def multiflip_mcmc_sweep(self,
n_steps=1000,
beta=np.inf,
niter=10,
verbose=True):
'''
Fit the sbm to the word-document network. Use multtiplip_mcmc_sweep
- n_steps, int (default:1): number of steps.
'''
g = self.g
if g is None:
print('No data to fit the SBM. Load some data first (make_graph)')
else:
clabel = g.vp['kind']
state_args = {'clabel': clabel, 'pclabel': clabel}
if "count" in g.ep:
state_args["eweight"] = g.ep.count
state = self.state
if state is not None:
state = state.copy(bs=state.get_bs() + [np.zeros(1)] * 4,
sampling=True)
else:
state = gt.NestedBlockState(g)
for step in range(n_steps): # this should be sufficiently large
if verbose:
print(f"step: {step}")
state.multiflip_mcmc_sweep(beta=beta, niter=niter)
self.state = state
## minimum description length
self.mdl = self.state.entropy()
## collect group membership for each level in the hierarchy
L = len(state.levels)
dict_groups_L = {}
## only trivial bipartite structure
if L == 2:
self.L = 1
for l in range(L - 1):
dict_groups_l = self.get_groups(l=l)
dict_groups_L[l] = dict_groups_l
## omit trivial levels: l=L-1 (single group), l=L-2 (bipartite)
else:
self.L = L - 2
for l in range(L - 2):
dict_groups_l = self.get_groups(l=l)
dict_groups_L[l] = dict_groups_l
self.groups = dict_groups_L
def state_from_blocks(
adata: AnnData,
state_key: Optional[str] = 'nsbm',
neighbors_key: Optional[str] = 'neighbors',
adjacency: Optional[spmatrix] = None,
directed: bool = False,
use_weights: bool = False,
deg_corr: bool = True,
):
"""
Returns a gt state object given an AnnData
Parameters
----------
adata
The annotated data matrix.
state_key
The key under which the state has been saved
neighbors_key
The key passed to `sc.pp.neighbors`
adjacency
Sparse adjacency matrix of the graph, defaults to
`adata.uns['neighbors']['connectivities']` in case of scanpy<=1.4.6 or
`adata.obsp[neighbors_key][connectivity_key]` for scanpy>1.4.6
directed
Whether to treat the graph as directed or undirected.
use_weights
If `True`, edge weights from the graph are used in the computation
(placing more emphasis on stronger edges). Note that this
increases computation times
deg_corr
Whether to use degree correction in the minimization step. In many
real world networks this is the case, although this doesn't seem
the case for KNN graphs used in scanpy.
Returns
-------
Nothing, adds a `gt.block_state` object in adata.uns
"""
bl_d = adata.uns['schist'][f'{state_key}']['blocks']
params = adata.uns['schist'][f'{state_key}']['params']
if params['model'] == 'nested' or params['model'] == 'multiome_nested':
blocks = []
for nl in range(len(bl_d)):
blocks.append(bl_d[str(nl)])
else:
blocks = bl_d['0']
if 'deg_corr' in params:
deg_corr=params['deg_corr']
recs=[]
rec_types=[]
if use_weights:
# this is not ideal to me, possibly we may need to transform
# weights. More tests needed.
recs=[g.ep.weight]
rec_types=['real-normal']
if 'recs' in params:
recs=params['recs']
if 'rec_types' in params:
rec_types=params['rec_types']
if adjacency is None:
if neighbors_key not in adata.uns:
raise ValueError(
'You need to run `pp.neighbors` first '
'to compute a neighborhood graph.'
