Uncovering the Gene Profile of Acute Myeloid Leukaemia Using Gradient Boosted Trees

Abstract Applying the iterative methodology for dimensionality reduction using categorical gradient boosted trees, as it has been defined in [9], on a dataset consisted of 12708 genes expressions coming from 5052 individuals from 105 studies, we classify whether a person has acute myeloid leukaemia (AML) or is healthy. A CatBoost model on the dataset with reduced dimensions of 72 genes reached a ROC- AUC score of 0.9973 and F1-score: 0.9983 using ten fold cross validation and ROC-AUC: 0.9988 and F1-score: 0.9988 on an inference dataset. The dimension of the genes used by the previous model is then further reduced by removing the genes that do not have any bibliographic reference to AML. A CatBoost model is trained on this final dataset consisting of 63 genes, providing a ROC-AUC score of 0.9941 and F1-score: 0.9973 on ten fold cross validation and ROC-AUC: 0.9942 and F1-score: 0.9964 showing that not all genes with a correlation to AML have been found yet. This work could be considered as complimentary to the work of [9] where they used probe-sets. We conclude that the iterative method defined in [9] can lead to the identification of the gene profile of AML and also to identification of genes associated to AML which have never been correlated to the disease before.