Session: 618. Acute Myeloid Leukemias: Biomarkers and Molecular Markers in Diagnosis and Prognosis: Poster II
Hematology Disease Topics & Pathways:
Research, Acute Myeloid Malignancies, AML, Adult, Translational Research, Bioinformatics, Diseases, Treatment Considerations, Metabolism, Myeloid Malignancies, Biological Processes, Technology and Procedures, Study Population, Human, Machine learning
Despite advances in cancer research, outcomes for patients with acute myeloid leukemia (AML) remain poor due to the heterogeneous nature of the disease (Kantarjian et al, 2021). Variations in mutations correspond to unique metabolic states (Simonetti et al, 2021) which are responsible in part for mediating variability in therapy response and relapse biology (Jones et al, 2018, 2020). To fully unravel patient heterogeneity, it is necessary to understand the metabolic changes across AML subtypes. Previous research has demonstrated the impact of clinical features on their circulating metabolome (Bar et al, 2020). However, a comprehensive survey of circulating lipids and metabolites has not yet been undertaken in AML. Therefore, we examined the predictive power of the circulating plasma lipidome in AML patients.
We quantified the circulating metabolism from 231 patients at diagnosis. Annotation of 88 clinical features was performed for all patients, including: age, sex, blood count characteristics, mutational status, ELN risk score, cytogenetics, therapy response, relapse status, and overall survival (OS). Metabolite levels in each plasma sample was interrogated by mass spectrometry, identifying 177 metabolites, and 1988 lipids. Metabolites and metabolic pathways were analyzed using MetaboAnalystR. Of the 88 features assessed, 35 had significant associations with specific metabolites and/or lipids. Further, 28 of the 66 AML mutations analyzed also had distinct metabolites and/or lipids associated with it, as well as distinct pathway enrichment and lipid structures. Some of these findings align with known AML biology such as an increase in glutamate metabolism in FLT3-ITD mutant samples (Gregory et al, 2016), representing the usefulness of this resource for future studies.
Predicting outcomes and therapy response in AML is continually evolving to account for patient heterogeneity. Given that metabolites and lipids can help describe heterogeneous patient features, we next assessed the ability of lipids and metabolites to predict OS by Cox regression models. Both were found to be associated with OS, with top lipids (C-index=0.590) having a stronger association over top metabolites (C-index=0.577). Notably, lipids demonstrated a predictive capacity comparable to the current clinical standard, ELN score (C-index=0.666), suggesting that lipids could serve as an alternative predictor for OS.
We next analyzed whether individual lipids or metabolites can be predictive of therapy response. Selecting the best treatment for patients upfront can reduce the burden on patients as well as the healthcare system. Thus, predicting response to chemotherapy at diagnosis can be of clinical use. Using orthogonal partial least squares – discriminant analysis we found that plasma lipid levels significantly separated patients that achieved a complete response compared to patients that did not achieve a clinical remission post chemotherapy treatment. Further, only lipids were found to be predictive of therapy response with the top features including cardiolipins, and phospholipids, highlighting the importance of lipid metabolism on AML outcomes.
Based on lipids’ predictive power, we tested the capability of the lipidomics data to predict therapy response in a clinically meaningful capacity. We built machine learning models for the response to therapy lipidomic data, with, and without clinical features (age, cytogenetic risk, white blood cell count, AML diagnosis), as well as clinical features alone. The data was split into training and testing sets (80%/20%) and feature selection was performed by removing correlated features. Four machine learning models were tested by nested cross-validation (CV): ExtraTreesClassifier (ETC), RandomForestClassifier, XGBoost, and support vector machine. The best model was selected based on the CV score, followed by hyperparameter tuning. The lipidomics data demonstrated exceptional performance using ETC on test data to predict therapy response. Lipids alone and lipids with clinical features performed nearly identically (AUCs 0.95 and 0.96), surpassing the performance of clinical features alone (AUC 0.56), suggesting that lipids could be prospective biomarkers for therapy response.
Together, these data demonstrate that the circulating metabolome can stratify heterogenous AML patient populations and predict outcomes in AML.
Disclosures: No relevant conflicts of interest to declare.