)
elif 'connectivities_key' in adata.uns[neighbors_key]:
# scanpy>1.4.6 has matrix in another slot
conn_key = adata.uns[neighbors_key]['connectivities_key']
adjacency = adata.obsp[conn_key]
else:
# scanpy<=1.4.6 has sparse matrix here
adjacency = adata.uns[neighbors_key]['connectivities']
g = get_igraph_from_adjacency(adjacency, directed=directed)
g = g.to_graph_tool()
gt.remove_parallel_edges(g)
if params['model'] == 'flat':
state = gt.BlockState(g, b=blocks,
state_args=dict(deg_corr=deg_corr,
recs=recs,
rec_types=rec_types)
)
elif params['model'] == 'ppbm':
state = gt.PPBlockState(g, b=blocks,
state_args=dict(deg_corr=deg_corr,
recs=recs,
rec_types=rec_types)
)
else:
state = gt.NestedBlockState(g, bs=blocks,
state_args=dict(deg_corr=deg_corr,
recs=recs,
rec_types=rec_types)
)
return state
def codon_circos(cmap='tab20', filetype="pdf", reverse=False):
cm = plt.cm.get_cmap(cmap)
cmappable = ScalarMappable(norm=Normalize(vmin=0, vmax=20), cmap=cm)
g_codons = gt.Graph(directed=False)
g_codons.vp.codon = g_codons.new_vertex_property("string")
g_codons.vp.aa = g_codons.new_vertex_property("string")
g_codons.vp.aa_index = g_codons.new_vertex_property("int")
g_codons.vp.aa_color = g_codons.new_vertex_property("vector<float>")
g_codons.vp.codon_index = g_codons.new_vertex_property("int")
g_codons.ep.syn = g_codons.new_edge_property("bool")
g_codons.ep.grad = g_codons.new_edge_property("vector<float>")
for aa_index, aa in enumerate(aa_order):
if aa == "X": continue
for codon_index, c in enumerate(sorted([k for k, v in codontable.items() if v == aa])):
v = g_codons.add_vertex()
g_codons.vp.codon[v] = c
g_codons.vp.aa[v] = aa
g_codons.vp.codon_index[v] = codon_index
g_codons.vp.aa_index[v] = aa_index
g_codons.vp.aa_color[v] = cmappable.to_rgba(aa_index)
for ref in g_codons.vertices():
for alt in g_codons.vertices():
if alt <= ref: continue
codon_ref, codon_alt = g_codons.vp.codon[ref], g_codons.vp.codon[alt]
if distance_str(codon_ref, codon_alt) != 1: continue
if codontable[codon_ref] != codontable[codon_alt]:
e_c = g_codons.add_edge(ref, alt)
g_codons.ep.syn[e_c] = False
x = cmappable.to_rgba(g_codons.vp.aa_index[ref])[:3]
y = cmappable.to_rgba(g_codons.vp.aa_index[alt])[:3]
if reverse: x, y = y, x
g_codons.ep.grad[e_c] = [0.0, *x, 0.75, 1.0, *y, 0.75]
for ref in g_codons.vertices():
for alt in g_codons.vertices():
if alt >= ref: continue
codon_ref, codon_alt = g_codons.vp.codon[ref], g_codons.vp.codon[alt]
if distance_str(codon_ref, codon_alt) != 1: continue
if codontable[codon_ref] == codontable[codon_alt]:
e_c = g_codons.add_edge(ref, alt)
g_codons.ep.syn[e_c] = True
syn_color = 0.0, 0.0, 0.0, 1.0
g_codons.ep.grad[e_c] = [0.0, *syn_color, 1.0, *syn_color]
assert g_codons.num_vertices() == 61
dist = gt.shortest_distance(g_codons)
r = max([max(dist[g_codons.vertex(i)].a) for i in g_codons.vertices()])
print('Codons graph radius : {0}'.format(r))
print('Codons : {0} transitions out of {1} possibles.'.format(g_codons.num_edges(), 61 * 60 / 2))
syn_array = g_codons.ep.syn.get_array()
print('Codons : {0} are synonymous and {1} are non-synonymous.'.format(sum(syn_array),
len(syn_array) - sum(syn_array)))
state = gt.NestedBlockState(g_codons, bs=[g_codons.vp.aa_index, g_codons.vp.codon_index], sampling=False)
t = gt.get_hierarchy_tree(state)[0]
tpos = gt.radial_tree_layout(t, t.vertex(t.num_vertices() - 1), weighted=True)
cts = gt.get_hierarchy_control_points(g_codons, t, tpos)
pos = g_codons.own_property(tpos)
gt.graph_draw(g_codons, pos=pos, edge_control_points=cts, edge_gradient=g_codons.ep.grad, edge_pen_width=2.5,
vertex_text=g_codons.vp.codon, vertex_anchor=0, vertex_font_size=9, vertex_pen_width=1.6,
vertex_color=(0.65, 0.65, 0.65, 1), vertex_fill_color=g_codons.vp.aa_color, vertex_size=25.0,
output="gt-codon-{0}.{1}".format(cmap, filetype))
ecolor[e] = 'green'
i += 1
g.edge_properties["weight"] = eweight
g.edge_properties["color"] = ecolor
'''
#pos = gt.planar_layout(g)
#pos = gt.radial_tree_layout(g, g.vertex(0))
for i in range(3):
pos = gt.arf_layout(g, d = 1, a = 5, max_iter=0) # good
#pos = gt.fruchterman_reingold_layout(g, n_iter=1000)
#pos = gt.sfdp_layout(g, C = 1)
#pos = gt.circle_layout(g)
#gt.graph_draw(g, pos = pos, vertex_text=g.vertex_properties["name"], vertex_font_size=20, vertex_size=10, vertex_color = 'white', vertex_fill_color = 'blue', vertex_text_position=0, output_size=(2000, 1000), output="imgs/small_graph_top_" + str(i) + ".pdf")
