Randomized trials and endpoints in advanced HCC: Role of PFS as a surrogate of survival

Negative phase III clinical trials

Until 2016, sorafenib was the only systemic agent able to significantly increase survival in patients with advanced HCC.⁷ This was despite numerous attempts to improve, or parallel (i.e., non-inferiority trials), its efficacy and develop new second-line therapies in the context of phase III trials (Table 1). Until then none of the 8 randomized clinical trials (RCTs) testing systemic treatments (vs. sorafenib in frontline^9–12,17 or placebo in second-line^19–21) were able to achieve positive results. Nowadays, 6 of 21 (29%) trials have been able to meet the primary endpoint and potentially change the standard of care. This success rate is lower than in other tumour types, with a reported success rate of 37%,⁴² and resonates with the difficulties of developing effective drugs in HCC. Negative HCC trials enrolled a total of 8,604 patients and consumed a significant amount of resources. Failed drugs include linifanib,¹¹ erlotinib,¹² brivanib,^10,19 sunitinib,⁹ doxorubicin,^17,27 everolimus,²⁰ tivantinib,²³ ADI-PEG 20²⁶ and radioembolization with Y-90^14,15 (Table 1). These drugs have different molecular targets, mechanisms of action, and include various treatment modalities (pharmacological vs. radiation-based). Thus, it is likely that multiple factors contributed to their failure, but we will dissect 3 key factors, reviewed in 43: a) limited antitumoral efficacy (or biological activity), b) significant toxicity, and c) lack of effective enrichment strategies for patient enrolment.

Limited antitumoral efficacy (or biological activity)

The first factor relates to limited antitumoral activity of the drug as per its main molecular targets. This implies that either they have a marginal role in HCC progression, or their selective inhibition is insufficient to induce a significant clinical benefit. For instance, evidence from murine models demonstrated that aberrant activation of MTOR signalling promotes liver cancer,⁴⁴ and that its selective abrogation has antitumoral effects in xenografts.⁴⁵ In human HCC, the MTOR pathway is deregulated in up to 45% of samples,^46,47 and yet, in a phase III trial, everolimus was unable to improve survival compared to placebo in second-line treatment, with an HR for OS of 1.05.²⁰ Data from phase II already suggested a modest median survival of 8.4 months⁴⁷ associated with marginal response rates. Moreover, the companion biomarker study for the phase III trial failed to find any robust predictive biomarker of response to everolimus.⁴⁸ Altogether, these data suggest that MTOR inhibition has no antitumoral activity in advanced HCC. Other drugs potentially falling under this category include the EGFR inhibitor erlotinib,¹² and the FGFR2/VEGFR inhibitor brivanib.^10,19 It has also been shown that Y-90 resin microsphere treatment is not superior to sorafenib. The published phase III trials^14,15 were negative, with an HR for OS of 1.1. It is important to note that failure to demonstrate superiority does not mean similar efficacy, which requires an ad hoc trial design for non-inferiority or equivalence, a concept that will be further discussed in this review.⁴⁹

