Efficacy assessment in phase I clinical trials: endpoints and challenges

Abstract

The scope of phase I clinical trials in oncology goes beyond the conventional safety evaluation–only objectives of these trials in other specialties. Rather, most first-in-human cancer trials have therapeutic intent, and efficacy signals observed in phase I trials can drive a go/no-go decision of advancing a new molecule to phase II testing. The complexity of efficacy assessment in the context of a small, heterogenous patient population and a complex study design requires a more liberal perspective compared to later trial phases when looking into efficacy endpoints. Classically, in later phase clinical trials, these endpoints would include the objective response rate, progression-free survival, and overall survival. However, new, evolving endpoints may be worth investigating when looking into the antitumor activity signals in phase I trials. Integration of all these endpoints into trial designs can improve the assessment of therapeutic efficacy during early drug development and guide decisions related to further advancement of novel molecules into later phases. In this review, we discuss the advantages and pitfalls of different classic efficacy endpoints when evaluated as part of phase I trials in oncology and describe how challenges in assessing the antitumor activity of new drugs can be overcome through incorporation of novel endpoints that have thus far proven successful in clinical trials.

INTRODUCTION

Early-phase clinical trials are designed to evaluate novel drugs in humans with the primary purpose of establishing the drugs’ safety and preliminary efficacy prior to moving to registrational trials. This plays a very important role in the oncology field given the challenging disease nature and lack of curative therapeutic options, especially in the advanced setting.¹ Outside oncology, early-phase clinical trials sometimes include healthy participants with the primary aim of evaluating drug safety, which remains the main objective in phase I oncology trials.² However, phase I clinical trials in cancer patients mostly include those with advanced disease, and the agents tested are frequently considered viable treatment options especially in patients with no alternative standard of care therapy.³ In addition, efficacy outcome data from phase I trials are commonly used to make strategic decisions on whether drugs are worthy of exploring in a more advanced phase II or III clinical trial (Supplementary Figure 1). ^{3, 4} This is crucial given the increase in the cost- and resource- burden of running late phase clinical trials over the past decades. ^7–9 From a registrational perspective, there have been instances where early drug approvals were based on results from phase 1 or phase 1/2 trials. For example, the first FDA approval of pembrolizumab in patients with unresectable or metastatic melanoma was based on data from the KEYNOTE-001 phase 1b trial ⁵.The Pharmacological Audit Trail constitutes a framework of critical items explored during the discovery and early development of an anticancer drug that are predictive of success.¹⁰ According to the Pharmacological Audit Trail, in phase I trials, in addition to safety, pharmacokinetic properties, and pharmacodynamic proof of mechanism in tumor tissue or surrogates (e.g., skin, hair follicles, peripheral blood mononuclear cells), initial efficacy signals constitute key elements for moving drugs forward to later trial phases. ¹¹ More and more, though, the go/go-no decision for moving a drug from phase I to phase II testing is contingent on the presence of efficacy signals in the phase I setting. Hence, diverse methods of identifying antitumor signals in dedicated expansion phases in the phase I setting are increasingly being used to test preliminary efficacy across different diseases in nested basket trials after dose optimization.

Despite interest in efficacy assessment as a pivotal question to be answered in phase I trials, even if proposed as a secondary endpoint, challenges exist in how to evaluate patients from an efficacy perspective in such an early and complex phase of clinical development. A confounding factor is the highly heterogeneous patient population enrolled in phase 1 trials including a broad range of tumor types, diverse genomic profiles and prior lines of therapy. Classic efficacy endpoints used in phase II and III trials (e.g., response rate, progression-free survival [PFS]) are commonly used to study efficacy in expansion cohorts in the phase I setting. Still, given the critical milestone of moving a drug from phase I to phase II testing, a more elaborate toolbox than the one currently in use is necessary given the resources needed in phase II trials. In this review, we elaborate on the utility of classic and evolving efficacy endpoints that can be used for outcome evaluation in early drug development. We highlight benefits and challenges related to these endpoints and identify possible areas of further research that can be helpful in advancing and validating these endpoints.

