
recs = load(site)
if len(recs) > 2:
print(f'\n\n{recs.keys()}\n\n')
rec = recs['trip0']
sig = rec['resp']
# calculates response realiability and select only good cells to improve analysis
r_vals, goodcells = signal_reliability(sig,
r'\ASTIM_*',
threshold=meta['reliability'])
goodcells = goodcells.tolist()
raster = src.data.rasters.raster_from_sig(
sig,
probe,
channels=goodcells,
contexts=meta['transitions'],
smooth_window=meta['smoothing_window'],
raster_fs=meta['raster_fs'],
part=part,
zscore=meta['zscore'])
# trialR shape: Trial x Cell x Context x Probe x Time; R shape: Cell x Context x Probe x Time
trialR, R, _ = cdPCA.format_raster(raster)
trialR, R = trialR.squeeze(), R.squeeze() # squeezes out probe
_, dPCA_projection, _, dpca = cdPCA._cpp_dPCA(R,
trialR,
significance=False,
dPCA_parms={})
dPCA_projection = dPCA_projection['ct'][:, 0, ...]
dPCA_weights = np.tile(dpca.D['ct'][:, 0][:, None, None],
[1, 1, R.shape[-1]])
dprime = cDP.pairwise_dprimes(dPCA_projection)
return dprime, dPCA_projection, dPCA_weights, dpca
def dPCA_site_summary(site, probe):
# loads the raw data
recs = load(site, rasterfs=meta['raster_fs'], recache=rec_recache)
sig = recs['trip0']['resp']
# calculates response realiability and select only good cells to improve analysis
r_vals, goodcells = signal_reliability(sig,
r'\ASTIM_*',
threshold=meta['reliability'])
goodcells = goodcells.tolist()
# get the full data raster Context x Probe x Rep x Neuron x Time
raster = src.data.rasters.raster_from_sig(
sig,
probe,
channels=goodcells,
contexts=meta['transitions'],
smooth_window=meta['smoothing_window'],
raster_fs=meta['raster_fs'],
zscore=meta['zscore'],
part='probe')
# trialR shape: Trial x Cell x Context x Probe x Time; R shape: Cell x Context x Probe x Time
trialR, R, _ = cdPCA.format_raster(raster)
trialR, R = trialR.squeeze(axis=3), R.squeeze(axis=2) # squeezes out probe
Z, trialZ, dpca = cdPCA._cpp_dPCA(R, trialR)
fig, axes = plt.subplots(2, 3, sharex='all', sharey='row')
for vv, (marginalization, arr) in enumerate(Z.items()):
means = Z[marginalization]
trials = trialZ[marginalization]
if marginalization == 't':
marginalization = 'probe'
elif marginalization == 'ct':
marginalization = 'context'
for pc in range(3): # first 3 principal components
ax = axes[vv, pc]
for tt, trans in enumerate(
meta['transitions']): # for each context
ax.plot(times,
means[pc, tt, :],
label=trans,
color=trans_color_map[trans],
linewidth=2)
# _ = fplt._cint(times, trials[:,pc,tt,:], confidence=0.95, ax=ax,
# fillkwargs={'color': trans_color_map[trans], 'alpha': 0.5})
ax.tick_params(labelsize=ax_val_size)
# formats axes labels and ticks
if pc == 0: # y labels
ax.set_ylabel(
f'{marginalization} dependent\nfiring rate (z-score)',
fontsize=ax_lab_size)
else:
ax.axes.get_yaxis().set_visible(True)
pass
if vv == 0:
ax.set_title(f'dPC #{pc + 1}', fontsize=sub_title_size)
ax.axes.get_xaxis().set_visible(True)
else:
ax.set_xlabel('time (ms)', fontsize=ax_lab_size)
## Hide the right and top spines
ax.spines['right'].set_visible(False)
ax.spines['top'].set_visible(False)
ax.tick_params(labelsize=ax_val_size)
# legend in last axis
axes[-1, -1].legend(loc='upper right',
fontsize='x-large',
markerscale=10,
frameon=False)
return fig, ax, dpca
def probewise_dPCA_dprimes(site, contexts, probes, meta):
"""
performs dimensionality reduction with dPCA done independently for each probe. Then uses the first context dependent
demixed principal component to calculated the dprime between context for all probes in the site. Calculates
significance using montecarlo shuffling of the context identity.