Drug toxicity

The second reason is significant drug toxicity, which is relevant in cirrhotic patients since liver dysfunction decreases the threshold for severe adverse effects. The best example is sunitinib,⁵⁰ which was unable to improve survival when compared to sorafenib in frontline.⁹ Despite having a molecular target profile similar to sorafenib, the trial was prematurely terminated due to futility and safety concerns relating to sunitinib. The median OS with sunitinib was 7.9 months, compared to the 10.2 months with sorafenib (HR of 1.3). Treatment-related deaths occurred in 3.2% and 0.4% of patients receiving sunitinib and sorafenib, respectively. When this trial was conducted, sunitinib was already FDA approved for advanced kidney cancer and gastrointestinal stromal tumours, where toxicity was not a major clinical issue. However, in patients with underlying liver disease the toxicity of sunitinib was severe enough to obscure any beneficial antitumoral efficacy. Sunitinib has a higher inhibitory potency than sorafenib, particularly regarding its anti-angiogenic activity via VEGFR and PDGFR inhibition.⁵¹ Angiogenesis is critical during liver fibrogenesis,⁵² so the strong and sustained anti-angiogenic effect achieved with sunitinib seems detrimental, favouring liver failure. Previous phase II trials testing sunitinib in HCC offer additional insights into the hepatic toxicity of this drug,^53–55 including up to 4/37 (11%) treatment-related deaths.⁵⁵ An adequate identification of toxicity signals at this stage could help mitigate this problem. Another example is the VEGFR/PDGFR inhibitor linifanib, tested against sorafenib in a frontline setting.¹¹ This trial was terminated early based on futility (median OS for linifanib and sorafenib were 9.1 and 9.8 months), but grade 3–4 adverse events were significantly more frequent in linifanib than in sorafenib, including hypertension (21% vs. 11%) and hepatic encephalopathy (7% vs. 3%). Besides the negative effect of toxicity in clinical outcomes, there is a subtler effect of non-lethal toxicity as it associates with dose reductions, which could also decrease antitumor potency.

Lack of trial enrichment strategies

A third reason for clinical trial failure is the lack of effective enrichment strategies for patient enrolment based on predicted biomarkers of response. Trial enrichment in oncology is closely linked to the concept of oncogene addiction. This term describes those molecular alterations, generally DNA mutations or chromosomal aberrations, required for cancer cell proliferation and survival.⁵⁶ There are numerous examples in oncology of survival benefits after a clinical trial testing a drug only in those patients with mutations in its target⁵⁷ (e.g., ALK rearrangements in lung cancer and response to crizotinib⁵⁸). Only 2/21 (10%) phase III trials in advanced HCC incorporated patient enrichment, likely due to: a) limited access to tumour tissue in patients already diagnosed by non-invasive criteria; and b) few druggable targets among the most common genetic alterations in HCC.^59,60 In fact, the most common mutations in HCC (TERT promoter, CTNNB1, TP53, AXIN1, ARID1A and ARID1B) are untargetable.^2,61 One trial evaluated tivantinib vs. placebo in second-line treatment for patients with high expression of MET assessed by immunohistochemistry.²³ This trial was based on a post hoc analysis of 37 patients from a previous phase II trial⁶² and failed to meet its primary endpoint with an HR for OS of 0.97. Arguably, the signal in the phase II trial was weak, but most importantly, recent data questions the specificity of tivantinib as a MET inhibitor.⁶³ It was also thought that MET was a prognostic factor, but the median survival of 9.1 months for the placebo arm in MET-high patients in a second-line setting challenges this concept.²³ The second trial tested the VEGFR2 monoclonal antibody ramucirumab vs. placebo in second-line (i.e. REACH-2) in patients with AFP higher than 400 ng/ml, and showed a significant improvement in OS vs. placebo (HR of 0.71²⁵). A difference with the tivantinib case is that the rationale for REACH-2 came from a post hoc analysis of the negative phase III trial in all-comers (i.e., REACH²¹) which enrolled 565 patients. This showed a robust improvement in OS with ramucirumab in patients with high AFP (p = 0.02). AFP is a well-known poor prognostic marker,^36,47 highly expressed in tumours with a supposed progenitor cell origin,⁶⁴ but it does not provide a neat link between any specific driver oncogenic event (i.e., structural DNA alterations or signalling pathways) and ramucirumab’s main molecular target. Experimental evidence identifies VEGFR2 as a marker of hepatic progenitors,⁶⁵ which could hypothetically explain the efficacy of ramucirumab in patients with high AFP.