SIZE-BASED ENDPOINTS: SIZE MATTERS (AND DECEIVES THE EYES)

Objective response rate

The most classic endpoint for efficacy assessment in clinical trials is the objective response rate (ORR) (Supplementary Table 1).¹² Different response criteria have been developed, but the response criteria most used currently—the Response Evaluation Criteria in Solid Tumors (RECIST)—were developed in 2000 and updated in 2009. RECIST represent an objective approximation of efficacy to determine whether changes in tumor diameter indicate antitumor efficacy.¹³ Such standardization of response assessment tools allows for combining and comparing data across tumor types and tumor sites and even different clinical trials.

Despite the wide use and validation of RECIST, the ability of RECIST to establish the long-term benefit of a drug has been a subject of debate. Moreover, reproducibility issues of RECIST and intrareader/interreader variability in measurement of the size of target lesions are not uncommon when using RECIST,^{14, 15} and studies have reported significant discrepancies in RECIST response rates between investigator assessments and central reviewers assessments.^{16, 17} For example, a difference in the sums of the tumor diameters of a few millimeters in patients with low disease burdens (which could be observer-dependent) can dramatically change the ORR given the artificial thresholds for classification between categories of response. Unclear criteria in the selection of target lesions may be another questionable practice when using RECIST, as variations in decisions on which lesions should be selected as targets can lead to variations in response assessment. ¹⁸. Efficacy assessment using RECIST may not be applicable in a large group of patients who have nonmeasurable disease despite significant disease spread (e.g., prostate cancer or breast cancer with bone only metastasis.

Other response criteria have been proposed to overcome the limitations of RECIST. For example, for patients with primary or secondary tumors affecting the central nervous system, the Response Assessment in Neuro-Oncology (RANO) criteria, including its different variations (e.g., RANO-LGG for low-grade gliomas, RANO-BM for brain metastases), have been used to evaluate patients for response. However, subjective parameters are included in this assessment (e.g., use of steroids, neurological symptoms such as seizure frequency) that can account for differences in practice rather than the efficacy of the investigational agent. Additionally, vascular leakage, which can vary with different drugs (e.g., bevacizumab, steroids), can greatly influence the assessment of gadolinium uptake.

In the era of immunotherapy, additional issues with tumor size assessment have emerged. For example, pseudoprogression and hyperprogression evolved as new concepts reflecting aspects of cancer immunobiology that remain unsolved.^19–21 To mitigate these challenges, RECIST criteria for immunotherapy were proposed and efforts are ongoing to normalize the therapeutic decisions across study sites and investigators ²². Nevertheless, this methodological approach is not without limitations. Most patients initially categorized as having pseudoprogression experience true progressive disease (even with a mixed response or onset of new lesions), leading to an ethical conflict related to continuous exposure to an ineffective treatment and the associated side effects. Additionally, controversies exist related to the frequency of pseudoprogression in patients with tumor types that are less immunogenic (e.g., colorectal cancer, pancreatic cancer, hepatocellular carcinoma, gastroesophageal cancers) ²³.

Disease control and clinical benefit rates

In an attempt to overcome limitations imposed by ORR, researchers sometimes use the terms disease control rate (DCR) and clinical benefit rate (CBR) to include cases with stable disease, which sometimes can be considered as a desirable outcome.²⁴ Several experts would argue against the use of stable disease as a favorable outcome given its broad definition. However, durable stable disease is frequently seen with newer agents, indicating a cytostatic rather than a tumoricidal mechanism of action. For example, the FDA approved tebentafusp showed a response rate of 5% in the phase 2 study evaluating its use in patients with previously treated metastatic uveal melanoma ²⁵. Durable stable disease may impact patient survival. For instance, in a randomized phase 3 trial in patients with IDH mutated cholangiocarcinoma, ivodesinib, an IDH1 inhibitor, resulted in improved PFS and OS despite a negligible overall response rate (2%)²⁶. However, a lack of consensus on how these DCR/CBR should be defined has led to heterogeneity in intertrial assessments of response. Some investigators would prefer to include only patients with stable disease for 4–6 months in their estimation of disease-control or clinical benefit rates, whereas others will include all patients who had stable disease as the best response after the first assessment of efficacy.

Clinically meaningful duration of response (DOR; discussed more in time-based endpoints section) is difficult to define, and the definition is probably specific to the tumor type and disease setting. Additionally, some investigators would argue that in some early-phase trials, prolonged stable disease was most often related to disease biology (indolent disease being arguably more frequent among phase I trial participants) rather than clinical benefit of investigational agents. Some evolving concepts, such as comparing PFS for immediate prior therapy with PFS for an investigational drug as described below, can be used to address such challenges.

