Seonglae ChoThrough an interpretability analysis of the epigenetics foundation model Pleiades, a new class of biomarkers for early Alzheimer’s diagnosis was identified. A range of interpretability methods were applied, including supervised probing, sparse autoencoders (BatchTopK SAE), and gradient attribution. The nine SAE features that most strongly drove disease classification were all correlated with DNA fragment length. Fragmentomics is a biomarker class that has been used mainly for cancer detection, and it has not been sufficiently explored in Alzheimer’s. However, the available low-grade/low-quality data is limited.
Using Interpretability to Identify a Novel Class of Alzheimer's Biomarkers
An AI model was trained to detect Alzheimer's from blood samples. We opened it up to understand how—and found that DNA fragment length patterns dominate its decision-making. We distilled this insight into a human-interpretable classifier that generalizes better than the biomarker classes previously reported in the literature when tested on an independent cohort.
https://www.goodfire.ai/research/interpretability-for-alzheimers-detection#
