# -*- coding: utf-8 -*-
"""This module describes a heat diffusion workflow for analyzing BEL networks with differential gene expression [0]_.
It has four parts:
1) Assembling a network, pre-processing, and overlaying data
2) Generating unbiased candidate mechanisms from the network
3) Generating random subgraphs from each unbiased candidate mechanism
4) Applying standard heat diffusion to each subgraph and calculating scores for each unbiased candidate mechanism based
on the distribution of scores for its subgraph
In this algorithm, heat is applied to the nodes based on the data set. For the differential gene expression experiment,
the log-fold-change values are used instead of the corrected p-values to allow for the effects of up- and
down-regulation to be admitted in the analysis. Finally, heat diffusion inspired by previous algorithms publsihed in
systems and networks biology [1]_ [2]_ is run with the constraint that decreases
edges cause the sign of the heat to be flipped. Because of the construction of unbiased candidate mechanisms, all
heat will flow towards their seed biological process nodes. The amount of heat on the biological process node after
heat diffusion stops becomes the score for the whole unbiased candidate mechanism.
The issue of inconsistent causal networks addressed by SST [3]_ does not affect heat diffusion algorithms
since it can quantify multiple conflicting pathways. However, it does not address the possibility of contradictory
edges, for example, when ``A increases B`` and ``A decreases B`` are both true. A random sampling approach is used on
networks with contradictory edges and aggregate statistics over multiple trials are used to assess the robustness of the
scores as a function of the topology of the underlying unbiases candidate mechanisms.
Invariants
~~~~~~~~~~
- Because heat always flows towards the biological process node, it is possible to remove leaf nodes (nodes with no
incoming edges) after each step, since their heat will never change.
Examples
~~~~~~~~
This workflow has been applied in several Jupyter notebooks:
- `Heat Diffusion Workflow <http://nbviewer.jupyter.org/github/pybel/pybel-notebooks/blob/master/algorithms/Heat%20Diffusion%20Workflow.ipynb>`_
- `Time Series Heat Diffusion <http://nbviewer.jupyter.org/github/pybel/pybel-notebooks/blob/master/algorithms/Time%20Series%20CMPA.ipynb>`_
Future Work
~~~~~~~~~~~
This algorithm can be tuned to allow the use of correlative relationships. Because many multi-scale and multi-modal
data are often measured with correlations to molecular features, this enables experiments to be run using SNP or
brain imaging features, whose experiments often measure their correlation with the activity of gene products.
.. [0] Hoyt, C. T., Konotopez, A., Ebeling, C., & Wren, J. (2018). `PyBEL: a computational framework for Biological
Expression Language <https://doi.org/10.1093/bioinformatics/btx660>`_. Bioinformatics (Oxford, England), 34(4),
703–704.
.. [1] Bernabò N., *et al.* (2014). `The biological networks in studying cell signal transduction complexity: The
examples of sperm capacitation and of endocannabinoid system
<https://www.sciencedirect.com/science/article/pii/S2001037014000282?via%3Dihub>`_. Computational and Structural
Biotechnology Journal, 11 (18), 11–21.
.. [2] Leiserson, M. D. M., *et al.* (2015). `Pan-cancer network analysis identifies combinations of rare somatic
mutations across pathways and protein complexes <https://www.nature.com/articles/ng.3168>`_. Nature Genetics,
47 (2), 106–14.
.. [3] Vasilyev, D. M., *et al.* (2014). `An algorithm for score aggregation over causal biological networks based on
random walk sampling <https://doi.org/10.1186/1756-0500-7-516>`_. BMC Research Notes, 7, 516.
