Pharmacogenomics in Early-Phase Oncology Clinical Trials: Is There a Sweet Spot in Phase II?

Many current oncologic drug development trials consider possible genetic factors governing a drug's response or toxicity. Originally centered on single genes (pharmacogenetics), pharmacogenomics now integrates information from the entire genome and can encompass both germline differences, such as single-nucleotide polymorphisms (SNP) and acquired changes (tumor mutations; refs. 1, 2). Pharmacogenomics focuses specifically on predictive genetic biomarkers of outcomes from pharmacologic interventions, as opposed to disease genetics, in which variants describe disease subtypes or prognostic information.

There is a clear incentive to incorporate pharmacogenomic factors during the drug development process, namely, to potentially characterize subpopulations within a disease that may uniquely respond to a given drug, even if the larger population does not. Conversely, identifying genetic predisposition for drug toxicity can improve overall tolerability of a drug through pharmacogenomic patient selection.

Although the impetus for including pharmacogenomics in oncology is clear, the development phase in which pharmacogenomics should receive attention is less clear. Should phase I and II trials routinely explore pharmacogenomic biomarkers? And, despite the sample size advantage of phase III trials, is phase III often too late? In this article, we examine the recent growth of pharmacogenomics in oncology trials, especially in early-phase studies, and examine several critical questions. We aim to show that phase II trials offer a potential “sweet spot” for routine incorporation of pharmacogenomic questions in oncology.

To assess the prevalence of pharmacogenomics in oncology drug development, we queried MEDLINE clinical studies of “antineoplastics” combined with the search term “polymorphism.” Limiting the results to studies since the year 2001 (after publication of the first human genome; refs. 3, 4) and to studies listed as phase I, II, or III clinical trials plus observational studies and meta-analyses, we found that the number of oncology studies incorporating pharmacogenomics is rapidly increasing (Fig. 1A).

To appropriately consider the timing of incorporating pharmacogenomics in oncology drug development, it is informative to evaluate phase III clinical trials first. Phase III clinical trials typically include large numbers of patients, facilitating discovery or validation of pharmacogenomic markers. We would, nevertheless, argue that phase III is too late to begin to consider pharmacogenomic effects. As an example, the role of KRAS mutation in anti–EGF receptor (EGFR) therapy in colon cancer is illustrative. Although anti-EGFR therapy was first approved by the U.S. Food and Drug Administration (FDA) in 2004 for use in patients with colon cancer, it was not until 2009 that a label change was instituted, stating that patients with mutated KRAS had no benefit from the addition of anti-EGFR therapy. The FDA label justifying the necessity of KRAS testing to inform anti-EGFR prescribing in colon cancer includes 7 different trials as supporting evidence; 6 of the 7 are phase III clinical trials, and 5,657 patients were included (5). Of those, only 3,287 patients had a KRAS result, and of those, 651 KRAS-mutant patients received anti-EGFR monoclonal antibody therapy, in retrospect, inappropriately (5). Clearly, many more patients were treated inappropriately in off-protocol situations during the 5 years preceding the label change. If earlier-phase trials had been able to more widely consider KRAS status, it is possible that fewer phase III trials would have been needed to show its importance. Not only could this have saved resources, but fewer patients would have been treated with an ineffective drug.

Consequently, we hypothesize that phase III is simply too late to consider a pharmacogenomic marker for the first time. Two other prominent oncology drug development examples recapitulate this point: Gefitinib failed commercially, in part, because predictive markers were not considered prior to the phase III studies (although in this case, the most useful predictive markers were not discovered until after the phase III studies were completed; refs. 6, 7). In contrast, for trastuzumab (8), the pivotal phase III study was limited to patients with HER2 overexpression (9). Subsequent analyses have shown that if investigators, instead, had chosen to treat all patients with breast cancer and simply analyzed the HER2-upregulated cohort separately, a much larger (and more expensive) study would have been needed to show the effect (10).