#gt.graph_draw(g, pos = pos, vertex_text=g.vertex_properties["name"], vertex_font_size=20, vertex_size=10, vertex_color = 'white', vertex_fill_color = 'blue', vertex_text_position=0, output_size=(2000, 1000), output="imgs/small_graph_top_" + str(i) + ".png")
state = gt.minimize_blockmodel_dl(g) # , deg_corr=True, B_min = 10
state.draw(pos=pos, vertex_shape=state.get_blocks(), vertex_text=g.vertex_properties["name"], vertex_font_size=20, vertex_size=20, edge_pen_width = 2, vertex_text_position=0, output="small_graph_top/small_graph_top_blocks_mdl_" + str(i) + ".pdf", output_size=(1500, 1000), fit_view=1.1)
state.draw(pos=pos, vertex_shape=state.get_blocks(), vertex_text=g.vertex_properties["name"], vertex_font_size=20, vertex_size=20, edge_pen_width = 2, vertex_text_position=0, output="small_graph_top/small_graph_top_blocks_mdl_" + str(i) + ".png", output_size=(1500, 1000), fit_view=1.1)
print(i)
#gt.draw_hierarchy(state, layout="sfdp", vertex_text=g.vertex_properties["name"], vertex_font_size=24, vertex_text_position="centered", edge_color=g.edge_properties["color"], output_size=(2000, 1000), output="small_graph_mdl.pdf", fit_view = 0.8, hide = 2)
print(vchrom)
print(np.array(vchrom))
state = gt.NestedBlockState(g, [np.array(vchrom), np.arange(0, 22)])
gt.draw_hierarchy(state, vertex_text=g.vertex_properties["name"], vertex_font_size=24, vertex_text_position="centered", output_size=(2000, 1000), output="small_graph_top/small_graph_top_mdl.pdf", fit_view = 0.8, hide = 2)
def nested_model(
adata: AnnData,
max_iterations: int = 1000000,
epsilon: float = 0,
equilibrate: bool = False,
wait: int = 1000,
nbreaks: int = 2,
collect_marginals: bool = False,
niter_collect: int = 10000,
hierarchy_length: int = 10,
deg_corr: bool = True,
multiflip: bool = True,
fast_model: bool = False,
fast_tol: float = 1e-6,
n_sweep: int = 10,
beta: float = np.inf,
n_init: int = 1,
beta_range: Tuple[float] = (1., 1000.),
steps_anneal: int = 3,
resume: bool = False,
*,
restrict_to: Optional[Tuple[str, Sequence[str]]] = None,
random_seed: Optional[int] = None,
key_added: str = 'nsbm',
adjacency: Optional[sparse.spmatrix] = None,
neighbors_key: Optional[str] = 'neighbors',
directed: bool = False,
use_weights: bool = False,
prune: bool = False,
return_low: bool = False,
copy: bool = False,
minimize_args: Optional[Dict] = {},
equilibrate_args: Optional[Dict] = {},
) -> Optional[AnnData]:
"""\
Cluster cells into subgroups [Peixoto14]_.
Cluster cells using the nested Stochastic Block Model [Peixoto14]_,
a hierarchical version of Stochastic Block Model [Holland83]_, performing
Bayesian inference on node groups. NSBM should circumvent classical
limitations of SBM in detecting small groups in large graphs
replacing the noninformative priors used by a hierarchy of priors
and hyperpriors.
This requires having ran :func:`~scanpy.pp.neighbors` or
:func:`~scanpy.external.pp.bbknn` first.
Parameters
----------
adata
The annotated data matrix.
max_iterations
Maximal number of iterations to be performed by the equilibrate step.
epsilon
Relative changes in entropy smaller than epsilon will
not be considered as record-breaking.
equilibrate
Whether or not perform the mcmc_equilibrate step.
Equilibration should always be performed. Note, also, that without
equilibration it won't be possible to collect marginals.
collect_marginals
Whether or not collect node probability of belonging
to a specific partition.
niter_collect
Number of iterations to force when collecting marginals. This will
increase the precision when calculating probabilites
wait
Number of iterations to wait for a record-breaking event.
Higher values result in longer computations. Set it to small values
when performing quick tests.
nbreaks
Number of iteration intervals (of size `wait`) without
record-breaking events necessary to stop the algorithm.
hierarchy_length
Initial length of the hierarchy. When large values are
passed, the top-most levels will be uninformative as they
will likely contain the very same groups. Increase this valus
if a very large number of cells is analyzed (>100.000).
deg_corr
Whether to use degree correction in the minimization step. In many
real world networks this is the case, although this doesn't seem
the case for KNN graphs used in scanpy.
multiflip
Whether to perform MCMC sweep with multiple simultaneous moves to sample
network partitions. It may result in slightly longer runtimes, but under
the hood it allows for a more efficient space exploration.
fast_model
Whether to skip initial minization step and let the MCMC find a solution.
