pnrobinson
commented
4 years ago @oleraj We are talking with Sandhya about a POET web portal to enable community annotation -- this is an example of what we need to fix. At first glance, the annotations for this disease are incomplete in our system. We are also expecting a lot of new annotations from European collaborators. At the moment, I think this is basically a limitation of the method -- it works well if the annotations are good but less so if the annotations are not. I will do an experiment of updating the annotations for this disease and if you like you could run this case through LIRICAL again. Obviously this does not fix all annotations, but it would show what could be gained. In general, Exomiser should be less sensitive to missing annotations because of its statistical model.
oleraj
And "Arthritis" and "Immunodeficiency" are now counted towards the diagnosis.
Hi Peter,
We did some more testing with LIRICAL and noticed some issues with the phenotype matching. One example is a CTLA4 patient that had this phenotype:
LIRICAL ranks CTLA4 near the top but the likelihood ratio is not very good.
The two matching phenotypes are 'Chronic diarrhea' and 'Eczema'. Some of the other phenotypes are maybe unrelated to the disorder and we noticed that they are counted against the diagnosis rather than being treated as neutral. I think that assumes the phenotypic description of disorders is perfectly complete, which is probably not the case. And it makes it tricky to run LIRICAL if any phenotype listed in our database for a patient not associated with the disorder counts against a diagnosis because it's hard to manually go through and see which ones would be directly related to the disorder. For cases with a dual diagnosis I can see potential issues as well, since a cluster of phenotypes might be related to one diagnosis and a separate cluster of phenotypes might be related to a separate diagnosis. But combining them together the unrelated phenotypes will be counted against both diagnoses so the LR would be low in both cases. I'm not sure if there's an easy way to deal with this. (As a side note, we tested two cases that had a dual diagnosis, both of which were ranked in the top 10 genes by Exomiser but the second diagnosis wasn't found in the top 10 for LIRICAL in either case.)
There are a few phenotypes that are related to the disorder that are also counted against the diagnosis, i.e., 'Arthritis' (HP:0001369), 'Immunodeficiency' (HP:0002721) and 'Arthralgia' (HP:0002829). "Autoimmune arthritis" is specifically listed in the OMIM description so it's surprising that 'Arthritis' is counted against the diagnosis. "Immunodeficiency" and "Arthralgia" aren't specifically listed in the OMIM description but immunodeficiency is implied since CTLA4 deficiency is an immunodeficiency and arthralgia is similar to arthritis. Any thoughts on this? Does LIRICAL require exact matching of phenotypes or can similar phenotypes also contribute to the score like for Exomiser?
We wondered where the annotations were from that LIRICAL is using so went to the list of phenotypes on hpo.jax website for this disorder and found that no form of Arthritis was included, which could be why it was counted against instead of in favor of the diagnosis, assuming this is what LIRICAL is using. We also noticed a lot of other annotations from OMIM that were missing from the description on hpo.jax, e.g., "Decreased levels of naive T cells" (probably should have a more specific phenotype for this, but seems like it should at least match to HP:0011839, "Abnormal T cell count"), "Decreased memory B cells" (again, should probably have a more specific phenotype in HPO for this but it seems like the closest existing might be HP:0010976, "B lymphocytopenia") , "Granulomatous lymphocytic interstitial lung disease" and "Lymphocytic infiltration of the brain" (not sure what these last two would match to in HPO). We wondered if these disease curations from OMIM were done before some of these HPO terms were added to the ontology, which is why they were missed in the curation? Or maybe they were just missed because not enough synonyms were present to pick them up?
FWIW, Exomiser was able to identify CTLA4 as the top-ranked gene for this patient, although it actually matched to a different disorder than the two that were tested in LIRICAL. Here are the disorder and matched phenotypes found in Exomiser:
Maybe LIRICAL doesn't use diseases from Orphanet?
Thanks,
Andrew