Stable disease: what does it really mean?

The category of stable disease is a very gray area with uncertain value in assessing a drug’s effect. Stable disease might not be stable, because the term may not capture changes that happen between the two artificial thresholds for response categories of −30% to +20% (stable disease would include patients who had between −30% decrease and 20% increase in target lesion sum of diameters). Moreover, changes in computed tomography (CT) density (necrosis) in patients treated with some agents may occur as a reflection of dynamic changes in tumor tissue after a given therapy. This effect in image density, typical of gastrointestinal stromal tumors (GIST) responding to treatment with KIT inhibitors, is depicted better using the Choi criteria, which have been correlated with survival in patients with GIST.²⁷ Moreover, as discussed below, metabolic response in geometrically stable lesions may be correlated with clinical outcomes.

Methods of visualization of changes in tumor size

Correlative plots, including waterfall plots (of greatest tumor shrinkage), swimmer plots (of duration of treatment), and spider plots (of tumor burden evolution over time), have been suggested as tools that can be more descriptive alternatives to classic assessment of efficacy in tables (Figure 1). Still, these plots are not free of the effects of interreader and intrareader variability and pose the risk of overestimating treatment efficacy, which can affect a go/no-go decision.²⁸ For many scenarios with no historical comparisons of response outcomes with currently approved agents (e.g., standard-of-care chemotherapies), these plots can be misguiding and can overestimate treatment response (confirmation bias). A more visually appealing and comprehensive approach to response evaluation is a 3D waterfall plot with lesion size changes on one axis and durability of response on the other. This incorporates time-based assessment of efficacy as discussed below and provides a multidimensional description of treatment effect in one graph.²⁹

Figure 1: Visualization methods for efficacy assessment in a mock dataset.

(A) Summary of RECIST best response in the mock dataset. (B) Waterfall plot showing best response per RECIST which provided more clarity on the spectrum of responses. (C) Waterfall plot showing liquid biopsy response suggesting that even in patients with SD as the best response, there was a deep ctDNA response. (D) 3D Waterfall plot showing a combination of size- and time-based assessments suggesting that even in patients with SD as the best response, multiple patients had prolonged PFS. (E) Spiderplot showing change in response over time. (F) Swimmer plot showing duration on treatment and response.

* Same color code has been used across different graphs with each color representing a specific individual.

TIME-BASED ENDPOINTS: TIME IS OF THE ESSENCE

Overall survival

Time-to-event endpoints like PFS and overall survival (OS) have classically been used to report therapeutic efficacy in cancer clinical trials.^{30, 31} Notably, the ultimate goal for cancer clinical trials to date remains improvement of OS ^30–33. The purpose of any therapy given to patients, especially in the metastatic cancer setting, is to prolong and improve the quality of life. Over the years, whether other endpoints used for efficacy assessment may be reflective of OS benefit has been a subject of debate. For example, the potential for disconnection between PFS and OS has been concerning to some regulatory agencies. This has been demonstrated with different therapeutic approaches, including cancer vaccines, immunotherapy, and cellular therapies, for which an increase in OS duration was not preceded by prolonged PFS ^{34, 35}. More often, improvement in PFS that was not translated into prolongation of OS was observed in some clinical scenarios (e.g., with treatment of BRCA-mutated metastatic pancreatic cancer with olaparib).³⁶ In any case, the use of OS for efficacy assessment in the phase I setting may be less relevant than its use in other phases, especially in the context of multiple confounding factors (including the effect of prior and subsequent lines of therapy on survival, indolent disease, and access to supportive care).

