Researchers at Goodfire trained an AI model, Pleiades, to detect Alzheimer's disease (AD) from blood samples by analyzing cell-free DNA (cfDNA). They discovered that patterns in DNA fragment lengths were key indicators in the model’s decision-making process. This insight allowed them to create a simpler, human-readable classifier that outperformed existing biomarker classifications when tested on an independent cohort. The model achieved an area under the receiver operating characteristic curve (AUROC) score of 0.84, indicating strong predictive capability. The study emphasizes the importance of fragmentomics, which examines how DNA fragments are structured and how these structures can signal disease presence. While this area has seen significant research in cancer detection, its application to neurodegenerative diseases like Alzheimer's has been less explored. The researchers used advanced interpretability techniques to pinpoint fragment length as a novel biomarker, demonstrating its potential for improving diagnosis. Pleiades itself is a robust 7-billion parameter epigenetic model trained on vast genomic datasets. It incorporates both genetic sequences and epigenetic modifications, allowing it to capture complex biological patterns. The model's training involved analyzing over 30 billion cfDNA fragments from 81 individuals, with half diagnosed with AD. By leveraging these internal representations, researchers developed a hierarchical attention transformer to effectively predict Alzheimer's, using cfDNA sequences that rival traditional proteomic biomarkers in accuracy.