Pharmacogenomic studies offer insight into how genomic markers correlate with drug response and hold potential for developing effective personalized cancer treatment regimes. Cell viability dose-response curves are often summarized by the area-over-the-curve (AOC) and used to determine which cells are sensitive to a given treatment, thereby discovering markers of response. However, uncertainty surrounds pharmacogenomic findings for several reasons: 1) A lower-than-expected level of agreement between any two pharmacogenomic studies when comparing similarly-assayed cell lines and drugs, 2) how inter-study correlation is measured, 3) the sensitivity measure reported and subsequent threshold used to call sensitive versus resistant cell lines, and 4) how the in vitro models recapitulate in vivo patient outcomes. For the F-phase I develop a statistical model to address the first three of these issues and apply the model to cell lines from the Cancer Cell Line Encyclopedia (CCLE) and Genomics of Drug Sensitivity in Cancer (GDSC) datasets to show that agreement is improved by conditioning on the posterior probability that a given drug is targeted and selecting the appropriate correlation metric for the underlying distribution of the data. I quantify uncertainty in calling sensitive versus resistant cell lines and report measures of inter-study agreement between other publicly-available pharmacogenomic datasets for common drugs and cell lines. I next provide findings from a collaborative effort assaying patient-derived organoids lines for individuals diagnosed with pancreatic cancer. I use my model to assess AOC and show through exploratory data analysis why the measure needs to be adjusted for growth rate confounding. I identify issues resulting from samples with low growth rates and motivate an empirical Bayes growth rate estimator. I show how the estimated growth rates compare with the other measures of sensitivity to illustrate the validity of the approach. I assess patient outcomes data with the organoid assays and report associations with patient time on treatment to motivate my second aim and the primary research direction for the K-phase of the fellowship. I show low correlations between the in vitro drug sensitivity measures and the in vivo patient sensitivity proxy measure. Given the low correlations I describe alternative measures of patient sensitivity and detail the first direction for my post-doctorate research. To that end I describe a deep learning model I am developing which takes as input clinical reports data and outputs cancer progression/response predictions. I use these labels to train a deep classifier which takes as input unstructured clinical notes and outputs a prediction of cancer progression along with an associated estimate of uncertainty. The input text is transformed into a vector embedding using word2vec and uncertainty is quantified via Bayes-by-backprop. Upon successful model training I will investigate using the predicted labels and uncertainty estimates to link to the pharmacogenomic in vitro sensitivity models to assess agreement between the models and patient outcomes.
Pharmacogenomic studies offer insight into how genomic markers correlate with drug response and hold potential for developing effective personalized cancer treatment regimens, yet uncertainties surround their findings for several reasons. I develop a statistical modeling framework to quantify the posterior probability of sensitivity, study agreement, and associated uncertainty, then apply the model to pancreatic cancer-derived organoid lines. I illustrate confounding with cell growth rate, develop a shrinkage estimator for growth-rate normalized sensitivity measures, and lastly present associations between in vitro pharmacogenomic models and in vivo patient outcomes to motivate the need for both in vitro experiments and in silico methods which accurately recapitulate patient outcomes.
