Skip to content

Which edges earn an arrow

A disease graph assembled from public sources mixes two kinds of edge, and only one of them is an arrow.

Some edges assert a mechanism with a direction: OmniPath's signed regulatory interactions, where a curator has established that A acts on B. Others assert only that two things occur together — a literature co-mention, membership of a shared pathway, an undirected protein interaction whose arrow exists because one protein happened to land in the first column of an API response.

Running do-calculus over the second kind assumes the very thing a causal method is supposed to establish. So NeoRx partitions the edges before identification runs, and identification runs only on the causal-admissible subgraph.

Associational edges are not discarded. They keep contributing to evidence scoring and robustness — see From verdict to target. They stop being arrows.

The rule

Two edge attributes decide it: edge_type and evidence_class.

>>> from causalphylo import CAUSAL_EDGE_TYPES
>>> sorted(CAUSAL_EDGE_TYPES)
['activates', 'causes', 'downregulates', 'inhibits', 'phosphorylates',
 'regulates', 'upregulates']

Those seven are directed and mechanistic: the source asserts which way the influence runs. An edge of one of these types is admissible on its type alone.

>>> from causalphylo import is_causal_admissible
>>> is_causal_admissible({"edge_type": "inhibits"})
True
>>> is_causal_admissible({"edge_type": "interacts_with"})
False
>>> is_causal_admissible({"edge_type": "participates_in"})
False

interacts_with is the instructive one. A STRING protein–protein interaction is real biology, and it has no direction. The arrow in the graph points whichever way the API happened to order the pair.

Gene to disease needs genetics

associated_with — the edge every association database produces — is admissible only under genetic evidence:

>>> from causalphylo import GENETIC_EVIDENCE_CLASSES
>>> sorted(GENETIC_EVIDENCE_CLASSES)
['genetic_association', 'somatic_mutation']
>>> is_causal_admissible({"edge_type": "associated_with",
...                       "evidence_class": "genetic_association"})
True
>>> is_causal_admissible({"edge_type": "associated_with",
...                       "evidence_class": "literature"})
False
>>> is_causal_admissible({"edge_type": "associated_with",
...                       "evidence_class": "known_drug"})
False

The warrant is Mendelian randomisation. A germline variant is randomised at conception, independently of the confounders that plague an observational comparison, so a genetic association carries a natural-experiment argument that a co-mention does not (Davey Smith & Ebrahim, 2003, Int J Epidemiol 32:1–22). That argument is what earns the arrow.

known_drug evidence is excluded for a different reason worth stating: it is the strongest-looking evidence class in Open Targets and it is circular here. A gene is linked to a disease by known_drug because someone already developed a drug for it. Admitting that edge would let the pipeline rediscover approved targets and call it a prediction.

Taking the subgraph

causal_subgraph keeps the admissible edges with their attributes and drops nodes left with none — a node participating in no causal relation cannot be a treatment, an outcome, or a confounder.

>>> import networkx as nx
>>> from causalphylo import causal_subgraph
>>> G = nx.DiGraph()
>>> G.add_edge("CCL5", "CCR5", edge_type="regulates",
...            evidence_class="regulatory")
>>> G.add_edge("CCR5", "HIV", edge_type="associated_with",
...            evidence_class="genetic_association")
>>> G.add_edge("TNF", "HIV", edge_type="associated_with",
...            evidence_class="literature")
>>> G.add_edge("TNF", "IL6", edge_type="interacts_with",
...            evidence_class="ppi")
>>> sub = causal_subgraph(G)
>>> sorted(sub.nodes)
['CCL5', 'CCR5', 'HIV']
>>> sorted(sub.edges)
[('CCL5', 'CCR5'), ('CCR5', 'HIV')]

TNF and IL6 are gone from the subgraph entirely. That is the mechanism behind the no_causal_path verdict in the worked example: once restricted, TNF has no route to the outcome.

Cycles

Biology has feedback loops, and a loop is not a DAG. Rather than silently breaking cycles — which would mean choosing a direction the data does not support — the admissible subgraph is reduced to its acyclic core, and what was removed is reported.

>>> from causalphylo import cyclic_components, acyclic_core
>>> H = nx.DiGraph()
>>> H.add_edge("A", "B", edge_type="activates")
>>> H.add_edge("B", "A", edge_type="inhibits")
>>> H.add_edge("B", "C", edge_type="activates")
>>> [sorted(component) for component in cyclic_components(H)]
[['A', 'B']]
>>> core = acyclic_core(H)
>>> sorted(core.dag.nodes)
['C']
>>> sorted(core.excluded)
['A', 'B']

A node inside a cycle is not a candidate target in that run, and the verdict for it says cyclic_component rather than reporting a result computed on a graph that was quietly rewritten. See Identification.

Widening this is a decision, not a tweak

CAUSAL_EDGE_TYPES and GENETIC_EVIDENCE_CLASSES are the whole interface between "what the sources said" and "what the causal engine may reason over". Adding a name to either silently widens what the backdoor criterion is allowed to treat as a mechanism.

The same frozensets appear in the snapshot layer as GENETIC_DATATYPES, for the same reason — the candidate frame for a dated run is drawn only from evidence types that confer admissibility. See Versioned snapshots.

This part is also a standalone package

Admissibility and identification have no biomedical model of their own: they read two edge attributes and graph structure. They have been extracted into CausalPhylo, which does this over any NetworkX graph with one runtime dependency. If identification over your own graph is what you want, depend on that rather than on NeoRx.

NeoRx depends on CausalPhylo rather than keeping a second copy, which is why every import on this page is from causalphylo import .... The admissibility vocabulary is still a decision with biomedical content — see the note above on widening it — but the rules that read it are not.