"""
from collections import defaultdict
import logging
import numpy as np
import random
from operator import itemgetter
from scipy import stats
from tqdm import tqdm
from pybel.constants import BIOPROCESS, CAUSAL_DECREASE_RELATIONS, CAUSAL_INCREASE_RELATIONS, RELATION
from pybel.struct.filters.node_selection import get_nodes_by_function
from pybel.struct.grouping import get_subgraphs_by_annotation
from ..constants import WEIGHT
from ..generation import generate_bioprocess_mechanisms, generate_mechanism
__all__ = [
'RESULT_LABELS',
'calculate_average_scores_on_graph',
'calculate_average_scores_on_subgraphs',
'workflow',
'multirun',
'workflow_aggregate',
'workflow_all',
'workflow_all_aggregate',
'calculate_average_score_by_annotation',
'Runner',
]
log = logging.getLogger(__name__)
#: The key for the heat diffusion score in the node's data dictionary
SCORE = 'score'
#: The default score
DEFAULT_SCORE = 0
#: The columns in the score tuples
RESULT_LABELS = [
'avg',
'stddev',
'normality',
'median',
'neighbors',
'subgraph_size',
]
[docs]def calculate_average_scores_on_graph(graph, key=None, tag=None, default_score=None, runs=None, use_tqdm=False):
"""Calculates the scores over all biological processes in the sub-graph.
As an implementation, it simply computes the sub-graphs then calls :func:`calculate_average_scores_on_subgraphs` as
described in that function's documentation.
:param pybel.BELGraph graph: A BEL graph with heats already on the nodes
:param Optional[str] key: The key in the node data dictionary representing the experimental data. Defaults to
:data:`pybel_tools.constants.WEIGHT`.
:param Optional[str] tag: The key for the nodes' data dictionaries where the scores will be put. Defaults to 'score'
:param Optional[float] default_score: The initial score for all nodes. This number can go up or down.
:param Optional[int] runs: The number of times to run the heat diffusion workflow. Defaults to 100.
:param bool use_tqdm: Should there be a progress bar for runners?
:return: A dictionary of {pybel node tuple: results tuple}
:rtype: dict[tuple, tuple]
Suggested usage with :mod:`pandas`:
>>> import pandas as pd
>>> from pybel_tools.analysis.heat import calculate_average_scores_on_graph
>>> graph = ... # load graph and data
>>> scores = calculate_average_scores_on_graph(graph)
>>> pd.DataFrame.from_items(scores.items(), orient='index', columns=RESULT_LABELS)
"""
subgraphs = generate_bioprocess_mechanisms(graph, key=key)
scores = calculate_average_scores_on_subgraphs(
subgraphs,
key=key,
tag=tag,
default_score=default_score,
runs=runs,
use_tqdm=use_tqdm
)
return scores
[docs]def calculate_average_scores_on_subgraphs(subgraphs, key=None, tag=None, default_score=None, runs=None, use_tqdm=False):
"""Calculate the scores over precomputed candidate mechanisms.
:param dict[tuple,pybel.BELGraph] subgraphs: A dictionary of {tuple node: pybel.BELGraph candidate mechanism}
:param Optional[str] key: The key in the node data dictionary representing the experimental data. Defaults to
:data:`pybel_tools.constants.WEIGHT`.
:param Optional[str] tag: The key for the nodes' data dictionaries where the scores will be put. Defaults to 'score'
:param Optional[float] default_score: The initial score for all nodes. This number can go up or down.
:param Optional[int] runs: The number of times to run the heat diffusion workflow. Defaults to 100.
:param bool use_tqdm: Should there be a progress bar for runners?
:return: A dictionary of {pybel node tuple: results tuple}
:rtype: dict[tuple, tuple]
Example Usage:
>>> import pandas as pd
>>> from pybel_tools.generation import generate_bioprocess_mechanisms
>>> from pybel_tools.analysis.heat import calculate_average_scores_on_subgraphs
>>> # load graph and data
>>> graph = ...
