Abstract
A high tumor mutation burden (TMB) is known to drive the response to immune checkpoint inhibitors (ICI) and is associated with favorable prognoses. However, because it is a one-dimensional numerical representation of non-synonymous genetic alterations, TMB suffers from clinical challenges due to its equal quantification. Since not all mutations elicit the same antitumor rejection, the effect on immunity of neoantigens encoded by different types or locations of somatic mutations may vary. In addition, other typical genomic features, including complex structural variants, are not captured by the conventional TMB metric. Given the diversity of cancer subtypes and the complexity of treatment regimens, this paper proposes that tumor mutations capable of causing various degrees of immunogenicity should be calculated separately. TMB should therefore, be segmented into more exact, higher dimensional feature vectors to exhaustively measure the foreignness of tumors. We systematically reviewed patients' multifaceted efficacy based on a refined TMB metric, investigated the association between multidimensional mutations and integrative immunotherapy outcomes, and developed a convergent categorical decision-making framework, TMBserval (Statistical Explainable machine learning with Regression-based VALidation). TMBserval integrates a multiple-instance learning concept with statistics to create a statistically interpretable model that addresses the broad interdependencies between multidimensional mutation burdens and decision endpoints. TMBserval is a pan-cancer-oriented many-to-many nonlinear regression model with discrimination and calibration power. Simulations and experimental analyses using data from 137 actual patients both demonstrated that our method could discriminate between patient groups in a high-dimensional feature space, thereby rationally expanding the beneficiary population of immunotherapy.