In this work, we develop a systematic approach for applying pathway knowledge to a multivariate Gaussian mixture model for dissecting a heterogeneous cancer tissue. The downstream transcription factors are selected as observables from available partial pathway knowledge in such a way that the subpopulations produce some differential behavior in response to the drugs selected in the upstream. For each subpopulation, each unique (drug, observable) pair is considered as a unique dimension of a multivariate Gaussian distribution. Expectation-maximization (EM) algorithm with hill-climbing is then used to rank the most probable estimates of the mixture composition based on the log-likelihood value. A major contribution of this work is to examine the efficacy of the EM based approach in estimating the composition of experimental mixture sets from cell-by-cell measurements collected on a dynamic cell imaging platform. Towards this end, we apply the algorithm on hourly data collected for two different mixture compositions of A2058, HCT116, and SW480 cell lines for three scenarios: untreated, Lapatinib-treated, and Temsirolimus-treated. Additionally, we show how this methodology can provide a basis for comparing the killing rate of different drugs for a heterogeneous cancer tissue. This obviously has important implications for designing efficient drugs for treating heterogeneous malignant tumors.