>>> candidate_mechanisms = generate_bioprocess_mechanisms(graph)
>>> scores = calculate_average_scores_on_subgraphs(candidate_mechanisms)
>>> pd.DataFrame.from_items(scores.items(), orient='index', columns=RESULT_LABELS)
"""
results = {}
log.info('calculating results for %d candidate mechanisms using %d permutations', len(subgraphs), runs)
it = subgraphs.items()
if use_tqdm:
it = tqdm(it, total=len(subgraphs), desc='Candidate mechanisms')
for node, subgraph in it:
number_first_neighbors = subgraph.in_degree(node)
number_first_neighbors = 0 if isinstance(number_first_neighbors, dict) else number_first_neighbors
mechanism_size = subgraph.number_of_nodes()
runners = workflow(subgraph, node, key=key, tag=tag, default_score=default_score, runs=runs)
scores = [runner.get_final_score() for runner in runners]
if 0 == len(scores):
results[node] = (
None,
None,
None,
None,
number_first_neighbors,
mechanism_size,
)
continue
scores = np.array(scores)
average_score = np.average(scores)
score_std = np.std(scores)
med_score = np.median(scores)
chi_2_stat, norm_p = stats.normaltest(scores)
results[node] = (
average_score,
score_std,
norm_p,
med_score,
number_first_neighbors,
mechanism_size,
)
return results
[docs]def workflow(graph, node, key=None, tag=None, default_score=None, runs=None, minimum_nodes=1):
"""Generate candidate mechanisms and run the heat diffusion workflow.
:param pybel.BELGraph graph: A BEL graph
:param tuple node: The BEL node that is the focus of this analysis
:param Optional[str] key: The key in the node data dictionary representing the experimental data. Defaults to
:data:`pybel_tools.constants.WEIGHT`.
:param Optional[str] tag: The key for the nodes' data dictionaries where the scores will be put. Defaults to 'score'
:param Optional[float] default_score: The initial score for all nodes. This number can go up or down.
:param Optional[int] runs: The number of times to run the heat diffusion workflow. Defaults to 100.
:param int minimum_nodes: The minimum number of nodes a sub-graph needs to try running heat diffusion
:return: A list of runners
:rtype: list[Runner]
"""
subgraph = generate_mechanism(graph, node, key=key)
if subgraph.number_of_nodes() <= minimum_nodes:
return []
runners = multirun(subgraph, node, key=key, tag=tag, default_score=default_score, runs=runs)
return list(runners)
[docs]def multirun(graph, node, key=None, tag=None, default_score=None, runs=None, use_tqdm=False):
"""Run the heat diffusion workflow multiple times, each time yielding a :class:`Runner` object upon completion.
:param pybel.BELGraph graph: A BEL graph
:param tuple node: The BEL node that is the focus of this analysis
:param Optional[str] key: The key in the node data dictionary representing the experimental data. Defaults to
:data:`pybel_tools.constants.WEIGHT`.
:param str tag: The key for the nodes' data dictionaries where the scores will be put. Defaults to 'score'
:param float default_score: The initial score for all nodes. This number can go up or down.
:param int runs: The number of times to run the heat diffusion workflow. Defaults to 100.
:param bool use_tqdm: Should there be a progress bar for runners?
:return: An iterable over the runners after each iteration
:rtype: iter[Runner]
"""
if runs is None:
runs = 100
it = range(runs)
if use_tqdm:
it = tqdm(it, total=runs)
for i in it:
try:
runner = Runner(graph, node, key=key, tag=tag, default_score=default_score)
runner.run()
yield runner
except Exception:
log.debug('Run %s failed for %s', i, node)
[docs]class Runner:
"""This class houses the data related to a single run of the heat diffusion workflow."""
def __init__(self, graph, target_node, key=None, tag=None, default_score=None):
"""Initializes the heat diffusion runner class
:param pybel.BELGraph graph: A BEL graph
:param tuple target_node: The BEL node that is the focus of this analysis
:param Optional[str] key: The key in the node data dictionary representing the experimental data. Defaults to
:data:`pybel_tools.constants.WEIGHT`.
:param str tag: The key for the nodes' data dictionaries where the scores will be put. Defaults to 'score'
:param float default_score: The initial score for all nodes. This number can go up or down.
