Feasibility and outcome of reproducible clinical interpretation of high-dimensional molecular data: a comparison of two molecular tumor boards

Precision oncology is expected to improve cancer treatment by taking into account molecular alterations [1]. Targeted treatment of well-defined molecular alterations has shown a clinical benefit, accordingly [2‐4]. The precision oncology process relies on many steps, including patient accrual, sample analysis, interpretation of results, and their clinical application [5]. The clinical interpretation of molecular data from tumor sequencing has been called the “bottleneck” of precision oncology [6]. Published guidelines address variant annotation and biomarker prioritization, whereas a complete interpretation workflow remains unstandardized [7‐11]. Multiple databases and search tools exist for the identification of biomedical literature to support biomarker associations [12‐14]. Yet, most databases contain non-overlapping literature [15, 16]. The vast biomedical literature and challenges in the variant interpretation process lead to inter-interpreter differences even with limited molecular data [17]. The use of multi-gene panels to simultaneously interrogate multiple genes of interest has become a standard in most cancer centers. In addition to gene-panel diagnostics, even more comprehensive analyses of genome or transcriptome are increasingly used [18‐21], thus further raising dimensionality and therefore complexity of the resulting data. These analyses hold promise to identify targetable alterations in patients where no well-defined biomarker will be identified by more targeted analyses. For unselected patient cohorts, a clear benefit with precision oncology has so far not been shown in prospective studies [22‐24]. These results contrast with the clear benefit of precision oncology strategies in patients with well-defined molecular alterations within a specific tumor histology [3, 4]. A few more recent trials have shown efficacy of biomarker-directed therapy even in histology-agnostic trials [2, 25]. In order to further expand these benefits to a larger and unselected patient population, reproducible and evidence-based strategies for the clinical interpretation of complex molecular data are required. In order to identify challenges for harmonized workflows, we compared treatment options identified by two independent molecular tumor boards for patients with identical exome and transcriptome sequencing, e.g., high-dimensional, molecular data.

Specific attention and additional research is required to improve the clinical annotation of molecular data, which is still unstandardized and inconsistent between molecular tumor boards [17]. The integration of high-dimensional molecular data can be expected to further complicate clinical interpretation but no real-world data currently exist on the scale or clinical impact of this scenario. Alleviating this imminent “bottleneck” [6] is expected to improve clinical decision-making and the prospective design of clinical trials for precision oncology.

In this work, we retrospectively analyzed the clinical interpretation of identical and high-dimensional molecular alterations of 40 patients by two molecular tumor boards that were prospectively sequenced within the DKTK-MASTER-program. This analysis yielded a mean overlap coefficient of 66%. Previous studies of recommendation heterogeneity yielded overall agreement rates between 40% [17] and 86% [26]. However, major differences between the studies have to be taken into account, when comparing these data. The average number of molecular alterations per patient was 8 in the study by Rieke et al. and 2.6 in the study by Koopman et al. In the here presented study, more than 300 alterations per patient, identified by whole-exome and RNA-sequencing, were clinically interpreted. Furthermore, the study by Koopman et al. assessed clinical interpretation in well-defined clinical situations of melanoma and NSCLC samples. The DKTK-MASTER study was designed to include patients without established treatment options, which is highlighted by the large number of patients with neuroendocrine neoplasms in this cohort, for which no standard of treatment exists in later lines of systemic therapy. Considering these molecular and clinical challenges in an unplanned retrospective analysis of an experimental sequencing study, an overall agreement rate of more than 60% should be viewed as highly encouraging. These results could be mediated by similar MTB workflows, with an interdisciplinary MTB discussion after prior manual annotation of molecular aberrations with evidence levels, following a structured search of databases. Further improvement could therefore be expected with ongoing harmonization efforts.