This approach tend to be faster and consume less memory, but may be
less accurate.
fast_tol
Tolerance for fast model convergence.
n_sweep
Number of iterations to be performed in the fast model MCMC greedy approach
beta
Inverse temperature for MCMC greedy approach
n_init
Number of initial minimizations to be performed. The one with smaller
entropy is chosen
beta_range
Inverse temperature at the beginning and the end of the equilibration
steps_anneal
Number of steps in which the simulated annealing is performed
resume
Start from a previously created model, if any, without initializing a novel
model
key_added
`adata.obs` key under which to add the cluster labels.
adjacency
Sparse adjacency matrix of the graph, defaults to
`adata.uns['neighbors']['connectivities']` in case of scanpy<=1.4.6 or
`adata.obsp[neighbors_key][connectivity_key]` for scanpy>1.4.6
neighbors_key
The key passed to `sc.pp.neighbors`
directed
Whether to treat the graph as directed or undirected.
use_weights
If `True`, edge weights from the graph are used in the computation
(placing more emphasis on stronger edges). Note that this
increases computation times
prune
Some high levels in hierarchy may contain the same information in terms of
cell assignments, even if they apparently have different group names. When this
option is set to `True`, the function only returns informative levels.
Note, however, that cell affinities are still reported for all levels. Pruning
does not rename group levels
return_low
Whether or not return nsbm_level_0 in adata.obs. This level usually contains
so many groups that it cannot be plot anyway, but it may be useful for particular
analysis. By default it is not returned
copy
Whether to copy `adata` or modify it inplace.
random_seed
Random number to be used as seed for graph-tool
Returns
-------
`adata.obs[key_added]`
Array of dim (number of samples) that stores the subgroup id
(`'0'`, `'1'`, ...) for each cell.
`adata.uns['nsbm']['params']`
A dict with the values for the parameters `resolution`, `random_state`,
and `n_iterations`.
`adata.uns['nsbm']['stats']`
A dict with the values returned by mcmc_sweep
`adata.uns['nsbm']['cell_affinity']`
A `np.ndarray` with cell probability of belonging to a specific group
`adata.uns['nsbm']['state']`
The NestedBlockModel state object
"""
if resume:
# if the fast_model is chosen perform equilibration anyway
# also if a model has previously created
equilibrate = True
if resume and ('nsbm' not in adata.uns
or 'state' not in adata.uns['nsbm']):
# let the model proceed as default
logg.warning('Resuming has been specified but a state was not found\n'
'Will continue with default minimization step')
resume = False
if random_seed:
np.random.seed(random_seed)
gt.seed_rng(random_seed)
if collect_marginals:
logg.warning('Collecting marginals has a large impact on running time')
if not equilibrate:
raise ValueError(
"You can't collect marginals without MCMC equilibrate "
"step. Either set `equlibrate` to `True` or "
"`collect_marginals` to `False`")
start = logg.info('minimizing the nested Stochastic Block Model')
adata = adata.copy() if copy else adata
# are we clustering a user-provided graph or the default AnnData one?
if adjacency is None:
if neighbors_key not in adata.uns:
raise ValueError('You need to run `pp.neighbors` first '
'to compute a neighborhood graph.')
elif 'connectivities_key' in adata.uns[neighbors_key]:
# scanpy>1.4.6 has matrix in another slot
conn_key = adata.uns[neighbors_key]['connectivities_key']
adjacency = adata.obsp[conn_key]
else:
# scanpy<=1.4.6 has sparse matrix here
adjacency = adata.uns[neighbors_key]['connectivities']
if restrict_to is not None:
restrict_key, restrict_categories = restrict_to
adjacency, restrict_indices = restrict_adjacency(
adata,
restrict_key,
restrict_categories,
adjacency,
)
# convert it to igraph
g = get_graph_tool_from_adjacency(adjacency, directed=directed)
recs = []
rec_types = []
if use_weights:
