Research
Projects
Advances in clinical metagenomics: predictive capability of the salivary microbiome using machine...

Advances in clinical metagenomics: predictive capability of the salivary microbiome using machine learning for the diagnosis of oral and systemic diseases

This project is in line with the scientific-technical priority of the AES 2021-2023 entitled "Digital tools, technologies and solutions for health and care, promoting the development and use of innovative techniques, technologies and tools to improve the quality of life", being included in the priority line: "Improvements in the processes of prevention, prediction, diagnosis and follow-up of diseases and monitoring of the therapeutic response".

The following hypothesis and objectives are proposed: We intend to demonstrate the predictive capacity of the salivary microbiome to discriminate with high diagnostic accuracy the presence of oral diseases (gingivitis, periodontitis and caries) and systemic diseases (oral cancer and colorectal cancer). For this purpose, starting from the data obtained from massive sequencing of the 16S rRNA gene on salivary samples and applying traditional and advanced Machine Learning (ML) and Deep Learning (DL) techniques, we propose the development of predictive models that will allow the classification of oral or systemic pathologies of study and models that can classify at the same time more than one pathology in a sample.

Five methodological phases are established: obtaining salivary samples from controls and patients; 16S metabarcoding of the samples; bioinformatic analysis of the metagenomic data; biostatistical analysis of predictive diagnostic modelling (ML and DL) and biological/clinical interpretation of the results obtained.

Objectives

From the data obtained by massive sequencing of the 16S rRNA gene on salivary samples and applying traditional and advanced Machine Learning and Deep Learning techniques, we propose the following specific objectives:

  1. Development of two-class predictive models that will allow us to classify oral or systemic study pathologies from a single sample per patient. Each two-class model will have as a reference or control the healthy oral and healthy systemic patients, while its ‘target’ will be each of the pathologies studied.
  2. Calculate the probability of having specifically each pathology through the correct calibration of the model that allows an adequate interpretation of the results.
  3. Evaluate how the performance of each model is affected by data from samples of pathologies other than those with which it was trained.
  4. Development of multi-label predictive models, i.e., models that can simultaneously classify more than one clinical pathology in a sample.
  5. Calculate the probability of having each pathology studied specifically through the correct calibration, allowing an adequate interpretation of the results.
  6. Evaluate and control the degree of over-fitting of the models to ensure their generalisable application in the clinical setting.
  7. To assess the influence of covariates (age, sex and smoking) on the predictive models previously used.