Identificação de metabólitos séricos e salivares por meio de análise metabolômica e técnicas de aprendizado de máquina para o diagnóstico do câncer colorretal

Abstract

Introduction: Colorectal cancer (CRC) is the third most prevalent neoplasm worldwide, as well as the second leading cause of cancer death, and is frequently associated with late diagnosis due to low adherence to gold-standard screening and diagnostic methods such as colonoscopy, which is considered invasive. Therefore, the search for serum and salivary metabolites that can aid in the development of screening tools and contribute to a diagnosis better accepted by patients is necessary. Objective: To analyze the abundance of serum and salivary metabolites using Liquid Chromatography coupled to Mass Spectrometry (LC-MS) and machine learning (ML) techniques for CRC diagnosis. Materials and methods: In the first manuscript, a cross-sectional study was conducted with 52 CRC patients and 47 healthy control individuals matched for sex and age, in which serum samples were analyzed. In the second manuscript, a cross-sectional study was conducted with 55 CRC patients and 46 individuals in the control group, matched for sex and age, in which saliva samples were analyzed. In both studies, samples were analyzed using LC-MS. A heat map analysis with hierarchical clustering and a volcano plot were used to show the differentially abundant metabolites between the groups. Significance criteria (p ≤ 0.05; expression variation ≥ 2.0) were applied. Subsequently, ML algorithm approaches were applied. Diagnostic accuracy was evaluated through sensitivity, specificity, and respective ROC curves. Results: In the first manuscript, differentially abundant metabolites were identified between serum samples from the groups, with increased levels of sn-glycero-3-phosphocholine, 18:1(13Z)(17Me), glyceryl 5-hydroxydecanoate, and piperidine in CRC patients, while D-tryptophan showed greater abundance in the control group. Among the ML models evaluated, Random Forest showed the best diagnostic performance, with a sensitivity of 92.31%, specificity of 91.49%, and an area under the ROC curve of 0.973. In the second manuscript, an increase in metabolites related to vitamin E, polyamines, glycerides, phospholipids, and sphingolipids was observed in cancer patients, while glycerophospholipids, eicosanoids, and 17-Hydroxy-5alpha,17alpha-pregn-1-en-3-one showed greater abundance in the control group. The Naïve Bayes model demonstrated excellent discriminatory power, with a sensitivity of 97.83%, specificity of 96.36%, and an area under the curve of 0.996. Conclusion: The results indicate that metabolomics, combined with ML, represents a promising technique for non-invasive CRC screening, with the potential to optimize diagnosis, improve patient acceptance and, consequently, clinical prognosis, provided it is validated in future studies.