Identifying cancer drug MoAs and cell-line properties using signaling cascade models and Bayesian analysis: From throw-away experiments to persistent information
Linus Görlitz, Michael Block, Jörg Lippert
Competence Center Systems Biology and Computational Solutions, Bayer Technology Services GmbH, 51368 Leverkusen, Germany
Objectives: Clarifying the Mode-of-Action (MoA) of a drug often is a highly non-trivial task. This can mostly be done by performing specifically designed experiments. Depending on the complexity of the drug mthe required number of experiments to undoubtedly solve this question may be significant. Often performing significantly less standardized experiments would be sufficient for identification of the MoA if on the one hand prior experiments allowed a deconvolution of cell-line properties and drug induced effects and on the other hand a methodology was available which could make systematic use of this prior knowledge. Both is possible using Bayesian analysis of the data. We illustrate our approach by studying the MAPK signaling cascade [1] which is responsible for the regulation of gene expression and prevention of apoptosis and is often altered in cancer cells.
Methods: We use an established model [1], extended to be able to simulate two possible MoA’s (inhibition of RAS or inhibition of EGFR) in our modeling software platform MoBi®. Inter-individual variation of expression levels for RAS, MEK, ERK and EGFR is estimated using CGAP data [2]. Noisy measurements of activated ERK percentage for each of 8-14 individuals are generated (incorporating variability) for the basic model, two mutants showing overexpression of RAS or the EGF-receptor and for one RAS and one EGFR-inhibitor. Identification of cell-line properties is performed using a Bayesian/MCMC approach [3] implemented in R [4] & MoBi®. During the first iteration an uninformative prior distribution of the four parameters RAS, MEK, ERK and EGFR is used. The identified posterior distributions are used as prior information at the second stage where the strength of the inhibition is additionally identified (uninformative prior). The results are compared to using uninformative priors on all six parameters.
Results: Identification of MoA during the second iteration was never possible using uninformative priors. Application of the posterior distributions of the first stage allowed identification of the MoA for three cell lines. In one case the model could not predict strength but could rule out the incorrect MoA.
Conclusion: This method provides the possibility to use previously conducted experiments independent of their experimental design for new questions. Thus we have a generic method to continuously and systematically enhance knowledge about cell lines and answer relevant questions with possibly less experiments.
References:
[1] Hatakeyama M, Kimura S, Naka T, Kawasaki T, Yumoto N, Ichikawa M, Kim JH, Saito K, Saeki M, Shirouzu M, Yokoyama S and Konagaya A. A computational model on the modulation of mitogen-activated protein kinase (MAPK) and Akt pathways in heregulin-induced ErbB signaling. Biochem J. 373:451-463 (2003).
[2] Cancer Genome Anatomy Project. http://cgap.nci.nih.gov/
[3] Hastings WK. Monte Carlo Sampling Methods Using Markov Chains and Their Applications. Biometrika. 57(1):97-109 (1970).
[4] R Development Core Team, R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria (2009).