Synthetic epigenomes for open cross-platform episignature classifiers

DNA methylation (DNAm) signature (episignature) classifiers support diagnosis of rare diseases. Yet episignature testing is confined to a dedicated array platform, and most classifiers are neither openly available nor fully reproducible. Laboratories cannot train their own classifiers as positive controls are scarce and cannot be shared under privacy regulations. Here, we develop and release MethaDory, a framework to train episignature classifiers using synthetic DNAm profiles generated in silico from control data and summary statistics of affected individuals from any platform of interest. We trained 188 classifiers spanning 102 rare disorders and validated 154 across 915 individuals with an episignature-associated diagnosis, 533 healthy controls profiled with DNAm arrays, Oxford Nanopore, and PacBio sequencing, together with 7 international centres. MethaDory achieved an average F1 0.9 on arrays, matching or outperforming existing approaches. We show MethaDory classifier application to long-read genome data, facilitating diagnoses in unsolved cases and in ultra-rapid diagnostic settings.