Using machine learning to increase the predictive value of humanized mouse models for the human immune response to YFV-17D

Abstract

Despite their utility as models for human systems, intrinsic differences between mouse and human biology limit direct translation of findings. Immunocompromised mice have been engrafted with functional human immune system components, or “humanized”, to better model the human immune response. Continued translational failure led to the development of a second-generation humanized mouse model with enhanced myeloid and natural killer (NK) cell compartments, producing a stronger immune response after exposure to the live attenuated yellow fever vaccine (YFV17D). Additionally, semi-supervised machine learning algorithms have been shown to further uncover translational insights from mouse models. I hypothesized that these strategies may be synergistic when combined, further improving the predictive value of these models. Here, I combine expression data from three humanized mouse models (NRG-HIS, NRG-HIS/Fluc, and NRG-HIS/Flt3LG) challenged with live attenuated yellow fever vaccine with machine learning (ML) models (k-nearest neighbors (KNN), random forest (RF), support vector machine (SVM), and neural network (NN)). Model predictions were evaluated for accuracy using F-scores and Matthews correlation coefficients. Several algorithms combined with mouse models made significantly better predictions about differential expression in the human immune response than models made with randomly classified samples. Semi-supervised NN, SVM, and RF algorithms combined with NRG-HIS/Fluc mice performed best for the tested human cohorts. Unexpectedly, the NRG-HIS/Fluc model outperformed the NRG-HIS model; the adenoviral vector itself may have increased production or recruitment of lymphocytes during infection. The best-performing models uncovered DEGs involved in detecting pathogens, innate immune response, and interferon signaling. TRIM22 and TRIM5, and potentially related genes, may be reliably uncovered by semi-supervised ML models applied to humanized mouse models, though further study is required for verification. Overall, modified humanized models, combined with ML approaches, can improve predictions of the human immune response.