Pleiades

Most genomic language models read only the four-letter sequence. Pleiades, developed by Prima Mente, is a family of whole-genome foundation models that treat DNA methylation as a first-class signal alongside sequence, learning the epigenetic layer that programs cellular identity and shifts during disease. The family spans three sizes — 90M, 600M, and 7B parameters — and is built as an autoregressive transformer decoder trained at single-nucleotide resolution.

The motivating application is cell-free DNA (cfDNA). Fragments circulating in plasma carry the methylation marks of the cells they came from, which makes them a route to reading tissue state from a blood draw — but the signal is diffuse, and the molecules are short. Pleiades approaches this by modelling sequence and methylation jointly at base resolution, so a single pretrained backbone can infer where a fragment originated and what regulatory state it reflects.

Scope is what distinguishes Pleiades from its neighbours. Where MethylAI predicts methylation at individual CpGs from sequence, CpGPT learns array-based methylation representations for imputation and clock fine-tuning, and LEAF-1 embeds cfDNA fragments for cancer detection and cell typing, Pleiades reuses one fixed backbone across regulatory prediction, generative fragment synthesis, cell-type-of-origin classification, and clinical diagnosis. It is a commercial, closed model: no code or weights have been released, and the work is a preprint that has not been peer reviewed.

#Key Features

• Joint sequence-and-methylation modelling: Genomic sequence and its methylation state are modelled together at single-nucleotide resolution, rather than treating methylation as a downstream label predicted from sequence.
• Stacked hierarchical attention: Combined with alignment embeddings, this lets the models capture genome-scale epigenetic context without relying on extended context windows.
• One backbone, many tasks: The same pretrained model supports genomic regulatory prediction, cell-type-of-origin classification, and clinical prediction, rather than being refit per dataset.
• Generative cfDNA synthesis: The models generate realistic cell-free DNA fragments, enabling in-silico experimentation on fragment populations.
• Interpretability grounding: Mechanistic interpretability analyses tie latent features to fragmentomic and epigenomic properties of cfDNA, linking the disease predictions back to measurable biology.

#Technical Details

Pleiades is pretrained on a proprietary corpus of human methylation and genomic data totalling 1.9 trillion tokens, drawing on methylation atlases spanning human cell types, cfDNA profiles, and reference genome sequence. The three model sizes — 90M, 600M, and 7B parameters — share an autoregressive decoder architecture, and the two architectural contributions — alignment embeddings and stacked hierarchical attention — are what let them model the epigenome precisely at single-nucleotide resolution without extended context lengths.

On clinical evaluation, Pleiades detects Alzheimer's disease from plasma cfDNA at an AUROC of 0.89 and Parkinson's disease at 0.84. Combining its cfDNA-derived signal with plasma pTau-217 raises Alzheimer's detection to an AUROC of 0.97, indicating that the epigenomic signal carries information complementary to an established protein biomarker. The authors note that prospective validation in larger and more diverse cohorts is required before clinical utility can be claimed. The preprint is released under a CC BY-NC-ND licence.

#Applications

The clearest application is minimally invasive neurodegenerative disease detection: classifying Alzheimer's and Parkinson's disease from a plasma sample, alone or alongside protein biomarkers, which is relevant to trial enrolment and to early diagnosis where confirmation otherwise depends on imaging or cerebrospinal fluid. Beyond diagnostics, fragment-level cell-type-of-origin classification serves any liquid-biopsy analysis that must attribute circulating DNA to its tissue source, the generative capability supports in-silico studies of cfDNA fragment populations, and the regulatory predictions apply to epigenomic research more broadly. Because neither code nor weights are public, these capabilities are accessible only through Prima Mente.

#Impact

Pleiades scales foundation-model pretraining to the human epigenome at whole-genome, single-nucleotide resolution and carries a single such backbone through to a clinical endpoint. Its architectural claim — that hierarchical attention plus alignment embeddings can substitute for very long context windows in epigenomic modelling — matters to anyone building genomic sequence models, where context length is a persistent constraint. Its influence on the open research community is limited by its closed status: the proprietary corpus and unreleased weights mean the results cannot be independently reproduced or extended, and the reported AUROCs are retrospective, awaiting peer review and prospective cohort validation.