PFS

Recently, PFS became commonly used as a endpoint for efficacy given its ability to somehow reflect a combination of response and durability, an issue discussed above. PFS is commonly defined as the time from starting treatment (cycle 1 day 1) to the time of disease progression or last follow-up ^{30, 31}. However, several critics have argued that PFS may not be a good surrogate for OS in cancer clinical trials and should not be used as the sole endpoint ³⁷. In addition, there are specific challenges in relation to PFS use in the phase I setting. First, some trials may not require progression to prior therapy to be eligible. Therefore, carryover effect from prior treatment may impact PFS in subsequent phase I. Second, some patients may be taken off treatment because of toxicity, which is not an infrequent occurrence in the phase I setting. Such patients should be censored when taken off treatment or when starting a subsequent treatment, as continuing to calculate PFS while undergoing another therapy may alter the results of efficacy (censoring bias). Third, the role of PFS in efficacy assessment may be controversial in patients who continue treatment beyond progression of their disease based on the investigator’s judgment of clinical benefit, which is a subjective assessment. This can be even more complicated in patients with oligoprogressive disease who remain on treatment despite having mixed responses and in patients receiving palliative localized therapy. In such cases, the definition of which event determines progression becomes debatable. In most cases, investigators would define progression as the time after the first signs of progression and before clinical deterioration, especially in patients who are expected to have stable or responsive disease after experiencing oligoprogression. Another issue in the phase I context is the heterogeneity of tumor behavior, which can lead to differences in natural history (e.g. aggressive tumors would have shorter PFS and indolent tumors would have longer PFS, and both can be included in the same analysis set). Moreover, selection bias in patients participating in phase I trials after failure of standard-of-care options can be another point of concern.

Another issue related to the increasing use of PFS as a main endpoint in cancer clinical trials, which is especially popular in modern phase I trials of immunotherapy, is that it requires substantially more resources than other endpoints to be appropriately assessed (number of patients, longer evaluation period, and either a control arm or a well-annotated historical control), which are often scarce in these scenarios. The use of historical controls recently supported FDA approval of selumetinib in a small study in children with neurofibromatosis with inoperable plexiform neurofibromas ³⁸. Randomization may be preferred to a historical control for comparison, but because randomization to controls is rarely available in phase I trials, PFS data in isolation should be interpreted with caution. Taken together, these findings demonstrate that interpretation of PFS and other time-based endpoints may be clouded by systematic biases (most frequently patient selection bias) and take much longer than interpretation of ORR.

Other time-to-event endpoints

One endpoint that has been increasingly used to overcome some of the challenges related to the calculation of PFS described above is time to treatment discontinuation, or time to treatment failure. This is defined as the duration from starting treatment (cycle 1 day 1) to treatment discontinuation regardless of the reason (e.g., toxicity, clinical deterioration, radiological progression) ³¹. Although this endpoint may be easy to calculate from a statistical perspective, whether it is reflective of therapeutic efficacy remains controversial.

Another increasingly used time-to-event endpoint is the DOR. Classically, this is the interval from first documented response to progression or death, whichever occurs first. The benefit of assessing DOR is that it accounts for the durability of response rather than rapid but short-lived responses. In one study, DOR in phase II trials was shown to be more sensitive than PFS and ORR in predicting benefits in ensuing phase III trials ^{39, 40}. However, DOR provides a limited picture of the benefit from an investigational agent and may not necessarily correlate with improved survival. The POLO trial randomized patients with advanced pancreatic cancer with germline BRCA mutations to maintenance Olaparib vs placebo. The study did not allow cross over (although 15% of patients received a PARP inhibitor after progression). Olaparib improved PFS, response rate and DOR but did not translate into improved overall survival in these patients.

NOVEL ENDPOINTS: OVERCOMING LIMITATIONS TO CLASSIC ENDPOINTS

Metabolic response

Another type of response that has drawn attention over the past few years is metabolic response. This effect relies on changes in functional imaging, which uses radionucleotide probes, as indicators of biological efficacy. This has been commonly implemented in studies of hematological malignancies such as lymphomas, for which fludeoxyglucose F18 (FDG) positron emission tomography (PET)/CT response criteria have been developed, validated, and applied in clinical practice to guide clinical decisions ⁴¹. For solid tumors, however, the concept of metabolic response is less established, although examples of its utility exist ^42–44. For example, PET response criteria in solid tumors (PERCIST) and sodium fluoride PET response criteria for solid tumors (NAFCIST) have been proposed, validated for multiple tumor types, and shown to correlate with clinical outcomes, even in patients with nongeometrically changing lesions ^45–53. This may add value to the efficacy assessment for new agents with a cytotoxic effect, providing context to patients with stable disease as per RECIST. Still, several challenges in modern drug development persist, such as the inability to discern inflammation, infection, and progressive disease when using FDG uptake, which is important in areas such as immuno-oncology and intratumoral therapies ⁵⁴. More recently, radiolabeled pharmaceuticals (e.g., zirconium-89–labeled atezolizumab) have been suggested as tools for predicting response to therapy with immune checkpoint blockade ⁵⁵. Data from a recent study reported that using a fibroblast activation protein inhibitor (FAPI)-based covalent targeted radioligand (CTR) tracer improved the diagnostic accuracy of medullary thyroid carcinoma compared to FDG PET-CT allowing for more accurate management of patients⁵⁶; and bringing excitement about further potential uses which can also include response evaluation. ⁵⁷

