Assigning evidence to actionability: An introduction to variant interpretation in precision cancer medicine

Modern concepts in precision cancer medicine are based on increasingly complex genomic analyses and require standardized criteria for the functional evaluation and reporting of detected genomic alterations in order to assess their clinical relevance. In this article, we propose and address the necessary steps in systematic variant evaluation consisting of bioinformatic analysis, functional annotation and clinical interpretation, focusing on the latter two aspects. We discuss the role and clinical application of current variant classification systems and point out their scope and limitations. Finally, we highlight the significance of the molecular tumor board as a platform for clinical decision‐making based on genomic analyses.


| PRECISION CANCER MEDICINE AND INCREASING DATA COMPLEXITY
Since cancer is a genetic disease, analysis of a tumor's genetic material with regard to sequence variants and structural alterations using nextgeneration sequencing (NGS) is playing a crucial role in modern oncology and pathology. 1,2 Additional layers of high-throughput molecular information, in the form of proteomic, methylomic, and transcriptomic data obtained from tumors or liquid biopsies, are increasingly used for diagnostic, prognostic, or predictive purposes in cancer patients. 3,4 Broad molecular characterization of cancer samples lays the foundation for personalized precision medicine, leading to a better understanding of tumor biology, and identification of targetable genetic alterations, thus improving patient outcomes. [5][6][7][8][9][10] The fact that an and/or are based on broad molecular profiling of tumors. With the increasing size of NGS-based panels and a transition to whole-exome or whole-genome sequencing (WES/WGS) of tumors, these complex biomarkers gain accessibility and relevance. In addition, several biomarkers can be combined to form a composite score with higher specificity and sensitivity to predict response or resistance to a specific cancer therapy.
The genetic and molecular context of a tumor entity has a significant modulatory influence on the prognostic and predictive value of MBs and has to be accounted for. 11,12 These newly developed complex and composite MBs need to be carefully evaluated and benchmarked in a clinical setting before they can be implemented in routine clinical practice.
WES and WGS are currently carried out predominantly in advanced stages of disease and in the context of precision oncology programs and prospective trials. [13][14][15][16] They inevitably lead to the discovery of numerous genetic alterations with the ensuing challenge to attribute them a biological and clinical relevance. While the clinical relevance for some MBs can be assessed quite easily with data from clinical studies restricted to one tumor entity (e.g., EGFR L858R in the context of a stage IV adenocarcinoma of the lung), the clinical relevance of the same mutation in a different tumor context (EGFR L858R in gastric adenocarcinoma) is less clear. This also applies, for example, to alterations of genes (beyond BRCA1 and BRCA2) that are involved in homologous recombination DNA repair and can indicate responsiveness to poly(ADP-ribose) polymerase (PARP) inhibitors or platinum-based chemotherapies. Increased reporting of such alterations due to broad molecular profiling of different cancer entities led to the development of several classification systems for assessing evidence and defining clinical relevance of MBs in cancer. [17][18][19] Challenges can also arise when a molecular analysis detects multiple therapeutic targets and MB-drug associations with different levels of evidence. Currently, molecular tumor boards (MTB) have an important role in such scenarios in order to evaluate relevance and possible interactions between individual MBs. 20,21 They can also put the molecular profile in the context of the specific disease and consider additional clinical factors. For MTB to work successfully, its members need to establish a harmonized approach for evaluating molecular alterations based upon a jointly agreed standard. While this overview is not intended to replace experience and self-study, we hope to provide interested readers from various fields (pathology, oncology, bioinformatics, and other involved disciplines) orientation in this quickly evolving field and equip them with a short guide for standardized assessment of MBs.

