Thyroid nodules are a common endocrine disease affecting approximately 50% of the global population. Fifteen to 30% cannot be confidently diagnosed as benign or malignant nodules by cytopathology before surgery, which may lead to overtreatment. Although molecular diagnostic tests for thyroid nodules have been developed, they are either rule-out or rule-in tests due to low predictive values and other limitations, such as RNA degradation. In this retrospective, blinded and multi-center study, we aimed to characterize benign and malignant thyroid tissues by using pressure cycling technology (PCT) coupled with data independent acquisition (DIA) mass spectrometry. In total, 1,984 formalin-fixed paraffin-embedded (FFPE) punches of five different histological types of thyroid nodules from 826 patients from four clinical centers were analyzed using the PCT-DIA method. In the discovery phase of the study, a classifier model (ThyroProt) to distinguish benign and malignant thyroid nodules based on a panel of 14 promising protein biomarker candidates was established from a patient cohort of 579 patients and 1,793 DIA data files by deep learning. In the blinded validation phase of the study, we tested ThyroProt on an independent cohort of 247 thyroid nodules with 494 DIA data files. ThyroProt correctly identified the benign or malignant status of 230 of 248 thyroid nodules with an accuracy of 92%, showing considerable superiority compared to published accuracies based on DNA and/or RNA analysis, such as 65% by Afirma gene expression classifier, 78% by genome sequencing classifier, and 84% by ThyroSeqV3, respectively. ThyroProt achieved high sensitivity, specificity, positive predictive value and negative predictive value of 90%, 94%, 94%, and 90%, respectively. More importantly, ThyroProt is applicable to FFPE tissues which are the most abundant and robust samples in clinical practice. In summary, ThyroProt is superior both in ruling-in and ruling-out cancer in thyroid nodules compared to extant diagnostic molecular tests.