Once a drug is administered to a patient, it begins a complex journey involving diverse molecular interactions in numerous biological contexts (see Figure 1
). These molecular liaisons are not only essential for the arrival/removal of an active drug to the diseased system, but they are also directly responsible for the clinical effects of drug response, resistance and risk. System perturbations, primarily in the form of genetic variation, are the primary determinants of the differential clinical outcomes to drug treatment that was seen across patient populations. Understanding the nature of these effects is key to deriving clinical utility in the form of predictive biomarker tests. Given that cancer is the most genetically heterogeneous of all diseases, a key question facing oncologists today is “how can I use this knowledge to choose the treatment(s) that provides my patient with the highest likelihood of curative outcome and minimal likelihood of drug resistance and toxicity?” The answer to this question lies in understanding the relationship between the molecular composition of the patient/tumor and the existing evidence of its clinical utility. Today, efforts to decipher such relationships provide the cornerstones of innovation in personalized medicine.
Three major classes of predictive biomarkers are currently driving the emerging practice of precision oncology. The first relies on the biological principle of oncogene addiction [1
]—the observed dependency of tumor cells on cancer driver mutations (CDM’s). CDM’s typically usurp the mechanisms that control cellular signaling, which often leads to the constitutive activation of cellular growth and survival pathways. The functional dominance of such mutations ensures that CDM’s are not only prime targets for therapeutic intervention, but they also serve as excellent predictive biomarkers. In terms of aberration type, they are most commonly observed as single nucleotide variants (SNV’s), insertions/deletions (InDel’s), and fusion proteins (for a sample of biomarkers approved for clinical decision making by the US FDA, see [2
]). Beyond the tumor specific effects on drug response, such variants can also affect toxicity to anti-cancer therapies when present as germline polymorphism.
More recently, a second class of predictive biomarkers has appeared that acts by influencing the sensitivity of a tumor to immune recognition. Perturbations in mismatch repair genes, for example, often referred to as deficient mismatch repair (dMMR), can be deleterious to genomic stability and loss of function mutations in associated genes are known to correlate with high rates of tumor mutational burden (TMB) and microsatellite instability (MSI). Tumors that have higher levels of such mutations are believed to express more neoantigens—a type of cancer-specific antigen—that may allow for a more robust immune detection, which results in more durable response to immune checkpoint inhibitors, such as Pembrolizumab. The FDA approved Pembrolizumab that is based on cumulative data from five clinical trials totaling 149 patients, for tumors with high levels of MSI [4
]. This tumor agnostic approval of Pembrolizumab was the first of its kind for a predictive cancer biomarker, which is a precedent that has recently been followed by a second cancer agnostic approval of Larotrectinib in patients with solid tumors harboring NTRK
gene fusions [5
Perturbations in the DNA damage repair machinery also lend themselves to another distinct therapeutic strategy that is based on the concept of synthetic lethality. Synthetic lethality provides the third class of predictive biomarker and it is based on the fact that cell death is more efficiently induced by the simultaneous loss of function of two or more key players in cellular signaling pathways. The most successful example involves the clinical use of PARP inhibitors in patients with deleterious or suspected deleterious germline mutations in the BRCA1/2
]. Current work is investigating the predictive relevance of other genes involved in homologous recombination DNA repair, such as ATM
, raising hopes that greater predictive resolution can be achieved, particularly for hereditary breast and ovarian cancer patients, through detailed analysis of their broader mutational profile.
2.1. The Case for Multi-Gene Diagnostic Testing
Taken together, such diverse sources of predictive biomarker knowledge emphasize the importance of transcending traditional single gene testing in favor of more comprehensive diagnostic strategies. Multi-gene testing provides critical context to biomarker findings beyond the longer-term discovery implications and optimized use of valuable and often scarce tumor samples, which ensures that key information is not omitted from the treatment decision process. This is particularly important in cases where multiple independent biomarkers predict different responses to the same therapy. The literature is replete with a diversity of drug resistance mechanisms that can negate single gene diagnostic findings (for an overview of some of these mechanisms, see Figure 2
). For example, Erlotinib is associated with the prolongation of progression-free survival (PFS) in non-small cell lung cancer (NSCLC) patients with certain EGFR
mutations, but with a shortening of PFS in patients with concurrent KRAS
Given such knowledge, it would be medically shortsighted to test a lung cancer patient for mutations in only one of these genes. Another early example comes from melanomas, where BRAF
V600E mutations are common. Nazarian et al. [8
] have shown that certain activating mutations in NRAS
(e.g., Q61K) and other upstream components, such as PDGFR
β, can lead to resistance to drugs, like Vemurafenib and Dabrafenib.
Identifying such “predictive modifier” mutations is critical in understanding the clinical implications of functional dependencies amongst detected mutations. Not only is this important for the identification of potential resistance mechanisms, but it also opens important opportunities for the rational design of personalized combinatorial treatments. For example, the ability to probe both TMB status and identify CDM’s may allow for physicians to rationally define evidence-based combination therapies that involve both immune checkpoint inhibitors, such as Pembrolizumab, and targeted agents, such as Erlotinib. Such rationally designed combinations of immunotherapies and targeted therapies are likely to become some of the most exciting areas of progress in the utility of tumor NGS data over the coming years.
