Information Content/goSim could be calculated for more GO terms?

[dorrenasun]

Dear authors,

I am trying to calculate semantic similarity between some GO terms based on the information content methods.

# GO.db retrieved by 2020-09-10, Bioconductor version 3.12
atGO<-godata('org.At.tair.db', keytype="TAIR",ont="BP")
goSim("GO:0120254", "GO:0120255", semData=atGO, measure="Jiang") # NA

The value returns NA as these GO terms are not directly annotated for Arabidopsis. However, their descendant terms were actually annotated to some genes in the database (GO.db retrieved by 2020-09-10, Bioconductor version 3.12):

length(GOBPOFFSPRING[["GO:0120254"]] %in% keys(org.At.tair.db,"GO"))  # 196
length(GOBPOFFSPRING[["GO:0120255"]] %in% keys(org.At.tair.db,"GO"))  # 97

and thus their IC values were actually feasible for calculation. Do you think it is possible to include such GO terms in the current IC & goSim() calculation?

Thank you very much.

[dorrenasun]

Dear Authors,

I figured out why is this problem: this is a conflict between the pre-calculated "gotbl.rda" in the data folder and the local version of GO.db. It seems that the function prepare_relation_df() in the source code is not executed locally during installation, so gotbl.rda remains as the same as when downloaded with the package. It would be really nice if you have any plan to fix it.

Thank you!

[GuangchuangYu]

Thank you for the report, and apologies for the delay. Your diagnosis was exactly right — this was a real bug, not a version conflict on your side.

computeIC() was deciding which terms to score from the gotbl table shipped in data/ instead of from the installed GO.db:

goids <- unique(godata[godata$Ontology == ont, "go_id"])   # gotbl, not GO.db

That table was last rebuilt in 2021-07, so any term GO.db has added since then was given no IC at all and the IC-based methods returned NA for it — which is what you saw with GO:0120254/GO:0120255.

The effect is broader than a couple of terms. Against the current GO.db and org.Hs.eg.db, the number of annotated terms that can actually be scored:

ontology	annotated	before	now	gained
BP	11293	11126	11293	+167
CC	1832	1787	1832	+45
MF	4777	4410	4777	+367

The missing terms also biased every other IC, because they were counted in the sum(gocount) denominator but could never contribute to an ancestor's descendant sum. 92–95% of the terms that were finite under both universes are unchanged; the mean absolute difference is 0.003–0.011, and about 0.3% were overestimated by more than 0.5 — all in the "too specific" direction, which is what that mechanism predicts.

Fixed in ca62608 on devel. computeIC() now takes its universe from the installed GO.db — names(getAncestors(ont)), which is exactly that ontology's term set including the root — unioned with the annotated terms, so a term the local annotation package still uses keeps its IC even if GO.db has retired it.

Your two terms, on human data (org.At.tair.db isn't installed here, but the mechanism is organism-independent):

hsGO <- godata(annoDb = "org.Hs.eg.db", ont = "BP", computeIC = TRUE)
goSim("GO:0120254", "GO:0120255", semData = hsGO, measure = "Jiang")
# before: NA
# now:    0.849

GO:0120254 = olefinic compound metabolic process, GO:0120255 = olefinic compound biosynthetic process, so 0.849 is the answer you would expect.

One related thing I found while fixing this and have not changed, so that it is on the record: the same stale gotbl is also the relation table the Wang method walks. There, a term newer than the table scores 0 rather than NA — a plausible-looking wrong answer instead of a missing one. Fixing it means either regenerating data/gotbl.rda or building the relation table from GO.db at runtime, which is a bigger decision than this fix, so I have left it separate.

[GuangchuangYu]

Fixed in ca62608 (devel): computeIC() now derives its term universe from the installed GO.db.