My target ranking changed and I cannot find the analysis change

What actually happened

A knowledge graph is not measured. It is assembled: somebody took Open Targets, a literature-mined association set, a pathway database and an in-house curation file, and merged them into nodes and edges. What the graph holds is a function of that list.

Add one source and edges appear. Edges appearing is not a cosmetic change: a gene that had three neighbours now has five, so every quantity computed from the neighbourhood moves. Degree moves. Any propagation or diffusion over the graph moves, and moves further than the new edges, because a score that spreads along edges spreads along the new ones too. A rank is a comparison between scores, so a rank that did not move numerically still moves positionally when the gene below it did.

Drop a source and the same thing happens in reverse, with a worse reading attached: an association that is absent from the graph looks like an association the evidence does not support.

Why nothing in the file says so

Look at a node. It has an id, a type, perhaps some attributes. Look at an edge: a source, a target, a type, maybe a score. Each of those is a fact about one piece of the graph. Which sources went into the graph is a fact about the whole of it, and there is nowhere in a node or an edge for it to live.

Graphs that carry per-edge provenance are better off — the edge says it came from DisGeNET — and still do not answer the question, because the list of sources that contributed is the union over every edge, and nothing computes that union and writes it down where a comparison would see it. Two graphs sitting next to each other on disk disclose their difference only to someone who reads all of both.

Why the answer is plausible rather than absurd

Because it is arithmetically correct. The scores are right for the graph they were computed on. The ranking is the right ranking of those scores. Every intermediate value would survive any audit you could run on the computation, because the computation is not where the problem is.

And the magnitude is in the plausible range. A new source does not usually rewrite a ranking; it moves things by a few places and swaps two or three entries near the cut. That is precisely the size of effect people attribute to a methodological choice, a parameter, a different random seed, or a genuine improvement — which is why the usual response to it is to write it up.

Where it bites hardest

A benchmark. You evaluate a method against a graph built in March and a colleague evaluates theirs against one built in September. The two numbers are compared in a table as though they measured two methods. They measured two graphs.

A comparison across diseases. Two graphs assembled with different source coverage, because the curation effort differed by therapeutic area, are compared for which disease has the better-connected target landscape. The answer is about the curators.

A reproduction. Someone reruns your pipeline a year later, against today’s assembly, and gets a different answer. Nothing is wrong with either run, and there is no artefact left over to tell them that.

What Ligase says

The sources a graph was assembled from are a dimension — source_set — and the loader reads it off the graph’s own nodes and edges rather than asking you for it. Two graphs that disagree on it do not join:

Three things are worth noticing.

Neither side was declared. Both say from the sources named on its nodes and edges, naming the file: the engine read the per-edge provenance and took the union, so you can open the file it names and check.

Both source sets are printed in full, and then the member that makes them differ is named on its own line: theirs also lists disgenet. For two graphs with a dozen sources each, that line is the only part worth reading.

And the closing question is the one a user can act on — was a source added or removed upstream? — rather than a restatement that the two sets differ. It points at the rebuild, which is where the change was.

How it is caught

Load each graph. Nothing is declared about the sources, because the graphs say:

import ligase as lg
 
march = lg.knowledge_graph("graphs/2026-03.json", organism="Homo sapiens")
september = lg.knowledge_graph("graphs/2026-09.json", organism="Homo sapiens")
 
# Nothing has run. This read both files' nodes and edges.
lg.check_step("compare", {"ours": march, "theirs": september})

That refuses with the message above. There is no flag that makes it pass, because the two graphs really are different graphs, and the fix is yours to choose: reassemble both from the same source list, or stop comparing them and say in the write-up that they are not comparable.

What the engine will do is carry the dimension forward. A step over a graph inherits its source_set, so a ranking computed from the September graph carries September’s sources, and anything that tries to score it against March’s is refused the same way — three steps later, on a file that no longer mentions a source anywhere.

For what else a graph carries, see knowledge graphs and building a graph into a pipeline; for how the axis sits alongside the others, the dimension catalogue.