"""
self.graph = graph.copy()
self.target_node = target_node
self.key = key or WEIGHT
self.default_score = default_score or DEFAULT_SCORE
self.tag = tag or SCORE
for node, data in self.graph.iter_node_data_pairs():
if not self.graph.predecessors(node):
self.graph.node[node][self.tag] = data.get(self.key, 0)
log.log(5, 'initializing %s with %s', target_node, self.graph.node[node][self.tag])
[docs] def iter_leaves(self):
"""Returns an iterable over all nodes that are leaves. A node is a leaf if either:
- it doesn't have any predecessors, OR
- all of its predecessors have a score in their data dictionaries
:return: An iterable over all leaf nodes
:rtype: iter
"""
for node in self.graph:
if self.tag in self.graph.node[node]:
continue
if not any(self.tag not in self.graph.node[p] for p in self.graph.predecessors_iter(node)):
yield node
[docs] def has_leaves(self):
"""Returns if the current graph has any leaves.
Implementation is not that smart currently, and does a full sweep.
:return: Does the current graph have any leaves?
:rtype: bool
"""
leaves = list(self.iter_leaves())
return leaves
[docs] def in_out_ratio(self, node):
"""Calculates the ratio of in-degree / out-degree of a node
:param tuple node: A BEL node
:return: The in-degree / out-degree ratio for the given node
:rtype: float
"""
return self.graph.in_degree(node) / float(self.graph.out_degree(node))
[docs] def unscored_nodes_iter(self):
"""Iterates over all nodes without a score"""
for node, data in self.graph.iter_node_data_pairs():
if self.tag not in data:
yield node
[docs] def get_random_edge(self):
"""This function should be run when there are no leaves, but there are still unscored nodes. It will introduce
a probabilistic element to the algorithm, where some edges are disregarded randomly to eventually get a score
for the network. This means that the score can be averaged over many runs for a given graph, and a better
data structure will have to be later developed that doesn't destroy the graph (instead, annotates which edges
have been disregarded, later)
1. get all unscored
2. rank by in-degree
3. weighted probability over all in-edges where lower in-degree means higher probability
4. pick randomly which edge
:return: A random in-edge to the lowest in/out degree ratio node. This is a 3-tuple of (node, node, key)
:rtype: tuple
"""
nodes = [
(n, self.in_out_ratio(n))
for n in self.unscored_nodes_iter()
if n != self.target_node
]
node, deg = min(nodes, key=itemgetter(1))
log.log(5, 'checking %s (in/out ratio: %.3f)', node, deg)
possible_edges = self.graph.in_edges(node, keys=True)
log.log(5, 'possible edges: %s', possible_edges)
edge_to_remove = random.choice(possible_edges)
log.log(5, 'chose: %s', edge_to_remove)
return edge_to_remove
[docs] def remove_random_edge(self):
"""Remove a random in-edge from the node with the lowest in/out degree ratio."""
u, v, k = self.get_random_edge()
log.log(5, 'removing %s, %s (%s)', u, v, k)
self.graph.remove_edge(u, v, k)
[docs] def remove_random_edge_until_has_leaves(self):
"""Remove random edges until there is at least one leaf node."""
while True:
leaves = set(self.iter_leaves())
if leaves:
return
self.remove_random_edge()
[docs] def score_leaves(self):
"""Calculate the score for all leaves.
:return: The set of leaf nodes that were scored
:rtype: set
"""
leaves = set(self.iter_leaves())
if not leaves:
log.warning('no leaves.')
return set()
for leaf in leaves:
self.graph.node[leaf][self.tag] = self.calculate_score(leaf)
log.log(5, 'chomping %s', leaf)
return leaves
[docs] def run(self):
"""Calculate scores for all leaves until there are none, removes edges until there are, and repeats until
all nodes have been scored.
"""
while not self.done_chomping():
self.remove_random_edge_until_has_leaves()
self.score_leaves()
[docs] def done_chomping(self):
"""Determines if the algorithm is complete by checking if the target node of this analysis has been scored
yet. Because the algorithm removes edges when it gets stuck until it is un-stuck, it is always guaranteed to
finish.