Molecular response: the oracle of liquid biopsy

One promising endpoint for assessment of response in early-phase clinical trials of new agents is the liquid biopsy. Liquid biopsy refers to a broad spectrum of cancer-related components (e.g., tumor cells, DNA, microvesicles) that can be detected in the circulation of patients with shedding tumors. Because the amount of shedding may correlate with the tumor burden, analysis of these blood elements may be used as a tool for the assessment of molecular response ^58–62. The noninvasive nature of a liquid biopsy may offer an opportunity for real-time tracking of response. Repeat sampling of body fluids and correlations of changes in liquid biopsy analytes with pharmacokinetics, pharmacodynamics, and radiological response may be meaningful independent endpoints in phase I trials. The holistic overview of a tumor’s genetic profile provided by liquid biopsy can also provide an opportunity for longitudinal tracking of the clonal evolution that may happen throughout the treatment course and identification of possible molecular mechanisms of resistance.²⁵

Circulating free DNA has been validated as a predictive and prognostic biomarker in different disease settings, including advanced cancer and early-phase clinical trials. ^{60, 63–65} For example, a study with a pan-cancer cohort of patients enrolled in phase I/II clinical trials demonstrated that dynamic changes in variant allele frequency at day 21 of treatment and at the time of first restaging were associated with clinical or radiological response.⁶⁰

Liquid biopsy can help inform the context of gray area of stable disease discussed above in terms of clinical benefit. An example of this concept was observed in patients with uveal melanoma treated with the bispecific anti-CD3/TCR antibody tebentafusp. Specifically, patients had an ORR of only 5%, but 57% of the patients had stable disease, most with significant circulating tumor DNA level reductions ²⁵. Based on these results, tebentafusp was successfully explored in randomized phase III trials and is currently approved by the U.S. Food and Drug Administration for treatment of advanced uveal melanoma. Additionally, the value of liquid biopsy in patients with suspected pseudoprogression, another gray area, can be very compelling. Multiple reports have suggested that the use of liquid biopsy in these patients can differentiate between those with pseudoprogression and those with true progression ^{60, 66, 67}.

Given the promising results of utilizing circulating tumor DNA in efficacy assessment, Liquid Biopsy RECIST (LB-RECIST) have been proposed. Standardization of efficacy reporting using liquid biopsy may be underway, and efforts to decrease heterogeneity in drug development settings are being explored.⁶⁸ For example, the U.S. Food and Drug Administration (FDA) issued a guidance report for the use of liquid biopsy in early-phase clinical trials, which suggests the need for further data to support the use of circulating tumor DNA level as an endpoint likely to predict long-term treatment outcomes. ⁶⁹ The FDA proposal also highlights the need for standardized protocols for sampling and processing liquid biopsy, accounting for possible differences in tumor shedding, and establishing technical reproducibility and robust analytical performance as possible areas of interest prior to wide-scale implementation of liquid biopsy as a response biomarker. Several initiatives have been launched including the ctMoniTR Project by Friends of Cancer Research which was started in 2018 and continues to present important findings in the context of monitoring treatment response using ctDNA. ⁷⁰

These liquid biopsy methods face significant challenges regarding their reliability when making go/no-go decisions in small trials in early development, however.⁶⁸ For example, liquid biopsy testing may not be technically feasible, especially in a low-resource setting. Differences in liquid biopsy assays’ analytical sensitivity and specificity may also be a point of concern. Differences in the performance of tumor-informed (using tissue NGS to design the assay for liquid biopsy) and tumor-agnostic (using a liquid biopsy assay that is independent on tissue alterations) assays must be clarified in the context of efficacy assessment of investigational agents and determining which one should be regarded as superior to the others⁷¹. In addition to their utility in evaluating response of patients to investigational agents, tumor-agnostic approaches can be used to detect evolving mechanisms of treatment resistance and clonal evolution, which can guide decisions to stop treatment or add a therapeutic agent to an ongoing regimen to target those resistance mechanisms. In comparison, tumor-informed approaches can vary in their complexity, ranging from easier, quicker assays (e.g., polymerase chain reaction technologies) to more complex personalized assays. However, tumor-informed approached would arguably be a more personalized method of assessment of tumors than tumor-agnostic approaches. Variability in results of liquid biopsy related to differences in technology, daily variations in circulating tumor DNA amount, clonal hematopoiesis, and changes associated with inflammation and infection, among others, can also be challenges in the wide-scale application of circulating free DNA in efficacy assessment. Other analytes used in liquid biopsies (e.g., methylated circulating free DNA, circulating tumor cells, exosomes) present similar challenges.

