As a rule, the benefits of chemotherapy are unevenly distributed among patients classified as having the same disease. Aside from quoting median survival and considering clinical variables (e.g. age, sex, performance status, and residual disease burden), for individual patients oncologists are at a loss to predict the magnitude of benefit of a particular chemotherapy strategy before treatment. Thus, clinical management follows the same assumptions that lead to clinical trial design and recruitment, namely that for the sake of study all m-MGMT patients are considered the same. But in time, patients classified as having the same disease have divergent outcomes, retrospectively making the assumption of sameness naive. By contrast, biosimulation promises to identify the anterior probability of treatment benefit thereby allowing physicians to tailor their clinical decision making for each patient.
In the case of TMZ and lomustine, 35–48% of m-MGMT patients would be predicted to get a major benefit from combination treatment compared to conventional TMZ or the alternative of single agent lomustine, perhaps justifying the inconvenience and extra toxicity. The remainder would be predicted to have either a negligible benefit or one small enough that it that might be difficult to justify considering the limited therapeutic index for some patients. With regard to the question of concurrent versus sequential therapy, some patients had relatively modest effects from each single agent, but apparent synergy when the agents were combined, thus favoring combination therapy over a sequential approach. In other circumstances, when the efficacy of both drugs was relatively high and the combination offered relatively little incremental benefit, sequential therapy would seem to be more appropriate. For some patients in this cohort, lomustine was predicted to have better efficacy as a single agent compared to TMZ, while the opposite was observed for another subgroup. One of the most surprising revelations of this study is that TMZ was predicted to cause disease progression rather regression for some and a decremental or anti-therapeutic effect when added to lomustine in 15% of patients. While these findings certainly reflect patient selection in this cohort, the results suggest caution is needed about generalizing positive conclusions from NOA-09 to individual patients.
Many other signaling pathways besides MGMT impact the efficacy and differential sensitivity to these agents suggesting that we should fight against the wish for simplicity and the preference for single biomarker rules. Biosimulation demonstrates that the complexity of signaling pathway dysregulation contributes to vast heterogeneity of drug response in this population. In the pathway analysis, a confluence of abnormalities involving NF-κB activation, hedgehog pathway activation, and MMRD generated a lack of TMZ benefit. On the other hand, HRD, base excision repair deficiency, MMRD, and epigenetic dysregulation contributed to lomustine’s predicted superiority for a subset of patients. Interestingly, MMRD, KMT2A-D loss, and/or EP300 loss of function mutations or deletions produced opposite effects conferring TMZ resistance and lomustine sensitivity in the same patient, thus accounting for lomustine superiority over lomustine for a subgroup of patients. By contrast, HRD and IDH1 mutations conferred sensitivity to both TMZ and lomustine. Interestingly, all the IDH1-mutated patients in this analysis were predicted to have a synergistic benefit from the combination treatment. However, for the majority of patients the co-occurrence of multiple genomic abnormalities creates an otherwise unpredictable mixture of drug responsiveness and resistance resulting in quite different treatment propositions in the clinic. As a result, the uniqueness of each GBM challenges the generalized conclusions about combination chemotherapy and creates a necessity for deeper more comprehensive molecular diagnosis that embraces the complexity of each patient’s disease and its potential divergence from generalized clinical trial conclusions.
Computational biological modeling affords not only an individualized, predictive measure of benefit, but also a quantitative one that simplifies our approach to the patient and overcomes the intrinsic complexity of deconvoluting multiple complex and potentially conflicting signaling pathway inputs. Remarkably, biosimulation embraces the uniqueness (N-of-1) and the complexity in each patient’s cancer. The current study predicts that immense variety exists in the benefit of TMZ and lomustine as single agents as well as in the magnitude of incremental benefit for adding the second drug. For the combination, biosimulation predicts the range of anticipated benefits extends from achieving more than 99% efficacy against the disease to nothing at all, thereby forging an imperative either to implement or shun the combination strategy. In effect, lomustine plus TMZ is essential medicine for some, but of no benefit at all for others, potentially much better or much worse than the average results reported in NOA-09. For a sizable population where the benefit of treatment falls between the extremes, biosimulation fosters an individualized benefit-risk definition that facilitates an informed discussion of therapeutic index, replacing bias and arbitrary opinion with personalized predictive scoring. Rather than risking under-treatment for the sake of safety or over-treatment so as not to miss the possibility for a superior outcome, biosimulation promises to empower neurooncologists to choose the most appropriate option based on the deepest understanding of each patient’s unique disease characteristics.