Analysis of surrogate end-points in HCC

In order to explore the concept of surrogate endpoints recapitulating OS in advanced HCC we have identified 21 RCTs assessing systemic therapies with or without loco-regional therapies in advanced HCC (12 in first-line and 9 in second-line) (Table 1) published between 2008 and 2018 through a MEDLINE search via PubMed using the keywords “advanced hepatocellular carcinoma”. Results were limited to “clinical trial, phase III”. Trials recently presented at international meetings (2016–2018) were also included despite the full manuscripts not yet being available. For each trial, data on sample size, radiological response, TTP, PFS and OS were collected. TTP, PFS and OS were determined in terms of HR using published data (values less than 1 denote a favourable result in the experimental group). In addition, ORR was established with the odds ratio calculated from the published radiological response (values greater than 1 denote a favourable result in the experimental group). For the purpose of the trial-level analysis, we first assessed the overall correlation between PFS and OS (R = 0.84; R² = 0.71) (Fig. 3A), and then the correlation of TTP and OS (R = 0.83; R² = 0.69) (Fig. 3B). Afterwards, we established a conservative threshold of positive predictive value for PFS since this is the most documented surrogate time-to-event endpoint in oncology, and the one showing a higher correlation with OS.⁸⁴ Finally, we explored the correlation between ORR assessed by mRECIST and survival in early, intermediate and advanced HCC.

Progression-free survival

PFS is a composite endpoint of 2 variables: death and evidence of radiological progression, usually defined by standard criteria as RECIST⁹¹ or mRECIST.⁹² International guidelines initially discouraged this endpoint in HCC due to the competing risk of dying due to progressed liver dysfunction despite a relevant antitumoral benefit.³³ However, this limitation has been mitigated since most trials in HCC have adopted restrictive inclusion criteria in terms of liver function (i.e., Child-Pugh A without decompensation). In this scenario, the likelihood of death as a result of liver decompensation (i.e., gastrointestinal bleeding, encephalopathy or ascites and spontaneous peritonitis) is 5% at 1 year.⁹³ When we evaluate the association between PFS and OS in HCC phase III trials, we observe a Pearson correlation (R) of 0.84 (Fig. 3A). This figure falls in the upper boundary of a moderate correlation (R between 0.7 and 0.85). When specifically analysing the positive predictive value of theoretical thresholds of PFS correlating with OS, only 3/7 PFS reported an HR ≤0.6 that was significantly associated with a positive survival benefit (in all cases with an HR for OS <0.8). Conversely, those 4 studies reporting a positive PFS with an HR between 0.6–0.8 were associated with no significant survival benefits (HR for OS between 0.87 and 1.05) (Fig. 3A). In our study, according to the linear regression equation obtained [log HR_OS = 0.072 + 0.487 × log HR_PFS], a threshold PFS HR of 0.6 (representing a 40% risk reduction) will decrease the risk of OS by ~17% (OS HR = 0.83) (see Fig. 3A). In summary, a moderate correlation has been established between PFS and OS in 21 RCTs in advanced HCC. A value of HR ≤0.6 is proposed as surrogate threshold effect,⁹⁴ and is likely to predict a clinically meaningful improvement in OS. Is worth mentioning that this rule is supported mainly by positive trials in the second-line setting comparing active drugs vs. placebo. We assume that such an association is retained in a frontline setting, where 2 active drugs are being compared, but any recommendation in that setting should be tempered due to the lack of confirmatory data. In conclusion, PFS has a moderate correlation at trial level with OS (R = 0.84). A conservative minimum surrogate threshold effect of HR ≤0.6 is highly predictive of a significant improvement in OS, whereas HR ranging from 0.6–0.7 is an uncertain surrogate.

Time to progression

This endpoint quantifies the time between trial allocation and radiological progression, usually defined by standard criteria as RECIST⁹¹ or mRECIST.⁹² Deaths are censored as non-radiological progressions at the time of death or at an earlier visit, with a cause-specific hazard, representing informative censoring. Symmetric repeated radiological measurements every 6–8 weeks are required to avoid missing moderate differences between treatment groups.³³ This recommendation was not followed in the SIRveNIB,¹⁵ SARAH¹⁴ and ADI-PEG 20²⁶ trials.

