Phosphorus Research: Challenges Associated with Increasing the Predictive Power of AMH in Controlled Ovarian Stimulation
Last week, Phosphorus attended the American Society of Human Genetics (ASHG) Annual Meeting in Orlando, Florida. The ASHG Annual Meeting is an important nexus for professionals in the field, where experts exchange new ideas and discuss recent developments. Phosphorus was proud to present its scientific poster, entitled “Challenges Associated with Increasing the Predictive Power of AMH in Controlled Ovarian Stimulation” on Thursday. The study on which the poster is based looked at genetic factors that may contribute to the etiology of infertility diseases. Select text from the study is below:
Women between the ages of 18 and 42 undergoing controlled ovarian stimulation, or COS (N=395), from five fertility centers were retrospectively enrolled in the study. Patients with a history of polycystic ovarian syndrome, cancer, chemotherapy, or bone marrow transplant were excluded from analysis. Response categories were defined as “low response” (<5 oocytes retrieved) and “high response” (>20 oocytes retrieved). A single genotype association analysis (N=217) was performed by logistic regression models, with anti-Mullerian hormone (AMH) and individual genotypes as the dependent variables. Multiple test correction was calculated from 216,615 computations (.005<p<.008).
We attempted to create a polygenic predictive model. Samples were split into a discovery set (70%) and a validation set (30%). We used the discovery set to develop ranked lists of genotypes by their predictive capacity. We did this in three ways, with two variations in the first two methods: Anneal 1 and 2, Rank 1 and 2, and for a negative control, we picked a random order. We then trained logistic regression models on the discovery set using these ranked lists and measured model performance on the validation set using the area under the ROC curve. We validated based off of the leave-one-out cross validation method.
Averages of participants’ clinical characteristics reflected the general population seeking fertility treatment. Single genotype analysis yielded no significant results after multiple test correction (p>5.7×103).
When the model was trained on the discovery set of samples, AMH was a better predictor of low response compared to AMH with the addition of genotypes (0.75 vs. 0.70; p<1.0×10-6; Figure 1). Similarly, AMH was a better predictor of high response than AMH with the addition of genotypes (0.80 vs. 0.70; p<1.0×10-6). In the polygenic predictive model, genotypes failed to add predictive value to AMH alone.
Though this small pilot study shows that genetics may not be a strong predicator of fertility outcomes, this does not discount that genetics could be involved in the etiology of fertility diseases. One limitation in this study was the samples size. Future studies should be comprised of large cohorts that have sequencing, as well as extensive phenotypic data that will help account for heterogeneous phenotypes. A major issue with creating such a data set is the multifactorial nature of infertility, as well as the variable treatment options in infertility cohorts. Furthermore, it is important to note that proper conservative interpretation of genetic data when pertaining to complex phenotypes like fertility is necessary to prevent erroneous associations and predictive algorithms.
We were unable to create a predictive algorithm using genotype data to better predict controlled ovarian stimulation in comparison to AMH. Though this small pilot study shows that genetics may not be a strong predicator of fertility outcomes, this does not discount that genetics could be involved in the etiology of fertility diseases.
View the poster below: