(222d) A Metabolomics Approach to Predict Chemotherapy-Induced Peripheral Neuropathy

Introduction: Vincristine is a core chemotherapeutic drug administered to pediatric Acute Lymphoblastic Leukemia (ALL) patients. Although it is effective and has been in use for more than 50 years now, it has a dose-limiting toxicity: vincristine-induced peripheral neuropathy (VIPN). VIPN is characterized by symptoms such as a painful, numbing and tingling sensation felt in the palm and feet, with potential long-term effects[1]. Predictors and mechanism of VIPN incidence remain unclear, especially during the early phase of the treatment. Historically, an empirical cap of 2 mg is set for vincristine dosage. Due to this, a cohort of patients not susceptible to VIPN receives sub-therapeutic treatment, while another cohort may experience severe neuropathy. Thus, it is of interest to find predictors/biomarkers which can discriminate between these cohorts based on VIPN severity and hence aid in treatment personalization.

Methods and Results: Here, we employed a metabolomics approach to find metabolites as biomarkers that can accurately predict VIPN in pediatric ALL patients. We performed untargeted metabolite profiling using tandem mass spectrometry on blood samples of ALL patients undergoing chemotherapy with vincristine as the primary drug. We analyzed the samples at three time points: day 8, day 29 and around 6 months during the treatment. We classified these patients as being susceptible to high or low neuropathy based on total neuropathy score calculated frequently throughout the course of the treatment. Following metabolite profiling, we used machine learning techniques to analyze the profile data and determine a small set of metabolites highly predictive for overall VIPN susceptibility at those time points. Firstly, we performed a principal component analysis which showed that the metabolite profiles cluster according to the time point, implying that they need to be analyzed independently, rather than longitudinally. Then, we used support vector classifier along with recursive feature elimination algorithm to find the small sets of predictive metabolites. Cross-validation models built with these selected metabolites as predictors had an area under receiver operating characteristics (AUROC) curve of approximately 0.95 at all the time points. Furthermore, we performed rigorous manual validation of the chosen metabolites’ chromatogram peaks to ensure that they were integrated correctly by the mass spectrometry instrument. After manual validation, we selected the correctly integrated metabolites and built predictive models with them. Models built using the day 8 and 6 months data were more accurate, with AUROC greater than 0.9 in both the cases. Thus, one could use these trained models at day 8 and 6 months during the treatment to predict VIPN susceptibility and subsequently adjust vincristine dosage. We further identified few of these metabolites using their mass, retention time and adduct information. Lastly, we found relevant biological pathways associated with them.

Conclusion: We found small sets of metabolites that could accurately predict VIPN in pediatric ALL patients. We then built predictive models with these metabolites which can potentially aid in vincristine dosage decision making for clinicians, and subsequently improve treatment efficacy for patients. This approach can further be applied to predict peripheral neuropathy as a result of other chemotherapeutic drugs as well.

References:

1. Gidding, C. E. M., et al. "Vincristine revisited." Critical reviews in oncology/hematology 29.3 (1999): 267-287.