Reply to Azuaje: Predicting effective combined therapies for heterogeneous tumors

We thank Francisco J. Azuaje (1) for his comments regarding our article (2). His letter offers an opportunity for further discussion on the clinical relevance of heterogeneity drawn from our study using a computational optimization approach based on random sampling of RNAi-based perturbations. Our objective was to learn how to improve the design of drug combinations for diverse tumor subpopulations. We experimentally modeled heterogeneity using RNAi, which is a versatile and robust tool, to generate potentially phenotypically distinct subpopulations. We emphasize, however, that our computational analysis and consequent conclusions are not restricted to only loss-of-function perturbations, as long as there are functional phenotypic distinctions (of relevance here is therapeutic response) among the various subpopulations (this can be at the genomic, transcriptional, or signaling level). Indeed, one can produce similar datasets by using perturbations other than RNAi (e.g., ORF cDNAs) to generate subpopulations and subject to a broad panel of targeted and chemotherapeutic treatments; the associated optimization models and analyses would be similar. The subpopulations are thus generalizable across diverse sources of heterogeneity. In addition, as first iteration of this work, we sampled subpopulation frequencies from a uniform distribution, but this can be extended to examine—for example, in a more clinically relevant setting—sampling based on a prior distribution of genetic alterations for a specific cancer type. Although these distributions may be extracted from public datasets (e.g., The Cancer Genome Atlas), information regarding therapeutic effects on subpopulations is less comprehensive for us to extend this in our present work. Nevertheless, we share interest in moving even closer toward such clinical relevancy with regards to sampling distributions, as we endeavored to describe in the Discussion section of our article (2).

With respect to Azuaje’s (1) comment on intratumoral biological redundancies, this possibility should not affect the general characteristics of drugs, which according to our sensitivity analysis revealed the most frequently used drugs as high in robustness and average efficacy. The underlying biological redundancies and their relevance to clinical heterogeneity represent a matter that goes back to our discussions above and in the text of our article (2).

In addition, we can point out here that we have already made primary data (efficacy and toxicity) available to the research community. On the website (as described in our article) that links to the GitHub webpage (https://github.com/boyangzhao/DrugComboOptHetTumor), the data can be found under matlab/datasets. We encourage the broader community to explore this and alternative approaches/assumptions.