Abstract In most standard cases, mutational burden is derived from samples where mutational calling is performed in tumor and matched normal specimens. While this is ideal to get a true sense of tumor-derived mutations, it creates limitations due to cost of extra sequencing or in many trials, unavailable for genomic profiling. Without this normal material, there are increased difficulties of eliciting the accurate somatic mutation calls away from natural variation. This is even further complicated by the fact that different tumor types have different spectra of mutations so a method built on an individual indication have risks of overfitting. We present here the results of our investigations of machine learning to build reliable germline variant estimates appropriate for use in calculating burden. In this draft, we examine two tumor mutation classification methods based on our latest TMB tool: random forest and neural network across multiple indications in TCGA. We compare calls made by these methods against burden scores that we have generated from our own paired tumor/normal method (sister submission) to determine accuracy. Using sampled tumor results from each indication, different models were built across both random forest and neural networks, showing accuracy in some indications, up to 98%. Looking to work to create and indication-independent classifier, validation results show we can reach an average of 89% accuracy where training is built across all indications. In all cases, cross validation was performed to prove the consistency of the classification method and test the algorithm on our latest tumor only data set against what was generated from paired results. Citation Format: Yiqing Tian, Wendell Jones, Natalie Mola, Victor J. Weigman. TMB calling in absence of matched normal: Learning germline behavior [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2018; 2018 Apr 14-18; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2018;78(13 Suppl):Abstract nr LB-380.