Given the findings described above, pending confirmatory studies demonstrating the role of liquid biopsy in the assessment of molecular response to investigational agents, the use of liquid biopsies in general and circulating tumor DNA dynamics in particular may be considered a supplemental method for demonstration of therapeutic efficacy that can complement other methods of efficacy assessment.

Radiomics: what the eyes cannot see

Radiomics is a rapidly evolving field that employs advanced computational techniques to extract extensive quantitative features from normal radiological images (such as CT, magnetic resonance imaging, and PET scans) ⁷². By extracting and quantifying numerous imaging features such as tumor shape, texture, and intensity using sophisticated algorithms, this technology has the potential to provide detailed insights into tumor characteristics that are not visible using traditional imaging techniques ⁷³. Tumors commonly exhibit significant spatial variability, with different regions responding differently to treatment. Radiomic features may quantify these differential responses, offering a more detailed assessment of tumor behavior than conventional methods, and may be crucial predictors of early treatment response or resistance (Figure 3). In phase I clinical trials, this technology could be particularly critical for detecting subtle incremental effects of therapy in patients with cancer for whom traditional methods of efficacy assessment may fail to reveal beneficial therapeutic outcomes.

Figure 3: Novel radiological methods for evaluation of response.

(Upper Panel) Volumetric progression of Lung and Liver Lesions. Longitudinal volumetric analysis of lung and liver lesions in Uveal Melanoma patients treated with roginolisib (PI3Kd inhibitor) over a 24-week period. (A-B) Percentage change in volume of lung (A) and liver (B) lesions over time, showing the range of responses across all lesions (RECIST 1.1 target and non-target). (C-H) Waterfall plots depicting percentage changes in individual patient lesions across three-time intervals: screening to week 8 (C, F), week 8 to 16 (D, G), and week 16 to 24 (E, H) for lung (C, D, E) and liver (F, G, H) lesions. Each bar in the waterfall plots represents an individual patient’s lesion growth or shrinkage during the specified period. (Lower Panel) Longitudinal analysis of a splitting lung lesion and selected radiomic features. Tracking of a lung lesion over 21 weeks. Top: Visual representation of the lesion’s progression. Bottom: Evolution of selected (normalized) radiomics features (red) vs volume (black). Non-Small Cell Lung Cancer treated with Inhaled chemotherapy. Image courtesy iOnctura SA (NCT04328844). Analysis and images performed by radiomics.bio

A particularly interesting research area is the concept of delta radiomics, which refers to the analysis of changes in radiomic features over time, generally from the start to the end of treatment ⁷⁴. By comparing imaging features obtained at different moments during treatment, delta radiomics can provide dynamic insights into how a tumor is responding to therapy ⁷⁴.

Despite the potential of radiomics, its integration into research faces several challenges. One key problem is a lack of standardization in image acquisition, processing, and analysis. Differences in imaging protocols and equipment can lead to inconsistencies in radiomic features, negatively impacting reproducibility of measurements. Researchers are making efforts to standardize the radiomics workflow, from image acquisition to feature extraction, to ensure robust, reliable radiomic analyses ⁷⁵. Another challenge arises from the high dimensionality of radiomic data. The vast number of features extracted by radiomics can lead to overfitting, especially in phase I trials, in which datasets are small. Advanced statistical methods and rigorous validation protocols are necessary to develop models for evaluation of radiomics that are both accurate and generalizable. Moreover, understanding the biological significance of the extracted data features in radiomics is essential for clinical integration of the concept into clinical management.

Tumor growth inhibition and the growth modulation index

Tumor growth inhibition and the growth modulation index (GMI) are two concepts developed that stem from the understanding that the natural history of cancer is accelerated by tumor growth (Figure 2). Although less commonly used for efficacy assessment than other methodologies described before, both concepts can provide information on intrapatient efficacy of investigational agents by using the patient as their own comparator. This could be very compelling in early drug development given the selection bias and patient heterogeneity (lines of therapy, molecular background, tumor type) in phase I trials. ⁷⁶

Figure 2: Tumor Growth Inhibition.

