crossNN is an explainable framework for cross-platform DNA methylation-based classification of tumors [sequencing-based methylome profiling]

DNA methylation-based classification of brain tumors has emerged as a powerful and indispensable diagnostic technique. Initial implementations have used methylation microarrays for data generation, but different sequencing approaches are increasingly used. Most current classifiers, however, rely on a fixed methylation feature space, rendering them incompatible with other platforms, especially different flavors of DNA sequencing. Here, we describe crossNN, a neural network-based machine learning framework which can accurately classify tumor entities using DNA methylation profiles obtained from different platforms and with different epigenome coverage and sequencing depth. It outperforms other deep and conventional machine learning models with respect to diagnostic accuracy and computational requirements while still being fully explainable. We use crossNN to train a pan-cancer classifier to discriminate more than 170 tumor types across all organ sites. Validation in an independent cohort of >5,000 tumors profiled using different microarray and sequencing platforms, including low-pass nanopore and targeted bisulfite sequencing, demonstrates the robustness and scalability of the model with 99.1% and 97.8% precision for brain tumor and pan-cancer model, respectively. Tumor genomic DNA was subjected to methylation array or sequencing-based methylome profiling. Platform-specific preprocessing was performed as indicated and beta values (microarray) or methylated allele frequencies (all sequencing platforms) were used for classification by crossNN after binarization. *************************************************************** Raw files for human/patient samples were not submitted to GEO due to concerns about submitting personally identifiable sequence data for open access. ***************************************************************