High-throughput gene expression profiling experiments, in conjunction with traditional dose dependent efficacy and toxicity assays, has emerged as a potentially powerful means for elucidating biological compound mechanism of action. Since the gene expression profiling data reflects the underlying network of causal relationships between gene products, it is possible to reverse engineer the functional relationships between genes and phenotype(s) as the underlying process of cellular response to drug treatments. This approach relies on representing observed entities as nodes and the interactions between them as edges in a network and then applying learning procedures to determine networks which best represent the data; and/or models that predict the causal relationships between molecular entities in the data. A new approach for determining compound mechanism of action is presented that utilizes a probabilistic modeling framework for reverse engineering an ensemble of models, from dose response molecular profiling and phenotypic data, and then systematically and rapidly interrogating these models via high-throughput forward simulations. A benchmarking study of our approach on synthetic systems suggests that a molecular profiling experiment with 40 compound dose response conditions may provide sufficient data to result in specificity and sensitivity of 99% of predicting quantitative changes in variables. To demonstrate the biological relevance of the approach, we apply it to predicting efficacy markers in cell lines with different genetic backgrounds for three kinase inhibitors.

Extended Abstract Status: Not Uploaded