| STEP 1: IDENTIFICATION OF TUMOR-SPECIFIC AND GERMLINE VARIANTS AND BIOMARKERS
The identification of genetic variants and MBs obtained from NGSbased analyses is now a fairly well standardized process, 22 for which a number of open source and commercial bioinformatic algorithms and corresponding software solutions are available. These have their individual strengths and weaknesses, but generate comparable results providing an overall robust basis for further analysis. 23 This is followed by the primary annotation using bioinformatic tools and querying sequence and variant databases. This process leads to annotation of each variant with a set of metadata, such as the position of the variant in relation to the gene or the genetic region, the predicted cDNA and amino acid sequence (Human Genome Variation Society [HGVS] nomenclature) and its prevalence in different populations and variant databases. Although standardization of this step is feasible, the databases in use differ widely between institutions. This part of analysis should include technical quality filtering of variants for further classification and interpretation. It should not provide filtering with regard to functional or clinical relevance, nor any prioritization of the variants. In "tumor-only" sequencing (i.e., without parallel analysis of a germline sample), filtering data to eliminate common single nucleotide polymorphisms is also a task of the primary annotation. In WES and WGS, germline and somatic data are generated directly.

| GERMLINE ANALYSIS AND GENETIC COUNSELLING
A significant fraction (9%-12%) of cancer patients carries a germline mutation predisposing for different hereditary cancer syndromes [24][25][26] and a majority of patients expresses a clear preference to be informed about the results of germline analyses. 27 It is thus necessary to consider possible germline involvement of (likely) pathogenic variants that may cause cancer predisposition, both with regard to therapy as well as to preventive measures for the patient and affected family members. 24,28 Somatic and germline variants cannot be discriminated by  predictors (e.g., FATHMM, CADD) and also literature-based, manually curated databases (e.g., OnkoKB, CIVIC, JAX-CKB). 31 Automated annotation using these databases is hampered by historical inconsistencies in gene names and transcripts, inconsistent application of existing standards as well as non-HGVS annotation used by some bioinformatic tools. 32,33 In addition, the databases have to be maintained and versioned in order to be used for automated search and reporting. Therefore, additional manual search of the published biomedical literature is essential and yields important information beyond that obtained from search algorithms and databases in a majority of cases. This task is currently highly reliant on human resources ( Figure 1) and needs to be standardized and harmonized addresses the risk of phenotypically healthy carriers of a genetic trait, it requires a very high probability for the assessment of a possible F I G U R E 1 Process of MB classification and interpretation. Sequential steps in this process consist of quality-assured molecular diagnostics, standardized variant annotation followed by functional classification and evaluation of clinical significance. Clinical decisions are made by the MTB pathogenicity (with more than 90% certainty), as further screening, preventive measures, and in some cases, prenatal diagnostics, are based on it. 29 On contrary, oncogenicity assessment of somatic variants might tolerate a more progressive approach. Its aim should be the selection of variants effecting gene or protein function with a sufficiently high probability to inform the therapeutic decisionmaking process for cancer patients whose tumors are profiled in advanced stages of their disease. This standardized evaluation then enables functional annotation and clinical interpretation using the published classification systems.
Assessing clinical relevance strongly relies on published clinicogenomic datasets that often provide an inaccurate genetic profile of a tumor entity with missing data on rare variants. This might be best illustrated by the various classes of BRAF point mutations and associated clinical datasets. The most common V600 mutations belong to the BRAF class I mutations and induce a RAS-independent, constitutively active BRAF kinase, which is able to activate the signaling pathway as a monomer. Class I mutations can be targeted therapeutically using BRAF-inhibitors, which is the accepted standard of care for several cancer entities based on robust clinical data stemming from randomized phase 3 trials. Class II mutations (e.g., K601E, L597Q, and G469A) signal as constitutively activated mutant dimers.
BRAF monomer inhibitors, such as vemurafenib, are less effective in inhibiting protein function of these mutants. Data for (B)RAF-inhibitor or MEK-inhibitor efficacy in tumors with class II mutations is mostly preclinical or coming from individual case reports. In contrast, class III mutations (e.g., D594G or G466V) show impaired kinase activity.
They lead to dimerization between wild-type BRAF and CRAF, induce allosteric CRAF activation, 34 and thus make BRAF-directed therapy obsolete. Nevertheless, CRAF-inhibitors or MEK-inhibitors might be effective in these cases albeit robust clinical evidence is missing. 35 All three classes of BRAF variants are oncogenic and might represent an actionable target, but only BRAF V600 mutations have a validated clinical role. Functional classification of a MB has direct implications for the therapeutic decision, but the assessment of its oncogenicity is independent from published data on sensitivity to a specific inhibitor, that is, actionability and thus a more stable feature of a the MB. Therefore, we believe that functional annotation should be the first step of the clinical evaluation process, leading to structured interpretation of somatic molecular alterations in the MTB ( Figure 1).

