Training with synthetic data provides accurate and openly-available DNA methylation classifiers for developmental disorders and congenital anomalies via MethaDory

Multiple developmental and congenital disorders due to genetic variants or environmental exposures are associated with unique genome-wide alterations in DNA methylation (DNAm). Consequently, these patterns referred to as DNAm signatures, can be leveraged for diagnostic purposes by developing artificial intelligence (AI) models that enable molecular subclassification of individuals. Notably, DNAm signature application has been particularly successful for diagnosing individuals affected by developmental disorders and congenital anomalies, especially those with defects in genes encoding the Mendelian epigenetic machinery. So far, over 100 DNAm distinctive signatures have been reported for these disorders. However, the translation into diagnostic practice remains challenging not only because of the scarcity of samples from these (ultra)rare disorders needed to train DNAm AI models, but also due to the privacy regulations that restrict the sharing of affected individuals data, lack of methods for standardization, limited replication across different centers, and the emergence of commercial entities with competing interests. In this study, we show that synthetic cases, meaning in silico cases generated from publicly available DNAm data from unaffected individuals, and summarized data derived from anonymized study cohorts of affected individuals with certain disorders, can be used to train DNAm classifiers. We demonstrate that these DNAm classifiers trained on a large cohort of synthetic cases have an improved performance compared to previously published classifiers trained on cohorts of affected individuals only, which typically are limited in size due to the rarity of these conditions. Furthermore, they improve the classification of variants with intermediate effect and mosaic cases and do not require any private affected individual data for training. Finally, to facilitate dissemination of these models, we release 169 synthetic cases-trained DNAm classifiers for 89 disorders with MethaDory, an open-access tool for simultaneous testing of these DNAm signatures.