Anna Georgieva

An integrated approach for inference and mechanistic modeling for advancing drug development

An important challenge facing researchers in drug development is how to translate multi‐omic measurements into biological insights that will help advance drugs through the clinic. Computational biology strategies are a promising approach for systematically capturing the effect of a given drug on complex molecular networks and on human physiology. This article discusses a two‐pronged strategy for inferring biological interactions from large‐scale multi‐omic measurements and accounting for known biology via mechanistic dynamical simulations of pathways, cells, and organ‐ and tissue level models. These approaches are already playing a role in driving drug development by providing a rational and systematic computational framework.

Nonuniform sampling for global optimization of kinetic rate constants in biological pathways

Global optimization has proven to be a powerful tool for solving parameter estimation problems in biological applications, such as the estimation of kinetic rate constants in pathway models. These optimization algorithms sometimes suffer from slow convergence, stagnation or misconvergence to a non-optimal local minimum. Here we show that a nonuniform sampling method (implemented by running the optimization in a transformed space) can improve convergence and robustness for evolutionary-type algorithms, specifically differential evolution and evolutionary strategies. Results are shown from two case studies exemplifying the common problems of stagnation and misconvergence

Using a Systems Biology Approach to Explore Hypotheses Underlying Clinical Diversity of the Renin Angiotensin System and the Response to Antihypertensive Therapies

In this chapter, we discuss how a systems biology approach can be used in drug development, by presenting an example of building and parameterizing a model of the renin angiotensin system (RAS) pathway. The RAS plays a pivotal role in regulating blood pressure (BP) and kidney function. We introduce a mathematical representation of the pathway in the systemic circulation and describe how we derived the parameters of the model from available clinical measurements and inferences about the homeostatic nature of the physiological system. The chapter includes a brief introduction to the implementation of RAS-modulating therapies, model validation and variability at the pathway level. The chapter also describes the process of extending the model from the systemic circulation to the kidney and the process by which the two models were connected. The work presented here is part of one regulatory pathway in a larger physiological model of BP regulation and renal function (in both healthy and disease states) that is used to generate and test hypotheses of the underlying physiology to investigate a range of clinical scenarios.

A SYSTEMS MODELING APPROACH TO UNDERSTANDING THE MECHANISMS OF RENAL PROTECTION OBSERVED IN THE AVOID STUDY: 5C.02

Objective: Use a systems-biology based modeling approach to identify and test mechanistic hypotheses that may explain the reductions in proteinuria and rate of GFR (Glomerular Filtration Rate) decline following the addition of aliskiren, a direct renin inhibitor (DRI), to Losartan, an angiotensin receptor blocker (ARB), as reported in the AVOID trial (24week study in diabetic nephropathy patients). Method: A RAAS-hypertension model was collaboratively developed by extending the well-established Guyton-Coleman model of blood pressure regulation to include 1) a detailed representation of the renin angiotensin aldosterone (RAAS) pathway and 2) a renal module that captures kidney damage caused by elevated glucose, renal angiotensin II, and blood pressure, leading to proteinuria and gradual decline in GFR. The model also includes a virtual patient (VP) population consisting of normotensives, hypertensives, and hypertensives with diabetic nephropathy. To parameterize the model, published and internal data on changes in RAAS biomarkers, blood pressure, GFR, and proteinuria following RAAS therapies were used. The AVOID study was simulated in the RAAS model using a cohort of diabetic nephropathy VPs selected to match baseline characteristics of patients in the trial. Results: The model was able to capture both the percent reduction in UACR (urinary albumin excretion rate) and differential rate of decline in GFR in the aliskiren arm compared to placebo observed in AVOID (Figure 1). In this poster we present model-based hypothesis testing that points to multiple mechanistic factors that can contribute to added benefits of a DRI in combination with ARBs, including the degree of baseline intrarenal RAAS upregulation, the relative effectiveness of ARBs in inhibiting intrarenal versus systemic AT1 receptors, and variation in the populations sensitivity to RAAS therapies. Conclusions: We have developed a RAAS model that can be used to test hypotheses, to predict outcomes of future clinical trials, and to identify patient populations most likely to benefit from RAAS therapies. Figure 1. No caption available.