:return: Is the algorithm done running?
:rtype: bool
"""
return self.tag in self.graph.node[self.target_node]
[docs] def get_final_score(self):
"""Return the final score for the target node.
:return: The final score for the target node
:rtype: float
"""
if not self.done_chomping():
raise ValueError('algorithm has not yet completed')
return self.graph.node[self.target_node][self.tag]
[docs] def calculate_score(self, node):
"""Calculate the score of the given node.
:param tuple node: A node in the BEL graph
:return: The new score of the node
:rtype: float
"""
score = self.graph.node[node][self.tag] if self.tag in self.graph.node[node] else self.default_score
for predecessor, _, d in self.graph.in_edges_iter(node, data=True):
if d[RELATION] in CAUSAL_INCREASE_RELATIONS:
score += self.graph.node[predecessor][self.tag]
elif d[RELATION] in CAUSAL_DECREASE_RELATIONS:
score -= self.graph.node[predecessor][self.tag]
return score
[docs] def get_remaining_graph(self):
"""Allows for introspection on the algorithm at a given point by returning the sub-graph induced
by all unscored nodes
:return: The remaining unscored BEL graph
:rtype: pybel.BELGraph
"""
return self.graph.subgraph(self.unscored_nodes_iter())
[docs]def workflow_aggregate(graph, node, key=None, tag=None, default_score=None, runs=None, aggregator=None):
"""Get the average score over multiple runs.
This function is very simple, and can be copied to do more interesting statistics over the :class:`Runner`
instances. To iterate over the runners themselves, see :func:`workflow`
:param pybel.BELGraph graph: A BEL graph
:param tuple node: The BEL node that is the focus of this analysis
:param Optional[str] key: The key in the node data dictionary representing the experimental data. Defaults to
:data:`pybel_tools.constants.WEIGHT`.
:param Optional[str] tag: The key for the nodes' data dictionaries where the scores will be put. Defaults to 'score'
:param Optional[float] default_score: The initial score for all nodes. This number can go up or down.
:param Optional[int] runs: The number of times to run the heat diffusion workflow. Defaults to 100.
:param aggregator: A function that aggregates a list of scores. Defaults to :func:`numpy.average`.
Could also use: :func:`numpy.mean`, :func:`numpy.median`, :func:`numpy.min`, :func:`numpy.max`
:type aggregator: Optional[list[float] -> float]
:return: The average score for the target node
:rtype: float
"""
runners = workflow(graph, node, key=key, tag=tag, default_score=default_score, runs=runs)
scores = [runner.get_final_score() for runner in runners]
if not scores:
log.warning('Unable to run the heat diffusion workflow for %s', node)
return None
if aggregator is None:
return np.average(scores)
return aggregator(scores)
[docs]def workflow_all(graph, key=None, tag=None, default_score=None, runs=None):
"""Run the heat diffusion workflow and get runners for every possible candidate mechanism
1. Get all biological processes
2. Get candidate mechanism induced two level back from each biological process
3. Heat diffusion workflow for each candidate mechanism for multiple runs
4. Return all runner results
:param pybel.BELGraph graph: A BEL graph
:param Optional[str] key: The key in the node data dictionary representing the experimental data. Defaults to
:data:`pybel_tools.constants.WEIGHT`.
:param str tag: The key for the nodes' data dictionaries where the scores will be put. Defaults to 'score'
:param float default_score: The initial score for all nodes. This number can go up or down.
:param int runs: The number of times to run the heat diffusion workflow. Defaults to 100.
:return: A dictionary of {node: list of runners}
:rtype: dict[tuple,list[Runner]]
"""
results = {}
for node in get_nodes_by_function(graph, BIOPROCESS):
results[node] = workflow(graph, node, key=key, tag=tag, default_score=default_score, runs=runs)
return results
[docs]def workflow_all_aggregate(graph, key=None, tag=None, default_score=None, runs=None, aggregator=None):
"""Run the heat diffusion workflow to get average score for every possible candidate mechanism.
