REFLECTions on Combination Therapies Empowered by Data Sharing

Molecular profiles used in clinical settings to guide treatment decisions have mostly focused on somatic alterations to DNA such as mutations, copy-number variations, and gene fusions. These alterations are curated individually from the perspective of therapeutic, diagnostic, or prognostic associations, as well as predicted or demonstrated biological effect (5). Although it is well established that such aberrations currently represent the cornerstone of precision oncology, these actionable events are limited in number, frequency, and predictive value. Importantly, the vast majority of clinically approved biomarkers do not enable the prioritization of drug combinations within the context of a cancer's overall molecular profile. REFLECT addressed this issue by complementing well-established DNA aberrations (named “master biomarkers”) with gene and protein expression data. These REFLECT signatures stratified patients based on recurrent events that could be linked to drug targets and mechanism of action. It is the combination of the master biomarkers and these multiomic signatures that allowed REFLECT to nominate drug combinations that are likely to yield additive therapeutic effect.

Although these initial results are promising, it is not clear whether both gene and protein expression are necessary for REFLECT to achieve useful prediction or whether gene or protein expression alone is sufficient to complement the DNA features. There are also other molecular profiles that have been generated in TCGA, including methylation and chromatin accessibility [Assay for Transposase-Accessible Chromatin using sequencing (ATAC-seq)], which have not been integrated in REFLECT. In clinical settings, however, the number of available assays is much more limited. Clinical assays are often targeted, limiting the number of genomic regions, genes, or proteins that can be profiled for a patient, which is likely to affect REFLECT's predictive value. Consequently, although REFLECT implementation may be straightforward in principle, there are multiple challenges that arise in clinical settings that are likely to affect the relevance of REFLECT predictions.

In their study, the authors elegantly showcased the value of REFLECT by comparing the therapeutic benefit of the drug combinations prioritized by the system with the remaining drug combinations tested in immortalized cancer cell lines, patient-derived xenografts, and patients. Although REFLECT clearly yielded benefit in this retrospective setting, it does not reflect how the tool will be used in clinical settings. In molecular tumor boards, patients are discussed individually, which implies that REFLECT predictions must be assessed at the patient level. How will the REFLECT rank drug combinations if multiple ones are deemed beneficial by the system? What is the quality of this ranking? How will the potential toxicities and the elected drug combinations be accounted for during the REFLECT-assisted treatment decision process? These questions will have to be answered before REFLECT can be used routinely.

The REFLECT computational pipeline is complex as it leverages multiple public pharmacogenomic resources and open-source algorithms to define coactionable targets and identify candidate drug combinations. Although this level of sophistication is a reflection of the molecular complexity of cancer and therapeutic responses, it also demands a high level of transparency for the oncology community to scrutinize, reproduce, and reuse the REFLECT system. The authors are to be commended for developing an open-source package (source code available on GitHub under a nonrestrictive license Apache 2.0 at https://github.com/korkutlab/reflect), providing a functional software environment using the Code Ocean platform (https://doi.org/10.24433/CO.1046303.v1) and a searchable web portal (https://bioinformatics.mdanderson.org/reflect/). There is still room for improvement, as the code released by the authors is restricted to the main REFLECT functionalities but does not allow for reproducing all the results reported in the study with a single command, therefore earning a Bronze reproducibility standard (6). It is clear, however, that the adoption of an open science approach enables the scientific community to better understand the inner workings of the system and for its members to test it in their own institutions. This is particularly important for rapid and wide adoption of REFLECT, which will allow for a more comprehensive assessment of its clinical relevance and limitations.

To fully capture the complexity of how therapies shape human cancer, multidimensional prediction models, such as REFLECT, will require comprehensive molecular profiles from millions of cancers linked to treatment outcomes. Although REFLECT relied heavily on preclinical models and available bulk genomics data, the concepts and framework introduced by this system represent a clear step toward a systematic approach to learning from multidimensional clinical and genomic data.

In addition to global growth and interconnection of clinical genomics programs, molecular profiling technologies continue to expand the breadth of analytes measured from an individual specimen as well as resolution, as single-cell analyses are now routine (Fig. 1). Gene panels continue to advance in size and increasingly incorporate RNA analysis (7). Comprehensive clinical whole-genome and transcriptome assays are now being launched by clinical laboratories (8). Predictions of drug response at the single-cell level have recently been reported, illustrating the potential for conceptual frameworks such as REFLECT to incorporate knowledge of intratumoral heterogeneity linked to predictions of response and resistance (9). Although clinical genomic data sharing to date has focused largely on DNA alterations from targeted panels, REFLECT has demonstrated the need to expand beyond targeted genomics to incorporate functional transcriptomic and proteomic measurements to guide the combined use of cancer therapies. However, the incremental value of each molecular assay remains to be demonstrated, particularly as the same cancer genome biology may manifest through multiple different biomarkers.

To move beyond preclinical models, there is an opportunity to transition from a static system to a learning system that can be updated with new evidence directly from patients treated with personalized predictions. Data sharing initiatives such as AACR Project GENIE (10) are an important step toward a global network of hospitals necessary for interchange of, and federated learning from, these data. This expansion raises additional challenges to monitoring that the original predictive value of the system remains the same or truly improves as the underlying data are updated. The need for continued validation will continue to drive the need for broad data sharing across a global network of hospitals generating and sharing comprehensive data using a common set of data standards usable by REFLECT and REFLECT-like systems.

T.J. Pugh reports personal fees from AstraZeneca, Canadian Pension Plan Investment Board, Chrysalis Biomedical Advisors, Illumina, Merck, and PACT Pharma and grants from Roche/Genentech outside the submitted work. B. Haibe-Kains is a paid consultant for and shareholder in Code Ocean Inc. and is part of the scientific advisory boards for IONIQ Sciences, Break Through Cancer, and CQDM.

T.J. Pugh and B. Haibe-Kains are supported by the Canada Research Chairs Program and the Gattuso Slaight Personalized Cancer Medicine Fund at the Princess Margaret Cancer Centre. T.J. Pugh is also supported by a Senior Investigator Award from the Ontario Institute for Cancer Research.