Human whole-epigenome modelling for clinical applications with Pleiades

Gene regulation in humans extends beyond the four-letter genetic code. DNA methylation, in particular, functions as a critical epigenetic regulator, dynamically programming cellular identity, adapting gene expression in response to environmental cues, and underpinning the onset and progression of numerous diseases. Here we present Pleiades, a series of whole-genome epigenetic foundation models spanning three sizes: 90M, 600M, and 7B parameters. Pleiades is trained upon an extensive proprietary corpus of human methylation and genomic data, totalling 1.9T tokens. We introduce alignment embeddings and stacked hierarchical attention techniques to provide precise epigenetic modelling without the need for extended context lengths. Collectively, these advances enable Pleiades to perform a diverse range of down-stream biological and clinical tasks, including genomic regulatory prediction, realistic generation of cell-free DNA fragments and fragment-level cell-type-of-origin classification, within a unified and scalable computational framework. We specifically apply Pleiades to the early detection of clinical Alzheimers disease and Parkinsons disease from plasma cell-free DNA, achieving high-accuracy detection (AUROC 0.89 for AD and 0.84 for PD) using a minimally invasive blood test. Combined with plasma pTau-217, Pleiades reaches an AUROC of 0.97 for AD, underscoring the promise of multimodal epigenomic and proteomic approaches. Using mechanistic interpretability, we ground Pleiades latent features in interpretable biological signals, relating its Alzheimers predictions to fragmentomic and epigenomic properties of cfDNA. These findings demonstrate the potential of genome-wide epigenomic modelling as a clinically translatable paradigm for diagnostics and precision medicine, though prospective validation in larger, more diverse cohorts will be required to confirm clinical utility.