It has been suggested that aggregate genomic measures such as mutational burden can be associated with response to immunotherapy. Arguably, the gold standard for deriving such aggregate genomic measures (AGMs) would be from exome level sequencing. While many clinical trials run exome level sequencing, the vast majority of routine genomic testing performed today, as seen in AACR Project GENIE, is targeted / gene-panel based sequencing.

Despite the smaller size of these gene panels focused on clinically targetable alterations, it has been shown they can estimate, to some degree, exomic mutational burden; usually by normalizing mutation count by the relevant size of the panels. These smaller gene panels exhibit significant variability both in terms of accuracy relative to exomic measures and in comparison to other gene panels. While many genes are common to the panels in AACR Project GENIE, hundreds are not. These differences in extent of coverage and genomic loci examined can result in biases that may negatively impact panel to panel comparability.

To address these issues we developed a deep learning framework to model exomic AGMs, such as mutational burden, from gene panel data as seen in AACR Project GENIE. This framework can leverage any available sample and variant level information, in which variants are featurized to effectively re-weight their importance when estimating a given AGM, such as mutational burden, through the use of multiple instance learning techniques in this form of weakly supervised data.

Using TCGA data in conjunction with AACR Project GENIE gene panel definitions, as a proof of concept, we first applied this framework to learn expected variant features such as codons and genomic position from mutational data (greater than 99.9% accuracy observed). Having established the validity of the approach, we then applied this framework to somatic mutation profiling data in which we show that data from gene panels can be calibrated to exomic TMB and thereby improve panel to panel compatibility. We observed approximately 25% improvements in mean squared error and R-squared metrics when using our framework over conventional approaches to estimate TMB from gene panel data across the 9 tumors types examined (spanning melanoma, lung cancer, colon cancer, and others). This work highlights the application of sophisticated machine learning approaches towards the development of needed calibration techniques across seemingly disparate gene panel assays used clinically today.

Citation Format: Jordan Anaya, Craig Cummings, Jocelyn Lee, Alexander Baras. Calibration Tool for Genomic Aggregates (CTGA): A deep learning framework for calibrating somatic mutation profiling data from conventional gene panel data [abstract]. In: Proceedings of the Annual Meeting of the American Association for Cancer Research 2020; 2020 Apr 27-28 and Jun 22-24. Philadelphia (PA): AACR; Cancer Res 2020;80(16 Suppl):Abstract nr 1093.