Therefore, although large phase III studies may, in some cases, provide the only detectable signal of a pharmacogenomic marker's relevance, phase III studies are confirmatory trials of a drug in selected patients. The evaluation of pharmacogenomic biomarkers should be no different. We believe that informative pharmacogenomic questions can, and should be, asked earlier in drug development.

If phase III is too late to begin consideration of pharmacogenomic biomarkers, how early in drug development can pharmacogenomics be incorporated? We examined our prior MEDLINE search for only phase I and II clinical trials and found 83 such studies over the past 10 years (57 phase II trials and 26 phase I trials). We first considered whether phase I trials are too early to informatively assess pharmacogenomic biomarkers. In fact, we found that only 19% of the phase I trials reported a “positive” pharmacogenomic finding. In contrast, for phase II studies, 70% identified a “positive” pharmacogenomic marker (Fig. 1B). Although publication bias may preferentially affect phase II publications, these data suggest that phase I is simply too early to be routinely informative for evaluating a pharmacogenomic marker. This is likely due to the small size of phase I studies, as well as the heterogeneity of doses used and patients entered. In phase II, at least by investigator-assessed terms, pharmacogenomic markers are commonly able to be associated with some clinical phenotype or endpoint. As a final comparison, the “success” of pharmacogenomic publications derived from phase III trials that met our search criteria is also shown in Fig. 1B. The proportion of positive studies is extremely high, surely representing some publication bias, because the publications in this group were almost always separate, secondary analyses of a given biomarker or biomarkers from a previously published phase III trial. In contrast, the above-examined phase II pharmacogenomic publications tended to typically report both the clinical results of the trial and the secondary pharmacogenomic findings within the same publication.

One caveat to the above generalizations, however, is that the utility of testing pharmacogenomic biomarkers in phase I may, in part, be affected by whether the biomarker is expected to correlate with drug response in comparison with toxicity. Specifically, for toxicity pharmacogenomic biomarkers, phase I may have some unique advantages because of detailed pharmacokinetic analyses and toxicity monitoring. In contrast, for response pharmacogenomic biomarkers, phase I studies in which tumor response rates are typically <5% would likely be low yield. Additionally, in some cases, the decision about when to incorporate pharmacogenomic investigations may depend upon development of knowledge for specific factors, such as the mechanisms involved in drug disposition and drug action, polymorphisms in those pathways, the frequency of polymorphisms, and whether the polymorphisms of interest have shown an impact on variability in drug exposure or response for other drugs.

These nuances notwithstanding, the above data prompt us to hypothesize that phase II may be a “sweet spot” for clinical incorporation of pharmacogenomics in oncology drug development. Generally speaking, phase I trials seem to be too early for routine discovery success, too thin by sheer patient numbers and complicated by the problem of dose and patient heterogeneity, and phase III can be too late (in which patients may be inappropriately treated). Just as for all other aspects of understanding drug activity, the appropriate “learning” phase is phase II.

To examine the advantages and challenges of incorporating pharmacogenomics into phase II studies, it is informative to examine a specific case. Of the 57 published phase II studies from our search, only 1 study stated its primary objective as to “identify potential polymorphisms related to a better and safer outcome” (11). We thought that this might be an especially interesting example to examine in depth.

The study was a phase I and II trial with a 27-patient dose-finding (phase I) portion, followed by a 60-patient phase II portion. It examined oxaliplatin/irinotecan/capecitabine as first-line therapy in metastatic colorectal cancer. In the phase II portion, the investigators examined 13 selected polymorphisms in 10 genes that were purportedly related to the study drugs' metabolism or efficacy (11). Only 1 of the 13 polymorphisms (located in GSTP1) was associated with response (P = 0.007); the other 12 did not meet the prespecified P < 0.05 significance level (11). The authors did not consider multiple testing, although they state that the results are exploratory. The issue of potential false discovery due to false association is an issue of paramount importance with pharmacogenomic data and is extensively considered elsewhere (12–14). Nevertheless, the GSTP1 SNP association from this study was, at least, a very interesting exploratory finding.

