In our germline WES pipeline, we use Exomiser for phenotype-driven variant prioritization. Phenotype input from clinicians is inconsistent, some cases come with 10+ documented features, others with just a primary diagnosis. We expand to 2-3 HPO terms. Anecdotally, sparse phenotype sets push the causal variant further down the ranked list, but I haven't found a clean benchmark on this. Specific questions:
- Has anyone measured how top-N recall degrades as a function of HPO term count? Any published numbers or in-house benchmarks you can share?
- Is there a practical minimum, e.g., below 5 HPO terms, prioritization is essentially noise?
- For clinical-notes-to-HPO extraction, what's working for you? We've tested Doc2HPO and ClinPhen, considering Phenotagger or an LLM-based approach.
Thanks in advance.
1 answer
A great point, Sameer! I would rather go for a simple phenotype0to-genotype correlation and jot down to the list of those variants that match the phenotype Although ClinPhen is a good one, matching with MESH/OMIM terms is a good choice!
If I were to assume you work with rare diseases, the HPO terms can also be matched with OephanDB/Genecards
Prash
Log in to answer this question.
what is the clinical context here? are these highly penetrant phenotypes in pediatric patient exomes, or?
on average how much clinical context do you have? a req form or full EHR access, or ?
Penetrance varies. also by disease. and we do have clinical notes and full EHR access