1. Get all biological processes
2. Get candidate mechanism induced two level back from each biological process
3. Heat diffusion workflow on each candidate mechanism for multiple runs
4. Report average scores for each candidate mechanism
:param pybel.BELGraph graph: A BEL graph
:param Optional[str] key: The key in the node data dictionary representing the experimental data. Defaults to
:data:`pybel_tools.constants.WEIGHT`.
:param Optional[str] tag: The key for the nodes' data dictionaries where the scores will be put. Defaults to 'score'
:param Optional[float] default_score: The initial score for all nodes. This number can go up or down.
:param Optional[int] runs: The number of times to run the heat diffusion workflow. Defaults to 100.
:param aggregator: A function that aggregates a list of scores. Defaults to :func:`numpy.average`.
Could also use: :func:`numpy.mean`, :func:`numpy.median`, :func:`numpy.min`, :func:`numpy.max`
:type aggregator: Optional[list[float] -> float]
:return: A dictionary of {node: upstream causal subgraph}
:rtype: dict
"""
results = {}
bioprocess_nodes = list(get_nodes_by_function(graph, BIOPROCESS))
for bioprocess_node in tqdm(bioprocess_nodes, total=len(bioprocess_nodes)):
subgraph = generate_mechanism(graph, bioprocess_node, key=key)
try:
results[bioprocess_node] = workflow_aggregate(
graph=subgraph,
node=bioprocess_node,
key=key,
tag=tag,
default_score=default_score,
runs=runs,
aggregator=aggregator
)
except Exception:
log.exception('could not run on %', bioprocess_node)
return results
# TODO reinvestigate statistical bootstrapping/resampling/distribution normalization
[docs]def calculate_average_score_by_annotation(graph, annotation, key=None, runs=None, use_tdqm=False):
"""For each sub-graph induced over the edges matching the annotation, calculate the average score
for all of the contained biological processes
Assumes you haven't done anything yet
1. Generates biological process upstream candidate mechanistic sub-graphs with :func:`generate_bioprocess_mechanisms`
2. Calculates scores for each sub-graph with :func:`calculate_average_scores_on_sub-graphs`
3. Overlays data with pbt.integration.overlay_data
4. Calculates averages with pbt.selection.group_nodes.average_node_annotation
:param pybel.BELGraph graph: A BEL graph
:param str annotation: A BEL annotation
:param Optional[str] key: The key in the node data dictionary representing the experimental data. Defaults to
:data:`pybel_tools.constants.WEIGHT`.
:param Optional[int] runs: The number of times to run the heat diffusion workflow. Defaults to 100.
:param bool use_tqdm: Should there be a progress bar for runners?
:return: A dictionary from {str annotation value: tuple scores}
:rtype: dict[str,tuple]
Example Usage:
>>> import pybel
>>> from pybel_tools.integration import overlay_data
>>> from pybel_tools.analysis.heat import calculate_average_score_by_annotation
>>> graph = pybel.from_path(...)
>>> scores = calculate_average_score_by_annotation(graph, 'subgraph')
"""
candidate_mechanisms = generate_bioprocess_mechanisms(graph, key=key)
#: {bp tuple: list of scores}
scores = calculate_average_scores_on_subgraphs(
subgraphs=candidate_mechanisms,
key=key,
runs=runs,
use_tqdm=use_tdqm,
)
subgraphs = get_subgraphs_by_annotation(graph, annotation)
subgraph_bp = defaultdict(list)
for annotation_value, subgraph in subgraphs.items():
subgraph_bp[annotation_value].extend(get_nodes_by_function(subgraph, BIOPROCESS))
final_scores = {}
for annotation_value, bps in subgraph_bp.items():
#: Pick the average by slicing with 0. Refer to :func:`calculate_average_score_on_subgraphs`
bp_scores = [scores[bp][0] for bp in bps]
final_scores[annotation_value] = np.average(bp_scores)
return scores