Abstract B1-55: Predictive modeling of drug sensitivity in non-small cell lung cancer

Oncogenic lesions that arise during cancer progression provide an attractive target for chemical intervention strategies. The extreme molecular heterogeneity of tumors, however, makes it difficult to identify authentic intervention targets and to link patients to the most appropriate treatment. Current therapeutic strategies are also limited in the range of biology they are able to target. To confront this challenge, we have launched a full scale investigation to identify the genetic lesions that arise during cancer progression together with a computational approach to link novel compounds to these lesions. A panel of non-small cell lung cancer (NSCLC) cell lines was screened with over 200,000 synthetic chemical compounds in a tiered HTS approach and with 4358 natural products fractions isolated from marine species. Iterative two sample Kolmogorov-Smirnov tests and regularized linear regression procedures were then used to link drug activity to the complexity of cancer genomes by systematically assigning enrollment biomarkers to each compound from measures of gene expression, gene mutation, gene copy number, and metabolomics datasets. These algorithms accomplish several tasks. First, they allow us to fit a mathematical model to each compound, which will enable us to predict sensitivities in untested cell lines and tumors. In addition, biomarkers that are specifying sensitivity to a compound could hint at a common, perturbed upstream or downstream biology in the sensitive versus resistant cells that will then allow us to extrapolate mechanism of action hypotheses for each of the uncharacterized compounds. Using this approach, we have found that new genetic vulnerabilities that are not currently actionable can be linked to novel chemicals. Experimental mechanism of action hypotheses can be derived from these chemical/biomarker relationships and were subsequently validated for a subset. Notably, we are also able to parse KRAS mutant cancers into multiple, distinct molecular subtypes defined by co-occurring mutations along with matched compounds. This indicates that KRAS lung cancers are representative of diverse mechanistic subtypes. Collectively, we are using this approach as a data driven way to parse mechanistic lung cancer subtypes as well as identify a mechanistically diverse cohort of therapies capable of contending with lung cancer heterogeneity together with enrollment biomarkers that can specify sensitivity. Citation Format: Elizabeth McMillan, Christopher DeSevo, Nathaniel Oswald, Jordan Hanson, Pei-Hsuan Chen, Ralph Deberardinis, Noelle Williams, John MacMillan, Bruce Posner, Michael White. Predictive modeling of drug sensitivity in non-small cell lung cancer. [abstract]. In: Proceedings of the AACR Special Conference on Computational and Systems Biology of Cancer; Feb 8-11 2015; San Francisco, CA. Philadelphia (PA): AACR; Cancer Res 2015;75(22 Suppl 2):Abstract nr B1-55.