Interpretable precision medicine for acute myeloid leukemia

Abstract

Precision medicine (PM) is a branch of medicine that defines a disease at a higher resolution using genetic and other technologies to enable more specific targeting of its subgroups. Because of its uses in clinical treatment and diagnostics, this field exemplifies the modern era of medicine. PM looks for not just the right drug, but also the right dosage and treatment regimen. PM encounters a variety of challenges, which will be explored in this dissertation. Large-scale sensitivity screens and whole-exome sequencing experiments (WES) have fostered a new wave of targeted treatments based on finding associations between drug sensitivity and response biomarkers. These experiments with the aid of state-of-the-art artificial intelligence (AI) algorithms are opening new therapeutic opportunities for diseases with unmet clinical needs. It has been proved that AI is capable of predicting novel personalized treatments based on complex genotypic and phenotypic patterns in tumors. The scientific community should make an effort to make these algorithms to be interpretable to humans so that the results could be easily approved by the medical regulators. The purpose of this thesis is to apply AI algorithms for precision oncology that are highly accurate, while guaranteeing that the predictions are interpretable by humans. This work is divided in three main sections. The first section comprises a new methodology to increase the predictive power of the discovery of novel treatments in large-scale screenings by exploiting that some biomarkers tend to appear in many treatments. This fact is called hub effect in gene essentiality (HUGE). Content of this section was published in [1]. The second section contains a novel interpretable AI method -called multi-dimensional module optimization (MOM)- that associates drug screening with genetic events and proposes a treatment guideline. Content of this section was published in [2]. Finally, the third section includes a detailed comparison of different recently published algorithms that attempt to overcome the barriers proposed by today's precision medicine. This study also includes two novel algorithms specifically designed to solve the challenges of applicability to clinical practice: Optimal Decision Tree (ODT) and Multinomial Lasso. The characterization of Interpretable Artificial Intelligence as approach with strong potential for use in clinical practice is one of the study's most significant achievements. We presen tunique methods for PM that are highly interpretable, and we summarize the needs that could be considered for constructing interpretable AI. We are confident that this method will transform the way PM is addressed, bridging the gap between AI and clinical practice.