Abstract 3774: Using hybrid population pharmacokinetic (PK)- scaled efficacy exponential tumor model to determine possible efficacious dose for an anticancer agent

Background: The proto-oncogene c-Met and its ligand HGF have both been shown to be frequently dysregulated in and correlate with poor prognostic outcomes in a number of human cancers (Christensen et al, Liu et al). The c-Met pathway can be activated by abnormal c-Met expression levels, activating mutations, gene amplifications autocrine/paracrine HGF stimulation, making inhibiting c-Met tyrosine kinase activity a viable therapeutic approach. INC280 is a highly selective small molecule c-Met inhibitor developed jointly by Novartis and Incyte for the treatment of solid tumors with activation of the c-Met pathway. Aim: The goal of the present work is to predict a possible efficacious dose for INC280 for treatment of solid tumors using a combination of a human population PK model with an efficacy exponential tumor model scaled from mouse. The use of population modeling that helps identify factors affecting drug behavior or explain variability within a target population has been endorsed by the regulatory agencies as a viable alternative to extensive clinical studies. A nice summary of tumor growth models and their use in Oncology can be found in Pharmacokinetics in Drug Development: Advances and Applications. Methods: To construct the human population PK model, data from 22 patients from a Phase I trial receiving multiple doses up to 150 mg of INC280 were used. An exponential tumor growth model was fitted to multiple dose data in GTL-16 MET-amplified tumor bearing mice. As the pharmacokinetics of INC280 in this mouse model was needed, a separate PK/PD model in mouse was first established using data from a single dose mouse study. The scaling from mouse to human was done by adjusting the model to account for reported values of tumor doubling in HCC human clinical trials as opposed to those observed in the mouse studies. Model parameters intrinsic to INC280 efficacy (Emax, EC50) were assumed constant across species. Results: A preliminary one compartment linear PK model described the data adequately. The dose predictions were conducted using the mean population PK parameters. Simulations were run for different initial tumor volumes, different EC50 values (to account for uncertainty for parameter estimate), and different tumor growth rates with doubling times of 30 days and 114 days (Kubota et al.). Efficacious dose was predicted to fall in the range of 100-400 mg total daily dose, depending on aggressiveness of tumor growth. Conclusions: The use of modeling and simulation to aid decision points in drug development is becoming increasingly common. Here we present an integrative modeling framework incorporating both preclinical and clinical data to provide dose estimation. The assumptions in the model are clearly stated and the model can be changed with the availability of new data. Citation Format: {Authors}. {Abstract title} [abstract]. In: Proceedings of the 103rd Annual Meeting of the American Association for Cancer Research; 2012 Mar 31-Apr 4; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2012;72(8 Suppl):Abstract nr 3774. doi:1538-7445.AM2012-3774

Computational inference of gene expression signatures differentiating between breast cancers responsive or resistant to letrozole.

Abstract #3024 Introduction: Genes which have been shown either to be differentially expressed or differentially changed between responders and non-responders to the neoadjuvant treatment with the aromatase inhibitor letrozole have been described. However, their network structure and interconnectivity have not been formally analysed.Bayesian network (BN) inference algorithms have demonstrated promise in the analysis of genomics data to recover fragments of biological pathways due to their ability to capture many types of relationships between variables and to handle noisy data.An inference score (4), normalized to be between -1 and 1, is a measure of the strength of the interaction (either positive or negative) between different nodes in the recovered network.
 Materials and methods: This computational study uses BN methodology to elucidate any possible network structure and refine a gene expression signature distinguishing responders from non-responders groups to letrozole treatment. It has been applied to 205 covariables (69 baseline expression, 45 day 14 expression and 91 changes in expression with treatment), identified via Random Forests as being differentially expressed between 37 tumours responding to neoadjuvant letrozole and 15 non-responders. The number of false positive relationships was limited by considering 4 different sub-sets of the original data 1) all 205 variables, but allowing for limited connectivity between them; 2) only the baseline variables 3) only the change variables with treatment and 4) only day 14 post-treatment variables. It was conjectured that the variables linking significantly to response status would appear in the overlap between lists 1) and 2), lists 1) and 3) and lists 1) and 4).
 Results: The analysis of the overall dataset highlighted the importance and the interconnectivity of ribosomal proteins. More specifically, the BN algorithm found two connected networks of 20 genes each with 13 and 14 ribosomal proteins and an average inference score of 0.65. In addition, BN links patient type (responder vs. non-responder) directly to genes that have been implicated in the natural history of breast cancer such as BCL-2 and KIAA0101 with inference scores of 0.4 and 0.45, respectively. The inferred network also contained a link with inference score of 0.4 between patient type and phosphatidylinositol glycan, whose oncogenic role has been suggested by others. The preliminary network analysis of the baseline covariables did not yield any meaningful pattern that differentiates the two patient types, except for the involvement of the cyclin E2 pathway with a “composite” inference score of 0.33.
 Conclusions: With initial encouraging results , further refinement of the BN analysis, as well as enrichment of the data should facilitate the discovery of a predictive baseline gene expression pattern. Citation Information: Cancer Res 2009;69(2 Suppl):Abstract nr 3024.