Scientific contributions

GR values: a robust quantification of drug sensitivity

Measurements of dose-response curves are fundamental to investigate the efficacy of therapeutic drugs and identify genes involved in sensitivity and resistance. However, one source of variability that strongly biases the traditional metrics of drug response, IC50 and Emax, and could potentially explain variations between studies is growth rate. To address this problem, I developed a new method to quantify drug response: the normalized growth rate inhibition (GR), from which I derived two metrics (GR50 and GRmax), that are independent of the duration of the assay and the division rate of the cell line in contrast to IC50 and Emax (see figure). As patient-derived cell lines generally grow slowly, the problem is exacerbated and thus the GR metrics are essential to properly compare results across established and patient-derived cell lines, which will be crucial for my research. Because the GR values are independent of the length of the assay, we can adjust the duration of the experiment to accommodate variations in growth rate between cell lines. I am now combining the GR with the Loewe additivity model to quantify and compare synergies across multiple cell lines. This new synergy metric will allow me to quantify differences in phenotype where, for example, two agents are cytostatic individually and cytotoxic when combined – something that other metrics fail to quantify. 

Relative Cell count vs. GR value

Dependence of drug response curves as a function of the number of divisions during the assay for a partial growth inhibition (model of constant drug effect on growth): IC50 and Emax increase with fewer divisions, but normalized growth rate inhibition (GR) metrics are constant (curves are collapsed).

Predicting drug sensitivity with pathway responses to ligands and kinase inhibitors

Most studies on drug sensitivity focus on genomic features. Here we investigated how measurements of the basal and perturbed states of signal transduction networks relate to drug response in breast cancer cell lines. We found that growth factor signaling predicts sensitivity to RTK and PI3K inhibitors. In particular, response to heregulin (HRG) is more predictive of sensitivity to inhibitors of the PI3K/AKT pathway than AKT activity. Further analyses showed that HRG response depends on the levels of ErbB2 and ErbB3. In this work, I developed new statistical approaches to overcome the problems of multiple hypothesis testing which will be essential for my future research. My results led to a better understanding of the relation between RTK signaling and kinase inhibitor sensitivity. In another project, I analyzed the transcriptional response of six cell lines to more than 100 different kinase inhibitors. By developing a new clustering approach, I found that the transcriptional signature of RTK inhibitors clusters with either inhibitors of the PI3K/AKT or the MAPK pathway depending on the cell line. Based on these signatures, I inferred the changes in transcription factor activity and identified synergetic drug combinations that we validated experimentally. To complete these projects, I mined public databases to quantify the polypharmacology of kinase inhibitors, which will be fundamental for modeling drug response.

Prediction of drug response based on signaling pathway responses

pERK level following HRG treatment (x-axis) and cumulative amounts of ErbB2 and ErbB3 (y-axis) overlaid with the GI50 value to the pan-PI3K/mTOR inhibitor GSK2126458 (color, red for resistant).

Determinants of fractional killing in response to TRAIL

The partial response of tumor cells to TRAIL and therapeutic antibody agonists is limiting the efficacy of this class of death-inducing agent as anti‐cancer drugs. Previous studies in the Sorger Lab raised the question of the role of non‐genetic cell‐to‐cell variability in incomplete killing by TRAIL. Based on single cell live microscopy data reporting caspase-8 activity and cell death in response to TRAIL treatment, I developed a model of initiator caspase dynamics. This model allowed me to identify a threshold in caspase activity that separates dying and surviving cells. Together with Dr. J. Roux, an experimental postdoctoral fellow, we used the model to predict how co-treatment with other therapeutic agents impacts on caspase-8 activity and cell death. We also identified treatment regimens that overcome overexpression of oncogenes FLIP-S/-L and BCL-2/-XL. 

Schematic of TRAIL response

Schematic of TRAIL response: reported caspase-8 activity (derivative of FRET reporter, dFR/dt) for surviving (in blue) and dying (in yellow) cells can be parametrized by k, the activation rate, and τ the time of activation. If activity reaches a threshold θT, the cell dies. Co-treatments (in red) affect k, τ and θT to increase cell death.

Quantification of the robustness of biological systems

During my PhD, I developed novel computational methods to quantify the robustness of biological networks. Robustness is a fundamental property of many biological systems and accounting for it in biological modeling is a challenge because it requires quantitative metrics capable of comparing robustness across models. By formalizing the evaluation of robustness, I was able to use it to discriminate between models and parameters, and in the absence of other criteria, decide that a model, or a parameter set, would be judged superior if it is more robust than other models to some class of perturbations. I applied my methods to different biological systems that are notionally robust or designed to be robust. For example, I identified causes of robustness in models of the cyanobacterial circadian clock, found the best network architecture to obtain stable oscillations in a model of the Suprachiasmatic nucleus, and designed models robust to external perturbations for synthetic circuits for pancreatic tissue homeostasis.

Modeling biological robustness

For given sets of parameters (dots) that fit the data, robustness to perturbations may vary (color intensity). For a biological systems know to be robust to two perturbations (red and blue), the parameters in both circles (top left) are the most plausible.