A diagnostic protein assay for differentiating follicular thyroid adenoma and carcinoma

Differentiating follicular thyroid adenoma (FTA) from carcinoma (FTC) remains challenging due to similar histological features separate from invasion. In this study, we aim to develop and validate DNA and protein-based classifiers for FTA/FTC differentiation. We collected 2443 samples from 1568 patients across 24 centers and applied next-generation sequencing, as well as discovery and targeted proteomics. Machine learning models were developed and compared utilizing DNA and/or protein features. The discovery protein-based model (AUC 0.899) outperformed the gene-based model (AUC 0.670). Consequently, we generated a protein-based model with targeted mass spectrometry and further validated it in three independent testing sets. The 24-protein-based model achieved high performance in the retrospective sets (AUC 0.871 and 0.853) and the prospective fine-needle aspiration biopsies (AUC 0.781). The classifier notably illustrated a 95.7% negative predictive value for ruling out malignant nodules. This study offers a promising protein-based approach for differential diagnosis of FTA and FTC. Highlights Genetic and proteomic profiling of follicular thyroid tumors from 1568 patients across 24 centers AI model based on proteins (AUC 0.899) outperformed that based on gene mutations (AUC 0.670) for differentiation of FTA and FTC Validation of the protein model in two retrospective cohorts (AUCs: 0.871, 0.853) and a prospective (AUC 0.781) The protein model has 95.7% negative predictive value for ruling out malignant nodules