We draw 2 conclusions from this example: 1 is that even phase II studies that posit pharmacogenomic aims as their primary purpose do not always adequately consider statistical issues. This conclusion reinforces the idea that any finding in phase II needs confirmation. On the other hand, it is interesting that 1 of the chosen polymorphisms in this study has a potential association with response. This finding makes it worthy of additional follow-up, and in that sense, the trial achieved its stated pharmacogenomic purpose.

In many phase II situations, this will be the case. The smaller size of phase II trials means that pharmacogenomic endpoints will necessarily be exploratory or hypothesis generating. This factor will be especially true if a variant is expected to have a modest impact (as is probably the case for most variants affecting complex genetic traits; ref. 15) or if multiple variants are tested (12, 13). However, statistical power limitations are not, in our view, prohibitive, because the phase II setting may be the first opportunity (at a fixed drug dose) to extensively evaluate pharmacogenomic variants as they may relate to a drug's development (Table 1).

Use of randomized phase II trials has been increasing (16, 17), and randomized designs intensify the importance of incorporating pharmacogenomic markers. Even if a follow-up phase III study is necessary, a pharmacogenomic marker that is not at least explored in a randomized phase II study is much less likely to be considered in the follow-up study. Randomization additionally offers a clear manner to specifically evaluate the marker with respect to drug (predictive), rather than just the disease (prognostic). This key aspect of randomization is essential for a genetic marker to be truly identified as a pharmacogenomic marker. Many published pharmacogenomic studies fail to include an untreated group and, thus, fail to convincingly show that the marker is truly predictive (18). The above study (11) is an illustration of this limitation, and it is, therefore, impossible to determine whether the GSTP1 marker is predictive of treatment response or simply prognostic for the natural course of the disease.

The randomized discontinuation trial (RDT) design (19) offers a unique and superb opportunity for incorporating pharmacogenomics. In RDTs, all patients are initially treated with the same drug and then “self-select” for continued therapy (if responding) or for randomization against placebo (if disease is stable). The design inherently invites the question of trying to better understand the response heterogeneity seen in patients with the same empiric “disease.” Pharmacogenomic factors likely will eventually explain a portion of this disease-response heterogeneity in oncology, alongside other acknowledged, important sources of interindividual variability including drug–drug interactions, drug resistance, excretory organ function, compliance, and comorbidity. Nevertheless, without incorporation of pharmacogenomic questions into RDT and other randomized phase II trials, the molecular underpinnings of response variability will likely continue to remain unanswered.

It is easy to appreciate the importance of discovering response-predictive pharmacogenomic markers in the phase II setting. Tumor-specific pharmacogenomic changes certainly fall into this category, and well-described dramatic examples (like EGFR and BRAF mutations predictive of therapy benefits in lung cancer and melanoma, respectively; refs. 20, 21) have reinforced this theme. Germline variants have also been associated with treatment response outcomes in oncology (22–24).

In contrast, the value of toxicity-predictive pharmacogenomic findings in phase II may not be as obvious, but should be considered just as important. Toxicity determinants are germline variants, predictive of an increased (or decreased) risk of an adverse event due to a specific drug or drug class. The 2 most well-characterized oncology examples of germline changes predictive of toxicity are UGT1A1 polymorphisms for irinotecan (25) and TPMT polymorphisms for 6-mercaptopurine (26). From the standpoint of pharmacogenomic discovery within phase II trials, one could argue that the probability of finding a positive pharmacogenomic association for a given toxicity will be more likely if the incidence of the toxicity is more common than the expected incidence of response. That scenario is quite common in oncology drug development. The ultimate value of toxicity-predictive markers is improved identification of patients at highest risk for adverse events, which is useful when deciding between several apparently equivalent therapies and for drug avoidance in patients with specific comorbidities. For pharmaceutical development decisions, pharmacogenomic risk stratification could mitigate concerns about moving a drug forward to phase III (or for approval) if the highest toxicity-risk patients could be preidentified and excluded.

