Abstract 191: A probabilistic analysis of somatic mutations indicates individual survival outcome classes with AUC near 1.00 for all tested cancer-drug combinations from TCGA and 4 immune checkpoint studies (all having ≥ 20 patients and an outcome ratio < 6)

Abstract

A new computational method to predict cancer treatment outcomes from somatic mutation data was tested. Using this method, treatment outcome success or failure for 78 different cancer-drug combinations (74 from TCGA & 4 from published immune checkpoint inhibitor studies) could be "predicted" for each patient with nearly perfect accuracy (AUC values from ROC curves at 1.000 or just below) based solely on individual patients9 somatic mutation information. Predictions worked for all examined cancer-drug combinations with information available for > 20 patients and with a treatment SUCCESS to FAILURE ratio between 1/6 and 6. Calculations disregarded outcome information about the patient for whom an outcome was being predicted, but so far only when calculating their own classification measure. More elaborate, independent calculations are being developed to eliminate the remnants of outcome information from one patient in classification measures calculated for other predicted patients, but these newer, more detailed calculations are ongoing. The methods avoid any (1) fitting of parameters to outcome or data, (2) use of linear algebraic methods, (3) determinations of scale factor values, and (4) use of some typically inaccurate types of experimentally estimated probability values. Instead, they use (1) more accurate metastatistics about an accurately determined type of probability value – the probability that the observed frequency of mutation for a gene differs from random in either separate population of the responder or of the non-responder patients – and (2) an analysis of some underlying causes of modeling bias – examining the sensitivity of how identifying non-random mutation frequencies can be perturbed by changes due to single patients. Statistics entailing extrapolation to an infinite sampling limit were avoided in favor of statistics more applicable to small finite samples. When one patient with a "known" outcome was deliberately varied, in a systematic non-random way, critical statistics exhibited consistent changes that differed depending on whether the varied patient belonged to the HIT or MISS outcome class and these changes remained consistent with outcome class when patients of "unknown" outcome were varied in a similar way. The analysis provided a quantitative mathematical explanation for why FLAG genes had appeared often in many GWAS and suggested that the mutational burden measure used often as a marker for checkpoint inhibitor studies might suffer from similar complications. Prospective studies are being planned. Citation Format: Jonathan Malcolm Friedman. A probabilistic analysis of somatic mutations indicates individual survival outcome classes with AUC near 1.00 for all tested cancer-drug combinations from TCGA and 4 immune checkpoint studies (all having ≥ 20 patients and an outcome ratio