2.2. Cancer Gene Panels and Analytical Software
Today, the most common clinical approach for multi-gene testing involves targeted gene panels. Targeted panels range from small, typically tumor-specific gene panels that start from circa ten genes, up to larger more tumor agnostic panels covering hundreds of genes (see Table 1
). While whole exome sequencing (WES) provides a more comprehensive view on the tumor genome it suffers important practical limitations when compared to targeted panels. Targeted panels permit a higher sequencing depth and are still more cost-effective as well as offer faster turn-around times than WES. These simple features contribute significantly to their basic clinical applicability, a fact that was recently validated by the FDA-approval of two independent NGS tests for solid tumors: Foundation Medicine’s F1CDx (324 genes), which has been approved as a companion diagnostic, and Memorial Sloan Kettering Cancer Center’s (MSKCC) MSK-IMPACT profiling test assay (468 genes). One of the key advantages of larger gene panels is the tumor agnostic nature of their clinical utility, which typically allows for any solid tumor type to be profiled in a single diagnostic assay.
In terms of clinical practice, the process of NGS-driven treatment decision support (TDS) begins with a patient-physician consultation. Once NGS testing is warranted, a tumor sample and/or a matched normal sample, together with a test requisition form, is sent to a diagnostic lab for sequencing. The resulting sequence data is uploaded into a genome informatics/TDS software, such as MH Guide (www.molecularhealth.com
), which then identifies and clinically interprets genomic variants, such as SNV’s, gene fusions, InDel’s, copy number alterations (CNA’s), TMB, or MSI. These systems function by comparing the identified variants against curated knowledge bases containing comprehensive information regarding approved and emergent predictive biomarkers. The findings are summarized in a concise report, allowing for physicians to quickly review current knowledge around the treatment implications of a variant, and the evidence of its clinical utility in the current patient’s case. Such reports automatically present the most clinically relevant biomarker knowledge first, prioritized in terms of the degree of available evidence (Figure 3
Today, the process of automated clinical interpretation is made possible through expert curated databases of predictive biomarker information. The capture and curation of this knowledge by biomedical experts is critical to the entire process and it serves a number of important roles. It ensures that all information regarding the effects of a biomarker, including validation and reliability levels, is captured in an accurate manner. Such quality and evidence measures are directly reported to clinicians in the diagnostic report, ensuring that they are explicitly clear about how relevant a biomarker finding might be for each patient. The aforementioned MH Guide system for example, employs three quality levels of clinical validity:
QUALITY LEVEL 1: Clinically endorsed biomarkers (i.e., recommended by key opinion leaders such as ASCO, EORTC, NCCN, ESMO, FDA, EMEA). This is considered as the information of highest relevance.
QUALITY LEVEL 2: Clinically observed biomarkers (i.e., observations coming from real patient data but not yet endorsed by a key opinion leader—including prospective, retrospective, registries, meta-analysis or n-of-one studies). This is considered as information of high relevance.
QUALITY LEVEL 3: Pre-clinically observed biomarkers observed in vitro or in vivo animal model systems or predicted in silico. This is considered as information of lower/unclear relevance.
Other essential information may also be captured during the curation/reporting process, including:
Detailed variant descriptions—e.g., the type of genomic aberration (such as SNP, Insertion, or Deletion, etc.)
The relevant drug or treatment
The effect of variant on treatment responsiveness—i.e., response, resistance or toxicity
The quantity of effect—e.g., strong, medium, weak
The observation context (i.e., the disease, disease stage, or model system)
A link to the source information and a grading of its reliability
By covering a broad spectrum of evidence levels, from emergent to clinically approved, the TDS systems facilitate the increasingly daunting task of deciphering whether identified biomarkers and associated therapies are likely to benefit a given patient, especially where more than one biomarker is detected. Such “expert systems” make this information available to physicians, which enable them to ascertain whether a variant was previously observed in similar patient cases, together with the observed clinical effects of drugs used in these contexts. While information about emergent biomarkers is certainly of lesser clinical utility than clinically endorsed biomarkers, they can provide an instructive data-point for clinicians treating late-stage cancer patients where no obvious treatment options exist. In such contexts, the information is often used to prioritize the clinical trial options for the patient. Given that predictive biomarker information tends to be disease specific, individual variants can have differential levels of both predictive power and evidence as one moves from one indication to another. This highlights the fact that the more molecular outcomes data we generate around biomarkers, the greater the challenge for physicians to clinically interpret a patient’s mutational profile becomes. The flipside is that as more clinico-molecular outcomes data is generated globally, the more we will be able to capitalize on the power of AI. This is an evolution that will decrease our current use of expert systems, which allows us to glean new knowledge and treatment pathways from real world clinical and molecular outcomes data.
However, today there are several challenges that still impede the effective application of AI in TDS. On the one level, it is important that we remain objective about the current applicability of the technology in the treatment setting. While machine-/deep-learning has shown great promise in areas, such as drug discovery and the automated analysis of diagnostic images [22
], it is still too early for broad reliable application in treatment decision making across multiple cancer types. This point is best illustrated by the challenges that are faced by the IBM Watson technology in this specific setting, most likely due to the lack of available clinico-molecular outcomes data upon which the system could “learn”. Access to such data, preferably for millions of patients, still remains a prerequisite for the reliable application of AI in genome guided treatment decision-making and we still face considerable challenges, especially with respect to data access, standardization, and privacy. Such technology, regulatory, and ethical issues, together hinder the broad community-wide availability of such data, and until these problems are effectively addressed, expert TDS systems will continue to play a dominant role in providing—and proving—the clinical utility of the genome-driven approaches to TDS.