# this is not ideal to me, possibly we may need to transform
# weights. More tests needed.
recs = [g.ep.weight]
rec_types = ['real-normal']
if n_init < 1:
n_init = 1
if fast_model:
# do not minimize, start with a dummy state and perform only equilibrate
states = [
gt.NestedBlockState(g=g,
state_args=dict(deg_corr=deg_corr,
recs=recs,
rec_types=rec_types))
for n in range(n_init)
]
for x in range(n_init):
dS = 1
while np.abs(dS) > fast_tol:
# perform sweep until a tolerance is reached
dS, _, _ = states[x].multiflip_mcmc_sweep(beta=beta,
niter=n_sweep)
_amin = np.argmin([s.entropy() for s in states])
state = states[_amin]
# dS = 1
# while np.abs(dS) > fast_tol:
# dS, nattempts, nmoves = state.multiflip_mcmc_sweep(niter=10, beta=np.inf)
bs = state.get_bs()
logg.info(' done', time=start)
elif resume:
# create the state and make sure sampling is performed
state = adata.uns['nsbm']['state'].copy(sampling=True)
bs = state.get_bs()
# get the graph from state
g = state.g
else:
states = [
gt.minimize_nested_blockmodel_dl(
g,
deg_corr=deg_corr,
state_args=dict(recs=recs, rec_types=rec_types),
**minimize_args) for n in range(n_init)
]
state = states[np.argmin([s.entropy() for s in states])]
# state = gt.minimize_nested_blockmodel_dl(g, deg_corr=deg_corr,
# state_args=dict(recs=recs,
# rec_types=rec_types),
# **minimize_args)
logg.info(' done', time=start)
bs = state.get_bs()
if len(bs) <= hierarchy_length:
# increase hierarchy length up to the specified value
# according to Tiago Peixoto 10 is reasonably large as number of
# groups decays exponentially
bs += [np.zeros(1)] * (hierarchy_length - len(bs))
else:
logg.warning(
f'A hierarchy length of {hierarchy_length} has been specified\n'
f'but the minimized model contains {len(bs)} levels')
pass
# create a new state with inferred blocks
state = gt.NestedBlockState(g,
bs,
state_args=dict(recs=recs,
rec_types=rec_types),
sampling=True)
# equilibrate the Markov chain
if equilibrate:
logg.info('running MCMC equilibration step')
# equlibration done by simulated annealing
equilibrate_args['wait'] = wait
equilibrate_args['nbreaks'] = nbreaks
equilibrate_args['max_niter'] = max_iterations
equilibrate_args['multiflip'] = multiflip
equilibrate_args['mcmc_args'] = {'niter': 10}
dS, nattempts, nmoves = gt.mcmc_anneal(
state,
mcmc_equilibrate_args=equilibrate_args,
niter=steps_anneal,
beta_range=beta_range)
if collect_marginals and equilibrate:
# we here only retain level_0 counts, until I can't figure out
# how to propagate correctly counts to higher levels
# I wonder if this should be placed after group definition or not
logg.info(' collecting marginals')
group_marginals = [
np.zeros(g.num_vertices() + 1) for s in state.get_levels()
]
def _collect_marginals(s):
levels = s.get_levels()
for l, sl in enumerate(levels):
group_marginals[l][sl.get_nonempty_B()] += 1
gt.mcmc_equilibrate(state,
wait=wait,
nbreaks=nbreaks,
epsilon=epsilon,
max_niter=max_iterations,
multiflip=True,
force_niter=niter_collect,
mcmc_args=dict(niter=10),
callback=_collect_marginals)
logg.info(' done', time=start)
# everything is in place, we need to fill all slots
# first build an array with
groups = np.zeros((g.num_vertices(), len(bs)), dtype=int)
for x in range(len(bs)):
# for each level, project labels to the vertex level
# so that every cell has a name. Note that at this level
# the labels are not necessarily consecutive
groups[:, x] = state.project_partition(x, 0).get_array()
groups = pd.DataFrame(groups).astype('category')
# rename categories from 0 to n
for c in groups.columns:
new_cat_names = dict([
(cx, u'%s' % cn)
for cn, cx in enumerate(groups.loc[:, c].cat.categories)
])
groups.loc[:, c].cat.rename_categories(new_cat_names, inplace=True)
if restrict_to is not None:
groups.index = adata.obs[restrict_key].index
else:
groups.index = adata.obs_names
# add column names
groups.columns = [
"%s_level_%d" % (key_added, level) for level in range(len(bs))
]
# remove any column with the same key
keep_columns = [
x for x in adata.obs.columns
if not x.startswith('%s_level_' % key_added)
]
adata.obs = adata.obs.loc[:, keep_columns]
# concatenate obs with new data, skipping level_0 which is usually
# crap. In the future it may be useful to reintegrate it
# we need it in this function anyway, to match groups with node marginals
if return_low:
adata.obs = pd.concat([adata.obs, groups], axis=1)
else:
adata.obs = pd.concat([adata.obs, groups.iloc[:, 1:]], axis=1)
# add some unstructured info
adata.uns['nsbm'] = {}
adata.uns['nsbm']['stats'] = dict(level_entropy=np.array(
[state.level_entropy(x) for x in range(len(state.levels))]),
modularity=np.array([
gt.modularity(
g, state.project_partition(x, 0))
for x in range(len((state.levels)))
]))
if equilibrate:
adata.uns['nsbm']['stats']['dS'] = dS
adata.uns['nsbm']['stats']['nattempts'] = nattempts
adata.uns['nsbm']['stats']['nmoves'] = nmoves
adata.uns['nsbm']['state'] = state
# now add marginal probabilities.
if collect_marginals:
# refrain group marginals. We collected data in vector as long as
# the number of cells, cut them into appropriate length data
adata.uns['nsbm']['group_marginals'] = {}
for nl, level_marginals in enumerate(group_marginals):
idx = np.where(level_marginals > 0)[0] + 1
adata.uns['nsbm']['group_marginals'][nl] = np.array(
level_marginals[:np.max(idx)])
# prune uninformative levels, if any
if prune:
to_remove = prune_groups(groups)
logg.info(f' Removing levels f{to_remove}')
adata.obs.drop(to_remove, axis='columns', inplace=True)
# calculate log-likelihood of cell moves over the remaining levels
# we have to calculate events at level 0 and propagate to upper levels
logg.info(' calculating cell affinity to groups')
levels = [
int(x.split('_')[-1]) for x in adata.obs.columns
if x.startswith(f'{key_added}_level')
]
adata.uns['nsbm']['cell_affinity'] = dict.fromkeys(
[str(x) for x in levels])
p0 = get_cell_loglikelihood(state, level=0, as_prob=True)
adata.uns['nsbm']['cell_affinity'][0] = p0
l0 = "%s_level_0" % key_added
for nl, level in enumerate(groups.columns[1:]):
cross_tab = pd.crosstab(groups.loc[:, l0], groups.loc[:, level])
cl = np.zeros((p0.shape[0], cross_tab.shape[1]), dtype=p0.dtype)
for x in range(cl.shape[1]):
# sum counts of level_0 groups corresponding to
# this group at current level
cl[:, x] = p0[:, np.where(cross_tab.iloc[:, x] > 0)[0]].sum(axis=1)
adata.uns['nsbm']['cell_affinity'][str(nl + 1)] = cl / np.sum(
cl, axis=1)[:, None]
# last step is recording some parameters used in this analysis
adata.uns['nsbm']['params'] = dict(
epsilon=epsilon,
wait=wait,
nbreaks=nbreaks,
equilibrate=equilibrate,
fast_model=fast_model,
collect_marginals=collect_marginals,
hierarchy_length=hierarchy_length,
random_seed=random_seed,
prune=prune,
)
logg.info(
' finished',
time=start,
deep=
(f'found {state.get_levels()[1].get_nonempty_B()} clusters at level_1, and added\n'
f' {key_added!r}, the cluster labels (adata.obs, categorical)'),
)
return adata if copy else None
def nested_model_multi(
adatas: List[AnnData],
deg_corr: bool = True,
tolerance: float = 1e-6,
n_sweep: int = 10,
beta: float = np.inf,
samples: int = 100,
collect_marginals: bool = True,
n_jobs: int = -1,
*,
random_seed: Optional[int] = None,
key_added: str = 'multi_nsbm',
adjacency: Optional[List[sparse.spmatrix]] = None,
neighbors_key: Optional[List[str]] = ['neighbors'],
directed: bool = False,
use_weights: bool = False,
save_model: Union[str, None] = None,
copy: bool = False,
# minimize_args: Optional[Dict] = {},
dispatch_backend: Optional[str] = 'processes',
# equilibrate_args: Optional[Dict] = {},
) -> Optional[List[AnnData]]:
"""\
Cluster cells into subgroups using multiple modalities.
Cluster cells using the nested Stochastic Block Model [Peixoto14]_,
performing Bayesian inference on node groups. This function takes multiple
experiments, possibly across different modalities, and perform joint
clustering.
This requires having ran :func:`~scanpy.pp.neighbors` or
:func:`~scanpy.external.pp.bbknn` first. It also requires cells having the same
names if coming from paired experiments
Parameters
----------
adatas
A list of processed AnnData. Neighbors must have been already
calculated.
deg_corr
Whether to use degree correction in the minimization step. In many
real world networks this is the case, although this doesn't seem
the case for KNN graphs used in scanpy.
tolerance
Tolerance for fast model convergence.
n_sweep
Number of iterations to be performed in the fast model MCMC greedy approach
beta
Inverse temperature for MCMC greedy approach
samples
Number of initial minimizations to be performed. The one with smaller
entropy is chosen
n_jobs
Number of parallel computations used during model initialization
key_added
`adata.obs` key under which to add the cluster labels.
adjacency
Sparse adjacency matrix of the graph, defaults to
`adata.uns['neighbors']['connectivities']` in case of scanpy<=1.4.6 or
`adata.obsp[neighbors_key][connectivity_key]` for scanpy>1.4.6
neighbors_key
The key passed to `sc.pp.neighbors`. If all AnnData share the same key, one
only has to be specified, otherwise the full tuple of all keys must
be provided
directed
Whether to treat the graph as directed or undirected.
use_weights
If `True`, edge weights from the graph are used in the computation
(placing more emphasis on stronger edges). Note that this
increases computation times
save_model
If provided, this will be the filename for the PartitionModeState to
be saved
copy
Whether to copy `adata` or modify it inplace.
random_seed
Random number to be used as seed for graph-tool
Returns
-------
`adata.obs[key_added]`
Array of dim (number of samples) that stores the subgroup id
(`'0'`, `'1'`, ...) for each cell.