| STEP 3: EVALUATING CLINICAL EVIDENCE OF BIOMARKER-DRUG ASSOCIATIONS
Identification and categorization of oncogenic MBs, as described in the previous section, is a prerequisite for further interpretation, but not sufficient to judge the potential therapeutic benefit. We focus here on selected aspects influencing the treatment decision, including the properties of a MB, quality and availability of precision oncology knowledge databases, evidence classification frameworks, and the integration of molecular data with additional clinical parameters.

| Precision oncology databases
Therapeutically relevant MB-drug associations can, in part, be found using the same databases as those used for the assessment of MB oncogenicity in Step 2. Standardization and harmonization of these knowledge bases is an ongoing effort, 33      In melanoma, the same MB-drug association would be classified as JCR Tier I-A, ESCAT I-A, and NCT m1A-Z ( Table 2).

| Variant classification systems
Intricacies of the classification are further demonstrated by data on vemurafenib in BRAF V600 mutated NSCLC. Vemurafenib is neither approved by the FDA nor by the EMA for NSCLC, but clinical studies and individual case reports support the effectiveness of the BRAF-inhibitor in this setting. 12,[42][43][44] Here as well, a distinction must be made between the common BRAF V600E and other BRAF V600 mutations. With BRAF V600E, the JCR leads to a Tier I-B recommendation (expert consensus without FDA approval). Evidence level I-C based on a basket trial 12 can be used for the ESCAT classification, since the only prospective study in this setting shows a low MCBS. 43 As this study presents clinically relevant response rates without survival data, evaluation according to ESCAT also allows the II-B classification. Judging the same evidence coming from a prospective, open-label study leads to m1A classification by the NCT framework.  Additional clinically relevant information can be obtained from these RNA-based assays in the form of (oncogenic) isoform expression, for example, FGFR2, that might be targetable with specific inhibitors. 45  TMB is used as companion diagnostics for immune checkpoint inhibitor therapy, while the HRD score is used as companion diagnostic for PARP-inhibitors. Implementation of these biomarkers into clinical practice is associated with scientific and technical challenges including:

T A B L E 2 Examples of divergent variant classifications based on JCR, ESCAT, and NCT classifications
• accurate definition of the biomarker in a manner that is optimized to answer the underlying clinical question, • implementation of technology and IT for high-throughput NGS and measurement of the biomarker in routine diagnostics, • standardization of laboratory and bioinformatic workflows to achieve valid and reliable measurement of the biomarker, • inter-assay calibration and optimization of cut-off points to translate the measured biomarkers levels into a clinical decision.
• inherent probabilistic nature, particularly when focused assays are being used that do not cover the full genomic footprint of interest.

| Tumor mutational burden
Originally, TMB was defined as the number of somatic missense mutations in the tumor exome. 48 58 Further studies evaluating the comparability of different platforms for routine diagnostics of HRD will be necessary. 59 Example: PAOLA-1 was a phase 3 trial that evaluated olaparib in combination with bevacizumab as first-line maintenance treatment for advanced high-grade serous ovarian cancer. 60 The addition of olaparib to bevacizumab provided a significant survival benefit, which was substantial in patients with HRD-positive tumors. In this study, One goal of precision oncology and molecular pathology should be the creation of databases and standards in order to quickly identify those therapeutic approaches that add value for patients and to separate them from ineffective or even harmful interventions. 63 In