To delineate the adequacy of TTP as a surrogate of OS in HCC we also conducted a trial-level meta-analysis to evaluate the correlation between TTP and OS in 21 RCTs (Fig. 3B). The Pearson correlation (R) was 0.83, which indicates a moderate association according to the IQWIG guidelines.⁸⁴ In 10 phase III trials there was a significant difference in TTP in favour of the investigational arm. However, these positive results in TTP were not reflected by superiority in OS in 5 (50%) trials. Brivanib¹⁹ and ramucirumab²¹ in second-line showed efficacy according to TTP (HR = 0.56 and 0.59, respectively), while not significantly improving the hard endpoint of OS (HR = 0.89 and 0.87, respectively). A trial comparing first-line lenvatinib vs. sorafenib showed a significant positive difference in TTP with levantinib (HR 0.63), which was not reflected by superiority in OS (HR = 0.92).¹³ This trial was positive since it was designed for non-inferiority (upper 95% CI lower than 1.08). Finally, linifanib¹¹ and hepatic infusion arterial chemotherapy (HAIC)¹⁸ in first-line failed to show any benefit in terms of OS (HR = 1.05 and 1.01, respectively) even though there was a clear benefit when measuring TTP (HR = 0.76 and 0.65, respectively). These results do not support the initial recommendation after the SHARP trial of using TTP as the optimal surrogate endpoint³³ in phase II trials, and reinforce the need for accurate evaluation of surrogacy in clinical trials. Based on the linear regression model obtained [log HR_OS = 0.083 + 0.491 × log HR_TTP], we can extrapolate that a therapy producing a 40% risk reduction in TTP will yield an estimated ~16% risk reduction in OS (HR = 0.84) (see Fig. 3B). Moreover, in order to directly compare the performance of PFS and TTP, we analysed the correlation between both surrogate endpoints, obtaining a Pearson correlation (R) of 0.99 (Fig. 4). Thus, in the modern era of HCC trial design, with minimal cirrhosis-related deaths (due to the inclusion of Child-Pugh A), there is a strong correlation between both endpoints. In fact, when we inferred the non-reported PFS HR of SHARP and the Asia Pacific trials according to the linear equation obtained by comparing both surrogate endpoints [log HR_PFS = 0.014 + 0.927 × log HR_TTP], the HRs values are close to 0.60, just at the previously proposed minimum threshold.

There are 2 other considerations regarding this endpoint. First, not all types of tumour progression necessarily have the same clinical meaning. Recent data also suggest that TTP may capture heterogeneous features, with essentially 2 types of progression at advanced stages.^95,96 In particular, survival after progression is significantly worse for patients who develop a new extrahepatic lesion and/or vascular invasion (median OS = 7.1 months) compared to those who progress due to the growth of existing intrahepatic/extrahepatic lesions or the development of a new intrahepatic lesion (median OS = 14.9 months). Second, factors including evaluation bias, trial attrition or informative censoring may weaken the association between the TTP and OS.²⁹ Finally, prolonged exposure to a given therapy might lead to a phenotypic change in tumours, thus, offsetting any initial advantage from the treatment captured by the surrogate endpoint.⁹⁷

Objective response rate

Tumour response in oncology trials is typically measured using RECIST (Response Evaluation Criteria in Solid Tumors).⁹¹ These criteria standardize methods for converting radiological observations into a quantitative and statistically tractable framework to define tumour response (i.e., a 30% decrease in the diameter of target lesion). ORR is the percentage of patients who achieve an objective tumour response. ORR by sensitive criteria in single-arm phase II trials could be a useful tool to prioritize treatments for testing in phase III trials. Disease control rate is the combination of ORR and stable disease, but it has 2 disadvantages that limit its adoption for regulatory approval: a) the definition of duration of stable disease varies between studies; and b) stable disease can reflect inherent characteristics of the tumour rather than treatment efficacy.

