Accepted_test

Application of Interpretive Machine Learning methods to find transcriptomic biomarkers of Alzheimer's disease

by Asainov D. | Kirillov B. | Department of Biochemistry, Biotechnology and Pharmacology, Institute of Fundamental Medicine and Biology, Kazan (Volga Region) Federal University, Kazan, Russia | Materials Technology Center, Skolkovo Institute of Science and Technology, Moscow, Russia

Abstract ID: 682

Event: BGRS-abstracts

Sections: [Sym 9] Section “Gene expression and human diseases”

Diagnosis of Alzheimer's disease (AD) is a lengthy process, that includes gathering of anamnesis, tests for the patient's mental status, invasive tests necessary to assess the level of beta-amyloid in the cerebrospinal fluid, and neuroimaging methods. However, early diagnosis using classical methods is complicated, indicates the serious potential of alternative methods of diagnosing AD. Such methods include machine learning models, which offer big opportunities. However, the majority of models possess a "black box" problem – the lack of information about how machine learning models make decisions. Explainable Machine Learning (xML, IML [2]) methods – explainable or interpretable machine learning – have been created to address this problem.

To train the models, we use the Ludwig framework focused on deep learning methods. The dataset we use is RNA expression data for the middle temporal gyrus of patients with diagnosed AD and controls. We use the Accumulated Local Effects (ALE) algorithm as part of the Alibi library to generate explanations for our predictions.

We have trained a number of models predicting the presence of AD. The explanations obtained with ALE allowed us to identify several transcripts that are particularly important input features for the trained models. Relevant genes of interest, found by our ALE test, are involved in blood coagulation, the occurrence of inflammatory responses, dopamine biosynthesis and maintenance of macrophage function.