Tissue-of-origin classification for cancer of unknown primary using DNA methylation data and the CpGPT transformer foundation model

Abstract

Cancer of unknown primary (CUP) accounts for approximately 2–3 percent of all cancer diagnoses, with incidence declining due to improved diagnostic technologies. Accurate identification of tissue origin is critical for treatment selection and prognosis, yet conventional immunohistochemistry achieves only 65.6 percent accuracy in metastatic tumors and requires highly skilled interpretation. DNA methylation-based classifiers such as EPICUP have demonstrated 87 percent accuracy in predicting tissue of origin for 216 CUP cases but require extensive tumor specific training datasets comprising 2,790 labeled samples from 38 tumor types. This study developed a novel approach using CpGPT, a DNA methylation foundation model pre-trained on over 1,500 datasets encompassing more than 100,000 samples, to classify tissue of origin through transfer learning without requiring tumor specific training data. We analyzed 3,600 primary tumor samples across 18 cancer types from The Cancer Genome Atlas using Illumina HumanMethylation450 BeadChip arrays. A novel ensemble embedding strategy was developed that generated 1,280-dimensional representations by performing ten independent iterations of random probe sampling (5,000 probes per iteration from 485,480 available CpG sites) and concatenating the resulting 128-dimensional CpGPT embeddings. Classification employed k-nearest neighbors (k=1) with L2 distance using FAISS similarity search, validated through leave-one-out cross-validation. Overall classification accuracy reached 87.56 percent (3,152/3,600 samples correctly classified), substantially exceeding immunohistochemistry benchmarks and producing results in a range consistent with the published performance of EPICUP, while utilizing transfer learning rather than tumor specific training. Per-class accuracy ranged from 75.0 percent for esophageal carcinoma to 98.5 percent for brain lower grade glioma, with 13 of 18 cancer types achieving accuracy above 85 percent. This work represents the first application of DNA methylation foundation models to cancer tissue-of-origin classification and demonstrates that transfer learning can achieve performance consistent with established classifiers while eliminating the need for large tumor specific training datasets, enabling rapid adaptation to new cancer types.