Predicting ionizing radiation exposure using biochemically-inspired genomic machine learning

Abstract

Background: Gene signatures derived from transcriptomic data using machine learning methods have shown promise for biodosimetry testing. These signatures may not be sufficiently robust for large scale testing, as their performance has not been adequately validated on external, independent datasets. The present study develops human and murine signatures with biochemically-inspired machine learning that are strictly validated using k-fold and traditional approaches. Methods: Gene Expression Omnibus (GEO) datasets of exposed human and murine lymphocytes were preprocessed via nearest neighbor imputation and expression of genes implicated in the literature to be responsive to radiation exposure (n=998) were then ranked by Minimum Redundancy Maximum Relevance (mRMR). Optimal signatures were derived by backward, complete, and forward sequential feature selection using Support Vector Machines (SVM), and validated using k-fold or traditional validation on independent datasets. Results: The best human signatures we derived exhibit k-fold validation accuracies of up to 98% ( DDB2, PRKDC, TPP2, PTPRE, and GADD45A) when validated over 209 samples and traditional validation accuracies of up to 92% ( DDB2, CD8A, TALDO1, PCNA, EIF4G2, LCN2, CDKN1A, PRKCH, ENO1, and PPM1D) when validated over 85 samples. Some human signatures are specific enough to differentiate between chemotherapy and radiotherapy. Certain multi-class murine signatures have sufficient granularity in dose estimation to inform eligibility for cytokine therapy (assuming these signatures could be translated to humans). We compiled a list of the most frequently appearing genes in the top 20 human and mouse signatures. More frequently appearing genes among an ensemble of signatures may indicate greater impact of these genes on the performance of individual signatures. Several genes in the signatures we derived are present in previously proposed signatures. Conclusions: Gene signatures for ionizing radiation exposure derived by machine learning have low error rates in externally validated, independent datasets, and exhibit high specificity and granularity for dose estimation.

Predicting Response to Platin Chemotherapy Agents with Biochemically-inspired Machine Learning

“…Log-loss models were initially constructed either by (a) a modified version of the misclassification-based method, or (b) using the BFS software described in Zhao et al (2018) 25 . Multiple signatures with low log-loss values can have different compositions, consistent with the possibility that there may be various diverse gene combinations that can give rise to signatures with satisfactory performance.…”

“…The log-loss minimized models generated by both methods had comparable compositions. The median GI 50 thresholded cisplatin model generated by the log-loss modified software [ ATP7B , BCL2L1 , CDKN2C , CFLAR , ERCC2 , ERCC6 , FAAP24 , FOS , GSTO1 , GSTP1 , MAP3K1 , MAPK13 , MAPK3 , MSH2 , MT2A , PNKP , POLD1 , POLQ , PRKAA2 , PRKCA , PRKCB , SLC22A5 , SLC31A2 , SNAI1 , TLR4 , TP63 ] shares 15/19 genes with the signature generated by the BFS software 25 [ ATP7B , BARD1 , BCL2 , BCL2L1 , ERCC2 , FAAP24 , FEN1 , FOS , MAP3K1 , MAPK13 , MAPK3 , MSH2 , MT2A , NFKB1 , PNKP , POLQ , PRKCB , SLC22A5 , SNAI1 ]).…”

“…The procedure is repeated until all genes have been evaluated. The gene subset with the lowest misclassification rate 6 or log-loss statistic 25 based on cross-validation is selected as the model for subsequent testing with patient GE and clinical data. K-fold cross validation of the misclassification-based models was performed using MATLAB software described in Zhao et al (2018) 25 .…”

“…SVMs minimized according to the log-loss classification function were also generated with both software described in Zhao et al (2018; uses multiclass compatible ‘fitcecoc’ function) 25 , and with a modified version of the software described above (using ‘fitSVMPosterior’ to compute posterior probabilities). Computed probabilities differ between ‘fitSVMPosterior’ and ‘fitcecoc’ (range: 0.02-0.04), thus the resultant models will differ between the two programs.…”

“…lower resistance thresholds), ‘fitSVMPosterior’ will warn that some classes are not represented, and thus those folds will not predict the labels for those missing classes. The log-loss models described in this manuscript were generated with the multiclass compatible ‘fitcecoc’ function software 25 .…”