Ayelet Zlotogorski-Hurvitz1,2, Ben Zion Dekel3, Dov Malonek3, Ran Yahalom4, Marilena Vered5,6. 1. Department of Oral Pathology and Oral Medicine, School of Dentistry, Tel Aviv University, 69978, Tel Aviv, Israel. 2. Department of Oral and Maxillofacial Surgery, Rabin Medical Center, Petah Tikva, Israel. 3. Department of Electrical and Computer Engineering, Ruppin Academic Center, Emek Hefer, Israel. 4. Department of Oral and Maxillofacial Surgery, The Chaim Sheba Medical Center, Tel Hashomer, Israel. 5. Department of Oral Pathology and Oral Medicine, School of Dentistry, Tel Aviv University, 69978, Tel Aviv, Israel. mvered@post.tau.ac.il. 6. Institute of Pathology, The Chaim Sheba Medical Center, Tel Hashomer, Israel. mvered@post.tau.ac.il.
Abstract
PURPOSE: To determine the Fourier-transform infrared (FTIR) spectra of salivary exosomes from oral cancer (OC) patients and healthy individuals (HI) and to assess its diagnostic potential using computational-aided models. METHODS: Whole saliva samples were collected from 21 OC patients and 13 HI. Exosomes were pelleted using differential centrifugation (12,000g, 120,000g). The mid-infrared (IR) absorbance spectra (900-5000 cm- 1 range) were measured using MIR8025 Oriel Fourier-transform IR equipped with a PIKE MIRacle ZnSe attenuated total reflectance attachment. Machine learning techniques, utilized to build discrimination models for the absorbance data of OC and HI, included the principal component analysis-linear discriminant analysis (PCA-LDA) and support vector machine (SVM) classification. Sensitivity, specificity and the area under the receiver operating characteristic curve were calculated. RESULTS: IR spectra of OC were consistently different from HI at 1072 cm- 1 (nucleic acids), 2924 cm- 1 and 2854 cm- 1 (membranous lipids), and 1543 cm- 1 (transmembrane proteins). The PCA-LDA discrimination model correctly classified the samples with a sensitivity of 100%, specificity of 89% and accuracy of 95%, and the SVM showed a training accuracy of 100% and a cross-validation accuracy of 89%. CONCLUSION: We showed the specific IR spectral signature for OC salivary exosomes, which was accurately differentiated from HI exosomes based on detecting subtle changes in the conformations of proteins, lipids and nucleic acids using optimized artificial neural networks with small data sets. This non-invasive method should be further investigated for diagnosis of oral cancer at its very early stages or in oral lesions with potential for malignant transformation.
PURPOSE: To determine the Fourier-transform infrared (FTIR) spectra of salivary exosomes from oral cancer (OC) patients and healthy individuals (HI) and to assess its diagnostic potential using computational-aided models. METHODS: Whole saliva samples were collected from 21 OC patients and 13 HI. Exosomes were pelleted using differential centrifugation (12,000g, 120,000g). The mid-infrared (IR) absorbance spectra (900-5000 cm- 1 range) were measured using MIR8025 Oriel Fourier-transform IR equipped with a PIKE MIRacle ZnSe attenuated total reflectance attachment. Machine learning techniques, utilized to build discrimination models for the absorbance data of OC and HI, included the principal component analysis-linear discriminant analysis (PCA-LDA) and support vector machine (SVM) classification. Sensitivity, specificity and the area under the receiver operating characteristic curve were calculated. RESULTS: IR spectra of OC were consistently different from HI at 1072 cm- 1 (nucleic acids), 2924 cm- 1 and 2854 cm- 1 (membranous lipids), and 1543 cm- 1 (transmembrane proteins). The PCA-LDA discrimination model correctly classified the samples with a sensitivity of 100%, specificity of 89% and accuracy of 95%, and the SVM showed a training accuracy of 100% and a cross-validation accuracy of 89%. CONCLUSION: We showed the specific IR spectral signature for OC salivary exosomes, which was accurately differentiated from HI exosomes based on detecting subtle changes in the conformations of proteins, lipids and nucleic acids using optimized artificial neural networks with small data sets. This non-invasive method should be further investigated for diagnosis of oral cancer at its very early stages or in oral lesions with potential for malignant transformation.