Our inability to biosimulate every patient with m-MGMT GBM represents a limitation of this approach. Approximately one-quarter of the patients in this study could not be modeled because of an insufficient number of genomic inputs, issues related to disease induction, or inability to achieve a steady state. Additionally, gaps exist in the model created by unknown consequences of genomic abnormalities which have yet to be elucidated by the research community. Implicitly, the success of any model is only as good as the completeness of the knowledge it is based on. While the CBM used in this study continues to evolve, we acknowledge the possibility that predicted outcomes of the patients who were not included in the analysis might have swayed the relative size of the subgroups identified in the study.
Another criticism of this work is that biosimulation awaits prospective clinical validation. This criticism conceals the viewpoint that aside from MGMT, the results of molecular profiling have no role to play in the management of the glioblastoma patient. Nevertheless, the proliferation of insights into the mechanisms of resistance and responsiveness, suggest this view is increasingly untenable. In fact, biosimulation is based on a scientific literature that includes many insights that are already commonplace and as such have the status of “common sense.” The association of IDH1 mutations with TMZ responsiveness, MMRD with temozolomide failure, and the favorable impact of HRD on chemotherapy response represent examples of this. Other scientific insights are less known in clinical circles, such as the impact of HOXA10 amplification, Hedgehog pathway or NF-κB activation on TMZ failure. The integration of both commonly known and unfamiliar insights into a biosimulation model bridges the chasm between the knowledge base of cancer biology and the clinic, and vastly simplifies the labor required to assess the significance of multiple genomic findings at the point of care. We do not disagree that the model used in this study should have prospective validation in a randomized cohort who received combination therapy versus single agent chemotherapy. Rather, we propose that there is an imperative to implement knowledge buried in the cancer biology literature. The consequences for diverse patient management when that knowledge is applied create urgency to complete this validation as soon as possible. If confirmed, neurooncologists would be able to inaugurate a new era of replacing the one size-fits-all decision making and the molecular naivete of the current status quo with a computational tool that maximizes the benefits of drug resources already at hand.
In the era of comprehensive molecular diagnosis, it is likely that more effective personalized clinical management can be accomplished. The findings of this study suggest that the best therapy available is not represented by a single approach applied to all patients, but is a matter of tailoring the treatment to the molecular underpinnings in each individual’s cancer. In the patient-centric world of molecular diagnosis, the best treatment is not a winner-take-all proposition, but a personalized approach derived from deep insights about the biology of each individual’s disease. The Bayesian approach to clinical decisions based on biosimulation creates the possibility of deriving actionable insight from comprehensive molecular diagnosis to fulfill the mission of giving the correct treatment to every patient. Lastly, molecular diagnosis often points to novel treatment possibilities for subgroups of patients. These include connecting cancers with HRD or STAG2 deficiency with PARP inhibitors, epigenetic dysregulation with histone deacetylase or EZH2 inhibitors, and DNA checkpoint abnormalities with ATM or ATR inhibitors, to name a few. The refinement of patient selection for the next generation of clinical trials based on deep molecular interrogation and comprehensive signaling pathway modeling promises greater success than has been part of the neuro-oncology journey so far. As such, biosimulation provides a practical solution for immediate patient care, but also the opportunity to evolve the armamentarium beyond conventional cytotoxic therapy and illuminate the next steps in the creation of individualized precision therapy for glioblastoma.