Graphical example of a differential tumor growth rate (TGR) for two different treatments administered for the same tumor type with the same size in the same setting. Of note is that even though the tumor continues to grow during both treatments, it does so to a markedly greater degree with treatment B than with treatment A.

Assessment of tumor growth inhibition integrates the tumor history (e.g., indolent versus aggressive disease) and biology in assessment of the efficacy of a new agent and pragmatically evaluates the cytostatic effect of the investigational agent in the context of the tumor’s natural history by comparing the slope of the tumor growth curve before and after treatment with the agent. This can be an intriguing concept in phase I trials as it makes the patient the unit of comparison, and it has been correlated with antitumor activity ⁷⁷.

The GMI is like tumor growth inhibition, with the difference being that the GMI takes into consideration the time factor in its calculation. Although several methods can be used to calculate the GMI, the most frequently used one is the ratio of the PFS duration for an investigational drug to the PFS duration for the most immediate prior therapy. A ratio of more than 1.3 (by consensus agreement among researchers) is considered indicative of therapeutic efficacy ^77–79. Multiple studies have demonstrated this ratio to correlate with clinical outcomes in clinical trials (e.g., trials of neurotrophic tropomyosin-receptor kinase inhibitors) ^80–82 and to facilitate assessment of therapeutic approaches in n-of-1 trials ⁸³.

Several problems may arise when applying these two methods to the clinical research context. For example, they end up being very dependent on the efficacy of the immediately prior therapy, leading to unfair or heterogeneous comparisons of efficacy. For example, testing a new agent when the prior line of therapy included a drug or regimen known to prolong survival (e.g., the FOLFOX regimen for colorectal cancer) may lead to underestimation of the efficacy of an agent tested in a phase I trial. Additionally, the difference in frequency and timing of scans between the two regimens being compared (e.g., every 6 weeks in some phase I trials vs every 3 months in standard of care setting) can contribute to detection bias or lead time bias. Attempts to standardize the assessment of GMI have been proposed, with some promising results already observed, such as in the context of treatment of patients with RET-altered cancers with RET inhibitors ⁸⁴. Although this example may not be directly relevant to first-in-human studies given the heterogeneity of the patient populations in them, the GMI concept has been used to identify the appropriate sequence of therapeutic regimens when the evidence of superiority of regimen over the other is unclear. For example, a phase III trial comparing the historically preferred ECX regimen followed by the FOLFIRI regimen at progression versus FOLFIRI followed by ECX in patients with gastric cancer suggested that ECX followed by FOLFIRI is the better therapeutic approach.⁸⁵ A recent phase II study used GMI as the primary endpoint while evaluating the efficacy of trabectedin in patients with retroperitoneal liposarcoma and leiomyosarcoma ⁸⁶. However, validation of different approaches to calculating the GMI and its correlation with long-term outcomes in patients with cancer may be needed. In any case, GMI calculations could complement the assessment of a new agent, especially in the absence of tumor shrinkage.

Other endpoints

In addition to the previously reported methodologies for evaluating efficacy in the phase I context described above, other tools have been reported to be correlated with clinical outcomes in patients with cancer. For example, use of 3D tumor volume in efficacy assessment instead of the 1D (one diameter as used in RECIST) or 2D (two diameters as used in RANO) volume has been proposed (Figure 3). In fact, 3D volumetric analysis has been shown to be superior to some of the classic size-based assessments (e.g., RANO-BM).⁸⁷ Computer-assisted tools with image recognition may allow for evaluation of the total tumor burden (rather than a limited number of target lesions as in RECIST) (Figure 3). This approach has been shown to correlate with OS in patients with intrahepatic cholangiocarcinoma who underwent transarterial chemoembolization. ⁸⁸ More precise measurements such as 3D tumor volume and total tumor burden evaluation could be facilitated by the introduction of artificial intelligence in clinical practice. Other time-based endpoints relevant to efficacy assessment include time to progression, duration of treatment, and time to new metastasis. Another endpoint increasingly of interest is the incorporation of symptoms and patient-reported outcomes into treatment efficacy assessment. This has been applied more often in phase II and III trials than in phase I trials, but the valuable insight into efficacy assessment that this can bring is also crucial to phase I trials. An example is reporting the improvement in symptoms in patients with desmoid tumors treated with nirogacestat in a phase III trial.⁸⁹

