A compact encoding of the genome suitable for machine learning prediction of traits and genetic risk scores

Genotype to phenotype prediction is a central problem in biology and medicine. Machine learning is a natural tool to address this problem. However, a person's genotype is usually represented by a few million single-nucleotide polymorphisms and most datasets only have a few thousand patients. Thus, this problem typically has many more predictors than the number of samples (patients), making it unsuitable for machine learning. The objective of this paper is to examine the efficacy of a compact genotype representation, which employs a limited number of predictors, in predicting a person's phenotype through the application of machine learning. We characterized a person's genotype using chromosome-scale length variation, a measure that is computed as the average value of reported log R ratios across a portion of a chromosome. We computed these numbers from data collected by the NIH All of Us program. We used the AutoML function (h2o.ai) in binary classification mode to identify the best models to differentiate between male/female, Black/white, white/Asian, and Black/Asian. We also used the AutoML function in regression mode to predict the height of people based on their age and genotype. Our results showed that we could effectively classify a person, using only information from chromosomes 1-22, as Male/Female (AUC = 0.9988 ± 0.0001), White/Black (AUC = 0.970 ± 0.002), Asian/White (AUC = 0.877 ± 0.002), and Black/Asian (AUC = 0.966 ± 0.002). This approach also effectively predicted height. In conclusion, we have shown that this compact representation of a person's genotype, along with machine learning, can effectively predict a person's phenotype.