ALLEVIATE: Artificial intelligence to improve acute lymphoblastic leukemia treatment

Acute lymphoblastic leukemia (ALL) is a heterogeneous form of cancer affecting the white blood cells, which includes multiple molecular subtypes, each with its own characteristics and prognosis. Although ALL is rare, it remains the most prevalent cancer among children and can also affect adults. Over the past few years, there has been an increasing focus on understanding the genetic factors involved as well as the use of data analysis and machine learning methods to enhance our knowledge of the genomic landscape, prognosis, and treatment responses in patients with ALL. This PhD aims to explore the use of complex data analysis and AI-driven methods to solve complex clinical issues by combining clinical risk factors and genetics.
Study I  investigated the variation in white blood cell count (WBC) in patients with ALL and explored potential predispositions from host genome variations. Spline functions were used to model WBC with its known associated features such as age, immunophenotype, and cytogenetics, followed by a genome-wide association study (GWAS) to associate with host genetics. Our results indicated an inverse correlation between age and WBC, which was stronger in specific patient subgroups of immunophenotype and cytogenetics. Although no significant associations (p < 5 · 10−8) were found between WBC and single nucleotide polymorphisms (SNPs), the top associated SNPs suggested effects on signal transduction, translation initiation, cell development, and proliferation. These findings suggest that host genome variants do not strongly influence WBC across different subsets of ALL. Further studies should focus on specific subsets and complex analyses to understand the factors contributing to WBC.
Study II focused on the prediction of asparaginase-associated pancreatitis (AAP) in pediatric patients with ALL, a. Using machine learning models and incorporating age, sex, and SNPs, we were able to improve the predictive performance up to ROC-AUC (area under the receiver operating characteristic curve) of 0.84 for a model ensemble on a held out test set. The inclusion of candidate SNPs related to pancreatitis and validated pancreatitis SNPs further enhanced the models’ performance. These findings offer the potential for individual-level risk assessment of AAP and hold implications for future prevention trials, especially regarding asparaginase re-exposure, allowing for more personalized and optimized treatment approaches.
Study III aimed to improve the prediction of relapse in patients with ALL by leveraging pharmacogenetics compared to a clinical model. To train machine learning models, clinical features of age, sex, immunophenotype and WBC available at diagnosis in addition to germline SNPs. Incorporating SNPs related to drug metabolism and biological pathways into the models improved relapse prediction performance, achieving an ROC-AUC of 0.70 compared to the clinical models’ ROC-AUC of 0.67. The best performing model included age, sex, immunophenotype, logWBC, and 18 SNPs from the Reactome Suppression of Apoptosis gene set. The best model in terms of MCC (Matthew’s correlation coefficient) included 39 SNPs related to the mercaptopurine action pathway. Patient stratification based on these two models revealed three subgroups with significant differences in prognosis even within groups from current risk stratification. These findings highlight the potential of SNPs to enhance relapse prediction, but needs further validation and refinement in larger patient populations and treatment protocols.
Study IV presents ongoing work to investigate the interplay and landscape of germline (inherited) and somatic (acquired) variation in pediatric ALL. Genomic events of germline and somatic variants, as well as gene fusions, were identified from whole genome sequencing and RNA sequencing, where a set of 5904 genes related to cancer predisposition and functional capabilities were used as scaffold for the investigation. Preliminary results show heterogeneous patterns affecting the landscape with differences between individual, but suggest an interplay with complementary functions. 
In summary, the studies presented in this thesis highlight the potential of complex data analysis and machine learning approaches in integrating diverse sets of data related to ALL and its treatment. By incorporating clinical features, genetic variations, and prior knowledge, these studies demonstrate the feasibility of machine learning in improving prognostic predictions and individual-level risk assessment for complex disease outcomes like AAP and relapse. The findings emphasize the value derived from combining heterogeneous patient data, for more precise strategies in the prevention and treatment of ALL. This thesis contributes to the broader goal of translating data science and machine learning to insights for precision medicine in ALL.