:param site: string identifying the site
:param probes: list of probe numbers
:param meta: dict with meta parameters
:return: dprime (ndarray with shape PC x Ctx_pair x Probe x Time),
shuffled_dprimes (ndarray with shape Montecarlo x PC x Ctx_pair x Probe x Time),
goocells (list of strings)
"""
raster, goodcells = _load_site_formated_raste(site, contexts, probes, meta)
# trialR shape: Trial x Cell x Context x Probe x Time; R shape: Cell x Context x Probe x Time
trialR, R, _ = cdPCA.format_raster(raster)
rep, unt, ctx, prb, tme = trialR.shape
transition_pairs = list(itt.combinations(contexts, 2))
# iterates over each probe
dprime = np.empty([len(transition_pairs), prb, tme])
shuffled_dprime = np.empty(
[meta['montecarlo'],
len(transition_pairs), prb, tme])
var_capt = list()
for pp in probes:
probe_idx = probes.index(pp)
probe_trialR = trialR[..., probe_idx, :]
probe_R = R[..., probe_idx, :]
# calculates dPCA considering all 4 categories
_, trialZ, dpca = cdPCA._cpp_dPCA(probe_R, probe_trialR)
var_capt.append(cdPCA.variance_captured(dpca, probe_R))
dPCA_projection = trialZ['ct'][:, 0, ...]
dprime[:, probe_idx, :] = cDP.pairwise_dprimes(
dPCA_projection,
observation_axis=0,
condition_axis=1,
flip=meta['dprime_absolute'])
# Shuffles the rasters n times and organizes in an array with the same shape as the original raster plus one
# dimension with size meta['montecarlo'] containing each shuffle
# calculates the pairwise dprime
print(f"\nshuffling {meta['montecarlo']} times")
rng = np.random.default_rng(42)
shuf_projections = np.empty([meta['montecarlo'], rep, ctx, tme])
shuf_projections[:] = np.nan
shuf_trialR = probe_trialR.copy()
for rr in range(meta['montecarlo']):
shuf_trialR = shuffle(shuf_trialR,
shuffle_axis=2,
indie_axis=0,
rng=rng)
shuf_R = np.mean(shuf_trialR, axis=0)
#saves the first regularizer to speed things up.
if rr == 0:
_, shuf_trialZ, shuf_dpca = cdPCA._cpp_dPCA(
shuf_R, shuf_trialR)
regularizer = shuf_dpca.regularizer
else:
_, shuf_trialZ, dpca = cdPCA._cpp_dPCA(
shuf_R, shuf_trialR, {'regularizer': regularizer})
shuf_projections[rr, :] = shuf_trialZ['ct'][:, 0, ...]
shuffled_dprime[:, :, probe_idx, :] = cDP.pairwise_dprimes(
shuf_projections,
observation_axis=1,
condition_axis=2,
flip=meta['dprime_absolute'])
# add dimension for single PC, for compatibility with single cell arrays
dprime = np.expand_dims(dprime, axis=0)
shuffled_dprime = np.expand_dims(shuffled_dprime, axis=1)
return dprime, shuffled_dprime, goodcells, var_capt
sim_aca = [sim_trialR[:, :, ctx, :] for ctx in range(S)]
sim_dPCA_proj = [cLDA.transform_over_time(a, dpca_axes) for a in sim_aca]
dPCA_ceil_d[ii, :] = cpd.dprime(sim_dPCA_proj[c0][:, 0, :],
sim_dPCA_proj[c1][:, 0, :])
# dprime floor: shuffle context(dim2) identity by trial(dim0), projection and dprime calculation
ctx_shuffle = shuffle(ctx_shuffle, shuffle_axis=2, indie_axis=0)
# trial-average data
shuf_R = np.mean(ctx_shuffle, 0)
# center data per neuron i.e. first dime
centers = np.mean(shuf_R.reshape((shuf_R.shape[0], -1)), axis=1)[:, None,
None]
shuf_R -= centers
# calculates dpca and projections
_, shuf_trialZ, _, _ = cdPCA._cpp_dPCA(shuf_R,
ctx_shuffle,
significance=False)
shuf_dPCA_proj = [
shuf_trialZ['ct'][:, 0, ctx, :]
for ctx in range(shuf_trialZ['ct'].shape[2])
]
dPCA_floor_d[ii, :] = cpd.dprime(shuf_dPCA_proj[c0].squeeze(),
shuf_dPCA_proj[c1].squeeze())
# plots figure
fig, axes = plt.subplots(1, 2, sharey=True)
axes = np.ravel(axes)
ci = 0.99
axes[0].plot(np.mean(LDA_floor_d, axis=0), color='red', alpha=1, label='floor')