EFFICACY OUTCOMES FROM PHASE I TRIALS SHOULD STILL BE EXTRAPOLATED WITH CAUTION

Despite the high relevance of efficacy assessment in phase I trials, it is, however, also important to note several challenges to extrapolating efficacy outcomes from phase I clinical trials. There have been instances of regulatory withdrawals following initial excitement and accelerated approvals based on data generated in phase I/II clinical trials. ⁹⁰ For example, olaratumab was initially approved by FDA and EMA in 2016 to be given in combination with doxorubicin for treatment of patients with metastatic soft tissue sarcoma based on improved OS in a randomized phase II trial. However, the ANNOUCE trial, a randomized phase III trial, failed to meet its primary endpoint of OS and led to withdrawal of market authorization in 2019.

The statistical interpretation of data in the context of small sample size and lack of randomized comparisons might also need adjustments to allow for not underestimating or overestimating efficacy outcomes. More importantly, efficacy outcomes that are used as proof-of-mechanism guiding further drug development can be distinct from those used for drug approvals and registrational purposes which may require a higher bar for establishing efficacy. However, a combination of efficacy outcomes from different endpoints can be helpful for both scenarios although further validation might be needed for ensuring appropriate interpretations in drug approval processes.

HOW TO INCORPORATE AND INTERPRET EFFICACY ASSESSMENT IN PHASE I TRIALS

When evaluating a new agent, not every individual efficacy endpoint will be able to depict the antitumor effect on its own and categorically assist in a go/no-go decision at the end of a phase I trial. However, the endpoints discussed herein can complement each other, as integrating different endpoints can provide improved assessment of efficacy. Selecting which endpoints to use may depend on the class and mechanism of action of investigational agents. In any case, using a combination of several efficacy endpoints may be worth the investment given the cost of and time required for testing low-performing agents in multiple phase II trials. A well-planned phase I trial including some of these endpoints of efficacy assessment can prevent discarding important agents. Examples of phase I trials integrating multiple endpoints in efficacy assessment are numerous. For example, a clinical trial of the investigational PRMT5 inhibitor, AMG193, showed consistent efficacy signals using size-based and time-based measurements; as well as associations with liquid biopsy dynamic changes. ⁹³ Integration of classic ORR and PFS with tumor volume growth rates was also reported in a phase I trial of Ivosidenib in patients with advanced glioma. ⁹⁴ However, the cumulative integration of all discussed endpoints together has not been previously explored. Scorecards can be developed that include different efficacy endpoints, use of which can help define the totality of evidence of the efficacy of new agents in phase I trials given the complexity of patient populations and study designs (Supplementary Table 2 and Supplementary Table 3). Understanding those concepts can be highly relevant to efforts for optimizing dosing of new therapeutic agents, including the transformative US FDA project Optimus ⁹⁵.

As Daniel Von Hoff said, “There are no bad anticancer agents, only bad clinical trial designs.” ⁷⁹ A new agent that is significantly active is easy to see from the very beginning of development, if used in the right population. Investigators know when they see it. The bispecific HER2/HER3 antibody, zenocutuzumab, showed encouraging activity in patients with NRG1 fusions despite earlier disappointment^{96, 97}. However, this first-hand experience gets diluted in trials with multiple sites enrolling small number of patients. Then, interpreting efficacy data in phase I trials is harder, and one needs to proceed with caution given the small numbers of patients and patient selection methods used in these trials. Good drugs and good technology applied to efficacy assessment in phase I trials will entail a better and richer late development program.

Conclusion:

Efficacy assessment in oncology phase I trials can be equally as important as safety assessment. Classic efficacy endpoints, alone, can have multiple challenges in the context of heterogenous patient populations. The integration of novel endpoints that leverage advances in available technologies into that paradigm can help to provide a better assessment of new molecules in phase I trials.

HIGHLIGHTS.

• Efficacy assessment in oncology phase I trials is crucial in the process of drug development.

• Multiple efficacy endpoints, beyond classic endpoints, have been proposed for efficacy assessment in phase I trials.

• Combination of data from multiple endpoints can enable better evaluation of efficacy outcomes.