In addition to the above considerations, one further specific challenge incorporating and interpreting pharmacogenomic markers in the phase II setting deserves mention: A potentially interesting or relevant pharmacogenomic polymorphism may be generally uncommon in the patient population being studied. This consideration often receives little attention, but the prevalence of the risk allele for a given polymorphism can vary considerably by ethnicity (27), or it may simply be uncommon in all populations. Falsely negative pharmacogenomic results in the phase II setting could arise because of this critical aspect. Therefore it is important to remember that when a potential pharmacogenomic marker is being selected for testing in a phase II study, if the number of patients with the susceptibility marker is expected to be low, and if the marker only confers a modest effect, then it is unlikely that a genotype–phenotype relationship will be seen. Such markers are probably simply inappropriate for testing in phase II studies. Appropriate power calculations considering the allele frequency in the intended population are, therefore, needed when designing such studies. Consideration of the expected ethnic background of the intended target population for a study should also be entertained. Phase II trial networks or cooperative groups can address this issue through increased geographic (and, therefore, hopefully ethnic) diversity. Nevertheless, it may still be impossible to appropriately power a phase II study to examine relatively uncommon pharmacogenomic variants. This challenge may be, ironically, even more true in randomized phase II studies in which only half the patients receive the drug for which the variant is potentially informative. In such cases, a common-arm approach (28) can be used (Fig. 2), although the logistical hurdles of gaining access to samples and data from the available pool of (potentially even international) trials is a practical and real limitation.

If a pharmacogenomic marker does show a signal of activity from a phase II trial, is a randomized prospective study always needed to confirm each potential candidate? Several authors have argued that this is neither universally necessary nor practical (29, 30), because completing a randomized prospective study for each of the candidates that has been or will be discovered in coming years would be cost prohibitive and infeasible, and in some cases, perceived as unethical (30). Cohort or case–control studies in large populations could provide confirmatory answers (29). As in the above EGFR example, entirely retrospective subset analyses of pooled data can be quite convincing.

Alternatively, innovative phase II designs within a given single study may accelerate confirmation of a potential marker. One such approach is the tandem, 2-step phase II design (Fig. 3; refs. 31, 32), also known as an enrichment design (33), which simultaneously allows one to test both a drug and a predictive pharmacogenomic biomarker in the same trial. This design is becoming increasingly used, and we believe it has the potential to inform and accelerate the incorporation of pharmacogenomic biomarkers in drug development. It, of course, requires that the pharmacogenomic marker can be rapidly tested, such that the patient can be treated in a reasonable time frame.

Finally, it must be acknowledged that most of the nominally positive, published associations from prior phase II studies have, to date, not yet changed clinical practice or even resulted in drug label changes. This situation is likely a reflection of the high standard that has been set for validating and then incorporating a pharmacogenomic finding. Nevertheless, we argue here that ultimately increasing the number of practice-changing pharmacogenomic findings will require more routine consideration of pharmacogenomics in earlier-phase clinical trials. We have tried to show that phase II is the most logical setting for such necessary discovery work, and it is most likely to provide the initial results worthy of confirmation testing.

Early-phase clinical trials in oncology are increasingly examining pharmacogenomics. Just as the complete sequencing of the first human genome heralded the advent of pharmacogenomic study, the near-future ability to routinely and cost-effectively sequence every patient's germline and tumor genome will only further the key role of pharmacogenomics in oncology therapeutics. We have shown that phase II clinical trials, in particular, have a favorable track record for showing positive pharmacogenomic signals, worthy of additional follow-up and validation. This positive direction encourages us to recommend increasing incorporation of pharmacogenomic biomarkers in phase II trial designs as secondary and exploratory endpoints, to allow the earliest identification of signals of predictive information. Because phase I is likely too early and phase III may be too late, phase II should, indeed, remain the sweet spot for pharmacogenomics in cancer therapeutics.

No potential conflicts of interest were disclosed.