Filter Go terms by type of relationship

[sfujio]

Hi,
I need to calculate the similarirty between GO terms of a protein set, but I am interested only in the "is_a" type of relationship.
I am a newbie in programming and haven´t quite yet figure out how the package handles the relation between the terms. I tried to understand by seeing the code of the functions but failed.

I would like to ask 1) if filtering for type of relationship between terms is possible, and 2) how is the information of the relationships is stored and obtained from GO.

Regarding point 2, maybe the package doesn´t need to know the type of relationship but at least it needs to know how the terms are related with each other in order to calculate some similarity measueres. I would like to know how is that information retrieved and handle.

Thank you for creating the package and for your time,

SF

[GuangchuangYu]

We used GO.db to query the GO term relationships.

Only calculate semantic similarity using is_a is not supported. GOSemSim only support filter out GO terms by dropCodes.

[GuangchuangYu]

To answer the second question first, since it explains the first: GOSemSim gets the term relationships from GO.db, and the relationship type is available there — but it is dropped before the similarity code ever sees it.

GOBPPARENTS (and the MF/CC equivalents) carries the type:

df <- as.data.frame(GOBPPARENTS)
names(df)          # "go_id" "go_id" "RelationshipType"
unique(df[[3]])    # "isa" "part of" "regulates" "negatively regulates" "positively regulates"

but getParents() / getAncestors() call AnnotationDbi::as.list() on it, which keeps only the parent IDs:

l <- AnnotationDbi::as.list(GOBPPARENTS)
l[["GO:0000001"]]  # "GO:0048308" "GO:0048311"  -- no relationship type

The package's own precomputed gotbl table does retain a relationship column with the same vocabulary, and the Wang method already consumes it (R/WangMethod.R normalises isa / part of and assigns the Wang et al. 2007 weights is_a = 0.8, part_of = 0.6, other = 0.7). So the information is present in two places; it just is not exposed to the IC-based and TCSS methods, which use the ancestor closure without regard to edge type.

The 2016 answer ("Only calculate semantic similarity using is_a is not supported") is still accurate as a statement about the API. Adding it would mean:

1. keeping RelationshipType when building the parent/ancestor mappings (use as.data.frame(GOBPPARENTS) rather than as.list());
2. carrying it through getAncestors() / getOffsprings(), which currently return plain character vectors;
3. adding a filter argument (e.g. relations = "is_a") that is applied when the closure is built.

Step 2 is the awkward one, because the ancestor mappings are cached as plain lists and consumed in C++ (infoContentMethod_cpp() takes anc_ as a list of character vectors). So it is a real change rather than a one-liner. I am leaving this open as a feature request — it is feasible, but it needs a design decision about the argument name and about whether the filter applies to the ancestor closure or only to direct parents.