Abstract PO-091: Modeling the effects of glucose accessibility on tumor cell growth: A comparison of mechanism-based and machine learning models

Abstract

Abstract Introduction: The altered consumption and utilization of glucose in tumor cells play an important role in the development of a tumor. We compare a mechanism-based model and four common machine learning approaches to predict how access to glucose (as affected by a GLUT1 inhibitor) influences tumor cell growth. Materials and Methods: Cytochalasin B was used as a GLUT1 inhibitor in MDA-MB-231 breast cancer cells. Mechanism-based model: We used a system of ordinary differential equations to describe the change of live tumor cell number N, dead tumor cell number D, real glucose concentration Greal, and accessible glucose concentration, Gacs over time t. Gacs is described as a function of Greal and the number of tumor cells. Given the measured time courses of N and D, the system estimates the proliferation rate kp, the death rate due to glucose depletion kd, the death rate due to bystander effect kbys, the consumption rate of glucose v, and the dose-related inhibition constant Ginh.Machine learning model: The initial number of live and dead tumor cells, glucose concentration, inhibitor concentration, and time of measurement were input as features, and N and D were selected as prediction targets. The features and prediction targets were trained on a linear regression model, a k-nearest neighbors regression model, a decision tree model, and a random forest model with the scikit-learn library in Python. Training and validation: The dataset (n = 300) was split into 75% training and 25% validation. All five models (i.e., one mechanism-based, and four machine-learning-based models) were trained to predict tumor cell growth for the validation set given the initial conditions. The residual sum of squares (RSS) and the coefficient of determination (R2) were compared to evaluate the performance of the models. Results: For the mechanism-based model, we estimated kp at 0.14 ± 0.003 day−1, kd at 0.041 ± 0.006 day−1, v at 4.48 ± 0.15 × 10−5 mM·cell−1·day−1, Ginh at 2.90 ± 0.12 × 10−5 cell−1 for 2 μM Cytochalasin B and 4.37 ± 0.18 × 10−5 cell−1 for 10 μM Cytochalasin B. The R2 value for the calibration of the mechanism-based model on the training data set was 0.97, while the values for the decision tree model, random forest model, k-nearest neighbor regression model, and the linear regression model were 1.00, 1.00, 0.94, and 0.70, respectively. The random forest model provided the best prediction on the validation set with an R2 of 0.92, followed by the decision tree, k-nearest neighbors regression, mechanism-based, and linear regression models with R2 values of 0.88, 0.86, 0.83, and 0.71, respectively. Conclusion: This study provides an example of a mechanism-based model that exhibits predictive ability comparable to that of a machine learning approach—but with the added benefit of explicitly accounting for underlying biological mechanisms. Mechanism-based modeling enables a number of further experiments and analyses (e.g., treatment optimization and experimental design) not possible using a machine learning approach. Citation Format: Jianchen Yang, Jack Virostko, David A. Hormuthm, II, Thomas E. Yankeelov. Modeling the effects of glucose accessibility on tumor cell growth: A comparison of mechanism-based and machine learning models [abstract]. In: Proceedings of the AACR Virtual Special Conference on Artificial Intelligence, Diagnosis, and Imaging; 2021 Jan 13-14. Philadelphia (PA): AACR; Clin Cancer Res 2021;27(5_Suppl):Abstract nr PO-091.