`adata.uns['schist']['multi_level_params']`
A dict with the values for the parameters `resolution`, `random_state`,
and `n_iterations`.
`adata.uns['schist']['multi_level_stats']`
A dict with the values returned by mcmc_sweep
`adata.obsm['CA_multi_nsbm_level_{n}']`
A `np.ndarray` with cell probability of belonging to a specific group
`adata.uns['schist']['multi_level_state']`
The NestedBlockModel state object
"""
if random_seed:
np.random.seed(random_seed)
seeds = np.random.choice(range(samples**2), size=samples, replace=False)
if collect_marginals and samples < 100:
logg.warning(
'Collecting marginals requires sufficient number of samples\n'
f'It is now set to {samples} and should be at least 100')
start = logg.info('minimizing the nested Stochastic Block Model')
if copy:
adatas = [x.copy() for x in adatas]
n_keys = len(neighbors_key)
n_data = len(adatas)
# are we clustering a user-provided graph or the default AnnData one?
if adjacency is None:
adjacency = []
if n_keys > 1 and n_keys < n_data:
raise ValueError(
'The number of neighbors keys does not match'
'the number of data matrices. Either fix this'
'or pass a neighbor key that is shared across all modalities')
if n_keys == 1:
neighbors_key = [neighbors_key[0] for x in range(n_data)]
for x in range(n_data):
logg.info(f'getting adjacency for data {x}', time=start)
if neighbors_key[x] not in adatas[x].uns:
raise ValueError('You need to run `pp.neighbors` first '
'to compute a neighborhood graph. for'
f'data entry {x}')
elif 'connectivities_key' in adatas[x].uns[neighbors_key[x]]:
# scanpy>1.4.6 has matrix in another slot
conn_key = adatas[x].uns[
neighbors_key[x]]['connectivities_key']
adjacency.append(adatas[x].obsp[conn_key])
else:
# scanpy<=1.4.6 has sparse matrix here
adjacency.append(
adatas[x].uns[neighbors_key[x]]['connectivities'])
# convert it to igraph and graph-tool
graph_list = []
for x in range(n_data):
g = get_igraph_from_adjacency(adjacency[x], directed=directed)
g = g.to_graph_tool()
gt.remove_parallel_edges(g)
# add cell names to graph, this will be used to create
# layered graph
g_names = g.new_vertex_property('string')
d_names = adatas[x].obs_names
for xn in range(len(d_names)):
g_names[xn] = d_names[xn]
g.vp['cell'] = g_names
graph_list.append(g)
# skip weights for now
# recs=[]
# rec_types=[]
# if use_weights:
# this is not ideal to me, possibly we may need to transform
# weights. More tests needed.
# recs=[g.ep.weight]
# rec_types=['real-normal']
# get a non-redundant list of all cell names across all modalities
all_names = set(adatas[0].obs_names)
[all_names.update(adatas[x].obs_names) for x in range(1, n_data)]
all_names = list(all_names)
# create the shared graph
union_g = gt.Graph(directed=False)
union_g.add_vertex(len(all_names))
u_names = union_g.new_vertex_property('string')
for xn in range(len(all_names)):
u_names[xn] = all_names[xn]
union_g.vp['cell'] = u_names
# now handle in a non elegant way the index mapping across all
# modalities and the unified Graph
u_cell_index = dict([(union_g.vp['cell'][x], x)
for x in range(union_g.num_vertices())])
# now create layers
layer = union_g.new_edge_property('int')
for ng in range(n_data):
for e in graph_list[ng].edges():
S, T = e.source(), e.target()
Sn = graph_list[ng].vp['cell'][S]
Tn = graph_list[ng].vp['cell'][T]
Sidx = u_cell_index[Sn]
Tidx = u_cell_index[Tn]
ne = union_g.add_edge(Sidx, Tidx)
layer[ne] = ng + 1 # this is the layer label
union_g.ep['layer'] = layer
# DONE! now proceed with standard minimization, ish
if samples < 1:
samples = 1
states = [
gt.NestedBlockState(g=union_g,
base_type=gt.LayeredBlockState,
state_args=dict(deg_corr=deg_corr,
ec=union_g.ep.layer,
layers=True))
for n in range(samples)
]
def fast_min(state, beta, n_sweep, fast_tol, seed=None):
if seed:
gt.seed_rng(seed)
dS = 1
while np.abs(dS) > fast_tol:
dS, _, _ = state.multiflip_mcmc_sweep(beta=beta,
niter=n_sweep,
c=0.5)
return state
states = Parallel(n_jobs=n_jobs, prefer=dispatch_backend)(
delayed(fast_min)(states[x], beta, n_sweep, tolerance, seeds[x])
