Modeling tumor heterogeneity and predicting effective combined therapies through computational optimization algorithms

Zhao et al. (1) report a computational method that models intratumor heterogeneity and its responses to drug combinations. The authors computationally simulated intratumor heterogeneity by mixing samples derived from a large collection of lymphoma cells, which were previously infected with specific single shRNA perturbations in vitro.

Each perturbation-specific sample represented a homogeneous cell subpopulation, whose responses (efficacy and toxicity) to individual drugs were known in advance. Using a multiobjective computational optimization technique (2), Zhao et al. (1) identify effective drug combinations for heterogeneous tumors, which are approximated as a mixture of cell subpopulations. Zhao et al.’s investigation is novel, sound, and elegant. However, I believe that two issues deserve further attention.

First, there is the possibility that Zhao et al.’s models overestimated the biological heterogeneity of the mixed samples simulated. A key question is how homogeneous their multisub-population tumors actually are, considering (i) the focus of their model on the expression silencing of single genes, and (ii) the random sampling of these subpopulations from a uniform distribution. It would be interesting to know how such cell-population diversity relates to other sources of heterogeneity, including those at the genomic, transcriptional, or signaling levels, as well as to specific phenotypic states (3). Moreover, the biological feasibility of the simulated mixed populations may be linked to mutually exclusive perturbations. Thus, there is a chance that the observed increase in the frequencies of selection of a few drugs, as reported in Zhao et al.’s (1) article, may be in part explained by underappreciated intratumor biological redundancies.

Second, although Zhao et al. deserve credit for publicly sharing code and analysis outputs, it would also be highly beneficial to make the primary (efficacy and toxicity) data, or at least a version of them, accessible to the community. This accessibility is important not only to support research reproducibility and alternative approaches to investigating drug combinations, but also to allow further theoretical investigations. The latter may include the exploration of alternative model assumptions, multiobjective optimization techniques, and parameter estimations.

I hope that the authors of this fascinating and nicely reported research will consider these observations to continue augmenting the impact of their research, as well as to enable related research efforts elsewhere.