The RECIST criteria were originally developed to evaluate cytotoxic agents. The generalization of targeted therapies has challenged this simplistic approach that relies on tumour shrinkage to indicate clinical efficacy. Sorafenib was associated with an ORR of only 2–3%, despite providing clear survival benefits.^7,8 Given the poor correlation between tumour response assessed with conventional tools and OS, a group of experts convened by the American Association for the Study of Liver Diseases (AASLD) proposed specific amendments to standard RECIST.³³ Further description of response and progression resulted in the criteria named modified RECIST (mRECIST), which ultimately incorporates the concept of viable tumour defined as the portions of tumour showing arterial enhancement.⁹² The mRECIST criteria in HCC have improved the sensitivity to quantify tumour response with targeted therapies: ORR of 9–17% with sorafenib,^10,13,18 10–12% with brivanib,^10,19 11% with regorafenib²² and 24% with lenvatinib.¹³ Retrospective studies have consistently demonstrated that patients who achieved an objective response on sorafenib had a longer survival than non-responders.^98–100 Recently, data from double-blind randomized trials assessing brivanib, nintedanib and lenvatinib further validated this association.^101,102 Thus, the association between tumour response and improved OS in patients with HCC at advanced stages complement what was already known in patients at early and intermediate stages treated with loco-regional therapies^103–109 (Table 2).

When we evaluate the trial-level correlation between ORR and OS (Fig. S1), the R weighted Pearson coefficient obtained is 0.54. This is significantly lower than the correlation obtained with PFS/TTP and OS as depicted in Fig. 3. There are 2 reasons for this: one is inherent to the use of odds ratios instead of HRs to compare differences in ORR. The accuracy of an odds ratio decreases for low values of ORRs. The second reason is that only a small proportion of patients within these trials achieved ORR (~10–20%), which is, in fact, the event that correlates with better survival.¹⁰¹ A direct comparison between RECIST and mRECIST for OS surrogacy through an independent meta-analysis of trials using either criteria would be ideal to define the role of ORR to predict OS in advanced HCC. This will also help determine the best tool for evaluating tumour response to systemic therapies. However, since only 5 RCTs reported response data using mRECIST, we did not sub-analyse this endpoint according to the tool used to evaluate response.

Some other questions remain unanswered. As observed for other solid tumours treated with efficacious targeted therapies,^110–112 the reported rates of responders are still suboptimal to estimate the maximum number of patients who would benefit from the treatment. In addition, the duration of response might be more clinically relevant than the extent of tumour reduction. Finally, the strategy to evaluate response might require a thoughtful revision when assessing immunotherapies. As shown in melanoma patients treated with checkpoint inhibitors, standard RECIST may not provide a reliable assessment of antitumor efficacy.¹¹³ In fact, response to immunotherapy may take longer compared to other agents and can even falsely mirror criteria for progression (i.e., pseudo-progression).¹¹⁴ Immune-related response criteria have been developed,^115,116 including the concept of “confirmation of progression” by a second scan obtained at least 4 weeks after progressive disease has been registered.

Despite all the challenges that evaluation of tumour response raises in oncology, and particularly in HCC, the importance of ORR as a surrogate endpoint is recognized by regulatory agencies and frequently used for accelerated drug approval. This was the case for nivolumab, approved in second-line based on an ORR of 18% by mRECIST and 14% by RECIST.^38,39 Remarkably, objective response to nivolumab has been associated with prolonged OS.⁶⁶ Overall, the fact that a high ORR in phase II trials was considered a robust criterion for drug approval,¹¹⁷ and further success in phase III trials,¹¹⁸ indicates that ORR should be considered as a primary endpoint for single-arm phase II studies. Related to this, early clinical trials are showing promising results with combinations of checkpoint inhibitors and targeted therapies, as measured by ORR. In front-line advanced HCC, combination treatments with lenvatinib plus pembrolizumab¹¹⁹ achieved an ORR of 46% by mRECIST and atezolizumab plus bevacizumab (n = 73)^120,121 34% by mRECIST. As a result, the later combination was granted breakthrough therapy designation by the FDA.¹²² Of note, most of the drugs approved under the accelerated programme reported ORRs exceeding 30%.¹²³