Pharmacogenetic analysis using artificial intelligence (AI) to identify polymorphisms associated with sub-optimal ovarian response and hyper-response.

Abstract

Purpose: To identify genetic variants associated with an increased likelihood of sub-optimal ovarian response or hyper-response by machine learning.

Methods: This retrospective observational study, conducted between March 2018 and April 2022, analyses 495 ovarian stimulations in oocyte donors. Only each donor's first ovarian stimulation was considered. The egg donors were healthy women aged 18 to 35 years. Donor characteristics and ovarian stimulation data were recorded, as well as genotypes of 31 polymorphisms previously identified as modulators of ovarian response. Models to predict the type of ovarian response (sub-optimal, normal, or hyper-response) were performed using 5 different classification machine-learning algorithms. The most important variables were determined by SHAP (Shapley-Additive-exPlanations) values.

Results: Despite being young with good ovarian reserves and using similar stimulation protocols, 15.15% of oocyte donors had a sub-optimal response (4-9 oocytes), while 27.27% showed a hyper-response (over 20 oocytes). The best predictive model was random forest, with an AUC of 0.822. Six significant genetic polymorphisms were identified: three in hormone receptors-oestrogen receptor (ESR2; c.*39G > A, c.984G > A), follicle-stimulating hormone receptor (FSHR; p.Asn680Ser, c.-29G > A), and AMH receptor (AMHR2; c.622-6C > T) and one in growth differentiation factor 9 (GDF9; c.398-39G > C). Four polymorphisms (ESR2, FSHR) were linked to sub-optimal response, while two (AMHR2, GDF9) were associated with hyper-response.

Conclusions: By using a predictive model to asses ovarian response, we identified six genetic polymorphisms associated with ovarian response. Women who carry these genetic variants may be suitable candidates for personalised ovarian stimulation treatments to help prevent inadequate responses.