Natural language processing enables efficient deep phenotyping for clinical genome sequencing analysis

ACMG 2026

Mar 10, 2026
Presentation

Franklin Lewis, Patrick McMullen, Pamela Rathbun, Anthony Wong, Kai Lee Yap, Jessica Stoll, Andrew Drackley, Alexander Ing

Introduction
Laboratories employ two primary prioritization workflows to analyze exome (ES) and genome sequencing (GS) data: genotype-driven analysis (GA) and phenotype-driven analysis (PA). GA flags variants suggestive of pathogenicity, such as predicted impact to protein function or de novo inheritance. PA ranks variants based on the gene’s putative relevance to a patient’s clinical presentation, represented by Human Phenotype Ontology (HPO) terms. The HPO database standardizes clinical terminology for gene/variant prioritization, but requires extraction of phenotypes from clinical notes and conversion to HPO terms. Manual phenotype construction is tedious, taking 29 minutes on average. This has led to exploration of natural language processing (NLP) as a potential alternative.

 

Methods
Sixty-seven patients with previously completed, diagnostic, clinical GS tests with monogenic etiologies were collected. Manual phenotype extraction for each case was completed based on notes from a clinical genetics provider and a consulting specialist written before the return of GS results. Notes were manually reviewed by genetic counselors, and HPO terms were assigned to all extracted patient phenotype descriptors. A rule-based NLP program performed phenotype extraction and HPO term conversion from the same notes. The NLP-derived list was manually curated to remove obvious errors. Three total term lists were produced: NLP-only, NLP+Curation, and Manual. Descriptive statistics were calculated for each. Term similarity and the effects of term curation were measured using a Jaccard index. The time spent on all manual processes was recorded.

 

The selected HPO terms from each list were used as inputs for Phen2Gene, a publicly available gene prioritization tool, with ranked gene lists as the output for each. For each phenotyping method, the causative gene’s median ranking and frequency with which it was ranked within the top 10, 50, 100, 500, 1000, or 5000 genes by Phen2Gene. Based on the causative variant[s] attributes, each case was categorized as requiring Genotype-Only (GO), Phenotype-Needed (PN), Phenotype-Possibly-Needed (PP) in the context of identifying the variant as diagnostic.

 

Results
Overall, NLP extracted significantly more terms on average (34) than did Manual extraction (31) prior to curation (p-value < 0.001), but removing erroneous terms from the NLP sets equated the average terms extracted. The average Jaccard index was 0.30 between Manual and NLP-only, and 0.32 between Manual and NLP+Curation, indicating an approximately 30% exact term overlap between methods. However, outcomes-based measures of gene ranking demonstrated similar abilities between phenotyping methods to place the causative gene top 10 or 50 on the respective generated gene lists, and a statistically significant difference was not observed for cases where the causative gene fell outside the top 5000. The median causative gene ranking position based on manual term extraction was lower (more optimal) than both versions of NLP extraction, though this difference was also not statistically significant.

 

The average time to curate an NLP list for a case was 3 minutes, significantly faster than the average manual phenotype extraction (29 minutes).

 

Fifty-five percent (37/67) of cases were categorized as requiring some element of phenotype-driven analysis (PN or PP) to isolate the causative variant, reinforcing the importance of accurate phenotyping.

 

Conclusion
Using this methodology, NLP had a similar ability to extract conceptually consistent HPO terms from clinical notes compared to a trained manual reviewer. Several recurrent NLP errors were noted, often due to incorrect semantic context. Manual review and curation of the NLP-derived terms improved the similarity between the NLP-only and Manual lists, suggesting that human review may alleviate concerns about NLP for ES/GS phenotype extraction. The time-saving potential of NLP for this phenotype extraction, with potentially comparable quality to a human reviewer, enables more capacity for human interpretation, and supports the utility of its potential implementation.