This analysis allows for a detailed look at challenges with the interpretation of complex molecular data for these efforts. Generally, more heterogeneous recommendations were found in the setting of biomarkers with low evidence levels and combination therapy, probably due to the wealth and heterogeneity of preclinical studies [15], differences in their appreciation [11], and a lack of controlled trials for combination targeted therapy due to combinatorial complexity [27]. More data for the rational use of drug combinations for precision oncology is warranted. Additionally, lower agreement rates were identified for gene expression data. This is probably caused by a lack of clinical studies for most of these alterations, which are additionally not well-defined because of a lack of established cut-offs. Despite these more obvious challenges, perfect matches were also not achieved for SNV. Most SNVs annotated in this cohort were not identified in well-characterized genes and therefore created challenges in their appreciation as potential biomarkers.

The agreement rate in our study was lower for identified predictive biomarkers than the corresponding treatment option. This can be explained by the frequent identification of several alterations that point towards identical treatment recommendations (e.g., BRCA mutation, ATM underexpression, FANCI, FANCA deletion for PARP inhibitors in the same patient) but were not always all named by both tumor boards. Yet, agreement rates of predictive biomarkers were significantly associated with agreement rates of treatment options. Therefore, the structured identification of potential (predictive) biomarkers from molecular data remains key to the identification of treatment options. Efforts for a harmonization of databases is expected to greatly aid with this [16]. Interdisciplinary teams will be increasingly important to extract the maximum of clinically relevant data from complex molecular profiles.

Some limitations of this study should be considered: Patients were discussed in parallel by the molecular tumor boards in Heidelberg and Berlin only in the beginning of the MASTER program, thus limiting the number of patients. Additionally, the recommendations reflect MTB practices of the inclusion years 2016–2018, which have evolved significantly since then, thus possibly underestimating current concordance rates. This limitation is important, since great efforts have been put into biomarker annotation and database development, and since efforts are ongoing [16, 28]. A follow-up analysis of the same patient data was not feasible within a blinded MTB setting. However, a more superficial second visit of the data suggested higher agreement rates between participating tumor boards.

The availability of new data and treatment options will eventually bring established targets into clinical routine and away from a more experimental MTB setting. An analysis of unclear and complex data can therefore be expected to remain a problem for precision oncology. Therefore, this analysis also has several strengths: Identical molecular data were available, limiting the effect of potential confounders. The retrospective, real-world design of this study reduced the risk of bias. Furthermore, the interpreted molecular data are highly complex, incorporating WES and RNA-Seq data, and thus allow for an analysis of a wealth of biomarkers and treatment recommendations even in the setting of two participating MTBs. High-dimensional molecular data beyond targeted sequencing are increasingly incorporated into precision oncology [18‐21]. The added value of whole-exome/genome and RNA-sequencing data is an important clinical question. In our cohort, 75–90% of 67 SNV or Indels that were used as predictive biomarkers would have been identified with large multigene-panels (e.g., MSK-IMPACT, TSO500). Yet, most structural/transcriptomic biomarkers that led to treatment recommendations or further supported them would not have been identified and the patient number is too small for a definitive analysis of the clinical impact. Therefore, this question should be answered in larger and ideally prospective trials. In the WINTHER trial, no improvement of outcome but a numerically better clinical benefit ratio could be shown with transcriptomic profiling [18].

An improved outcome could be shown in the I-PREDICT and WINTHER trials, as well as a real-world data analysis, for patients that received treatment that was better matched to their tumor’s molecular profiles [18, 23, 29]. In our retrospective analysis in a small cohort of patients, concordant treatment recommendations were also associated with an improved overall survival. Since patients with more well-defined molecular alterations were more likely to receive reproducible recommendations and effective treatment, this finding might reflect that patients receiving treatment that is well-matched to their unique molecular tumor profile achieve a greater clinical benefit—thus mirroring the results from the I-PREDICT and WINTHER trials in a setting with complex molecular data. Yet, this analysis should be viewed with caution given the low number of patients, retrospective analysis, and selection bias. Prospective trials with a focus on interpretation practices (such as matching scores and reproducibility of treatment options) are warranted to validate these exploratory findings.

Reproducible and evidence-based interpretation of complex molecular data is feasible with the use of structured workflows. Additional attention is required for the interpretation of data beyond genomic analyses and biomarkers with preclinical evidence levels as well as for the introduction of rational combination therapies.