Machine learning approaches to predict the human microbiome and disease associations

The human microbiome contains trillions of microorganisms and their genetic materials, called microbiota, that affect human health and are known to have an effect on the formation or prevention of diseases. Machine Learning analyzes to be performed on the data sets created with the genetic sequence samples taken here will enable the early diagnosis of many diseases. For this purpose, in this study, modern machine learning techniques including k-Nearest Neighbor, Light Gradient Boosting Machine, Logistic Regression, Random Forest, Support Vector Machine and XGBoost were used to predict Inflammatory Bowel Disease and its subtype Crohn's Disease with raw sequence samples taken from human gut microbiota. The dataset obtained from public databases preprocessed and converted into raw sequence data. These genetic sequences were first converted into k-mer based De Bruijn graphs and then into a structured form with the method developed in this thesis. Three analyzes were performed using this data set. In the first analysis, only individuals with Crohn's Disease were considered patients and the data set was left unbalanced. In the second analysis, the data set in the first analysis was balanced with the SMOTE algorithm. In the final analysis, samples with all types of Inflammatory Bowel Disease were considered patients and the data set was balanced accordingly. The performances of Machine Learning models were measured and compared with 5 separate metrics. Considering the accuracy scores, the highest performing models were LGBM (76.47%) for k-mer length 5 in the first analysis, RF (79.52%) for k-mer length 6 in the second analysis, and SVM (95.11%) for k-mer length 5 in the final analysis. The McNemar test performed in the last part showed that the analyzes were statistically significant.