for x in range(samples))
logg.info(' minimization step done', time=start)
pmode = gt.PartitionModeState([x.get_bs() for x in states],
converge=True,
nested=True)
bs = pmode.get_max_nested()
logg.info(' consensus step done', time=start)
if save_model:
import pickle
fname = save_model
if not fname.endswith('pkl'):
fname = f'{fname}.pkl'
logg.info(f'Saving model into {fname}')
with open(fname, 'wb') as fout:
pickle.dump(pmode, fout, 2)
# prune redundant levels at the top
bs = [x for x in bs if len(np.unique(x)) > 1]
bs.append(np.array([0],
dtype=np.int32)) #in case of type changes, check this
state = gt.NestedBlockState(union_g,
bs=bs,
base_type=gt.LayeredBlockState,
state_args=dict(deg_corr=deg_corr,
ec=union_g.ep.layer,
layers=True))
logg.info(' done', time=start)
u_groups = np.unique(bs[0])
n_groups = len(u_groups)
last_group = np.max(u_groups) + 1
if collect_marginals:
# note that the size of this will be equal to the number of the groups in Mode
# but some entries won't sum to 1 as in the collection there may be differently
# sized partitions
pv_array = pmode.get_marginal(union_g).get_2d_array(
range(last_group)).T[:, u_groups] / samples
groups = np.zeros((union_g.num_vertices(), len(bs)), dtype=int)
for x in range(len(bs)):
# for each level, project labels to the vertex level
# so that every cell has a name. Note that at this level
# the labels are not necessarily consecutive
groups[:, x] = state.project_partition(x, 0).get_array()
groups = pd.DataFrame(groups).astype('category')
# rename categories from 0 to n
for c in groups.columns:
ncat = len(groups[c].cat.categories)
new_cat = [u'%s' % x for x in range(ncat)]
groups[c].cat.rename_categories(new_cat, inplace=True)
levels = groups.columns
# recode block names to have consistency with group names
i_groups = groups.astype(int)
bs = [i_groups.iloc[:, 0].values]
for x in range(1, groups.shape[1]):
bs.append(
np.where(
pd.crosstab(i_groups.iloc[:, x - 1], i_groups.iloc[:,
x]) > 0)[1])
state = gt.NestedBlockState(union_g, bs)
del (i_groups)
groups.index = all_names
# add column names
groups.columns = [f"{key_added}_level_{level}" for level in range(len(bs))]
# remove any column with the same key
for xn in range(n_data):
drop_columns = groups.columns.intersection(adatas[xn].obs.columns)
adatas[xn].obs.drop(drop_columns, 'columns', inplace=True)
adatas[xn].obs = pd.concat(
[adatas[xn].obs, groups.loc[adatas[xn].obs_names]], axis=1)
# now add marginal probabilities.
if collect_marginals:
# add marginals for level 0, the sum up according to the hierarchy
_groups = groups.loc[adatas[xn].obs_names]
_pv_array = pd.DataFrame(
pv_array, index=all_names).loc[adatas[xn].obs_names].values
adatas[xn].obsm[f"CM_{key_added}_level_0"] = _pv_array
for group in groups.columns[1:]:
ct = pd.crosstab(_groups[_groups.columns[0]],
_groups[group],
normalize='index',
dropna=False)
adatas[xn].obsm[f'CM_{group}'] = _pv_array @ ct.values
# add some unstructured info
if not 'schist' in adatas[xn].uns:
adatas[xn].uns['schist'] = {}
adatas[xn].uns['schist'][f'{key_added}'] = {}
adatas[xn].uns['schist'][f'{key_added}']['stats'] = dict(
level_entropy=np.array(
[state.level_entropy(x) for x in range(len(state.levels))]),
modularity=np.array([
gt.modularity(union_g, state.project_partition(x, 0))
for x in range(len((state.levels)))
]))
bl_d = {}
levels = state.get_levels()
for nl in range(len(levels)):
bl_d[str(nl)] = np.array(levels[nl].get_blocks().a)
adatas[xn].uns['schist'][f'{key_added}']['blocks'] = bl_d
# last step is recording some parameters used in this analysis
adatas[xn].uns['schist'][f'{key_added}']['params'] = dict(
model='multiome_nested',
use_weights=use_weights,
neighbors_key=neighbors_key[xn],
key_added=key_added,
samples=samples,
collect_marginals=collect_marginals,
random_seed=random_seed,
deg_corr=deg_corr,
# recs=recs,
# rec_types=rec_types
)
logg.info(
' finished',
time=start,
deep=(
f'and added\n'
f' {key_added!r}, the cluster labels (adata.obs, categorical)'),
)
return adatas if copy else None