The Impact of Survival Benefit and Study Design on FDA Approval for Anticancer Drugs Over the Past Decades.

Introduction

Drug development in oncology continues to expand, representing the most significant portion of the global pharmaceutical industry's collective R&D spending. 1 More than 2000 anticancer drugs have been under development since 2020. The number of oncology trial initiation per year has remained historically high since 2021, with 29% up from 2017. The fierce competition among oncology clinical trials has resulted in the lengthy and costly development of anticancer drugs. 2 Despite the oncology community's tremendous effort, most registrational clinical trials are unsuccessful, and the success rate from Phase 3 initiation to regulatory approval remains around 30%‐40%, less than other therapeutic areas. 3 , 4 , 5 , 6 The oncology community seeks to improve the probability of success; thus, the factors associated with development success have attracted attention.

It is intuitively expected that the magnitude of efficacy, such as tumor response rate or survival benefit, will be related to the probability of regulatory approval for anticancer drugs. In this context, several studies investigated the efficacy magnitude of US Food and Drug Administration (FDA)‐approved drugs to explore the positive factors associated with FDA approval. 7 , 8 In addition, several predictive models for regulatory approval were suggested using Phase 1 clinical efficacy outcomes, mainly using overall response rate 9 , 10 and AI‐powered algorithms. 11 However, these previous studies focused only on the efficacy aspect of the investigational drugs. Therefore, it remains to be seen whether the relationship between efficacy and regulatory approval considers other backgrounds, for example, the safety profile of the anticancer drug, therapeutic context, and study design.

According to FDA guidance on benefit–risk assessment, the FDA makes regulatory decisions for New Drug Applications (NDAs) and Biologics License Applications (BLAs) based on the clinical efficacy, safety, therapeutic context, severity of the target disease, evidence level, and uncertainties regarding the drug's benefits and risks. 12 Therefore, considering these background factors, the relationship between efficacy and regulatory approval should be reviewed. This study initially evaluated the association of such groups of background factors to the probability of FDA approval. Second, it explored the extent of the survival benefit associated with the FDA approval by adjusting all other background factors.

Methods

This study collected data from publicly available databases and journals; therefore, institutional review board approval was not required per 45 CFR §46. We searched Clinical Trial.gov for all registered Phase 3 studies measuring survival outcomes initiated from January 1, 1990 to January 1, 2021. Among the identified 3463 studies, 208 studies with 158,250 participants were extracted as the eligible randomized controlled studies in which anticancer drugs were compared with control drugs for overall survival (OS) as primary or coprimary endpoints. We tracked the FDA approval status by the FDA website (drugs@FDA) for each pivotal study and determined that 79 (38.0%) trials resulted in FDA full approval and 129 (62.0%) trials did not lead to approval. This study followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting guideline. 13 All study data were extracted by one author (Koji Ishizuka) using pre‐specified forms and verified by a second author (Shunsuke Ono). All statistical analyses were performed using Stata statistical software (StataCorp, 2022; Stata Statistical Software Release 17. College Station, TX: StataCorp LP).

The study characteristics and clinical outcomes were collected from literature identified for 208 trials. We defined three categories of explanatory variables: (1) factors of therapeutic context, (2) factors of study design, and (3) factors of study outcomes.

The factors of therapeutic context intended to depict the high‐level background behind the registrational studies, that is, the targeted diseases, the treatment line, characteristics of the study drugs, and the timing factor. In this category, we collected primary classification of cancer (carcinoma/sarcoma/hematological malignancy), cancer type (27 types), treatment line (neoadjuvant/adjuvant/chemoradiation therapy, first line, and second line or later), type of study drug (cytotoxic agents, molecular‐targeted agents including small molecule and ADC, and biologics/protein/hormone), novelty of the study drug (follow‐on drug or first‐in‐class drug), and year of study initiation (1999‐2008, 2009‐2013, and 2014‐2018).

The study design factors consisted of the characteristics of the control drugs, the study arm's design, and the primary endpoint selection. The control drugs are characterized by two aspects: type of control drugs and age of control drugs in the market. The classification of the control drug consisted of four categories: (1) cytotoxic agents, (2) molecular‐targeted small molecules, (3) biologics/protein/hormones, and (4) placebo/best supportive care. If two drugs were used in the control arm as a combination regimen, we conducted hierarchic classification by the order of this category. For instance, if the combination regimen contained a cytotoxic agent (category 1) and a biologics (category 3), we classified the combination regimen as category 1: cytotoxic agents, considering its more severe side effect than the other categories. To determine the age of control drugs in the market, we referred to the first publication regarding the control drug cited by the National Comprehensive Cancer Network (NCCN) guidelines. In addition, we confirmed the first approval date of the control drug in the United States. For the single agent control arm, the age of control drugs was derived by calculating the publication date of the Phase 3 study subtracted by the first publication date or the first approval date of the control drug, whichever comes first. For the combination regimen, we used the first publication date for the subtraction. The design of the study arm was classified into four categories, which were commonly seen in the oncologic clinical trials as follows: (1) a study drug plus the standard of care (SOC) were compared with the SOC (e.g., ramucirumab plus docetaxel vs docetaxel), (2) a study drug was directly compared to the SOC drug(s) in a head‐to‐head design (e.g., nivolumab vs docetaxel), (3) a study drug was compared with placebo or no cancer treatment for the blank space in the treatment line (e.g., durvalumab vs placebo after chemoradiation), and (4) a study drug was added to the existing SOC regimen to replace one of the drugs in the SOC regimen and compared to the conventional SOC regimen (e.g., pembrolizumab plus chemotherapy vs cetuximab plus chemotherapy, more details are in Table S1 ).

The study outcomes were collected to depict the most crucial clinical efficacy outcomes and key safety profiles observed in the registrational studies. This category included hazard ratio (HR) for OS, median OS of the control drug (less than 9 months or 9 months or more), and the study drug's safety profile compared to the control drug. As the Common Terminology Criteria for Adverse Events (CTCAE) grade 3 or higher adverse events (AEs) are generally regarded as severe toxicity in oncology, we compared the frequency of the total CTCAE grade 3 or higher AEs between the study drug and control drug and classified into three categories: improved group (study drug had ≥5% less ≥G3 AE than control drug), equivalent group (study drug had similar [+/−5%] ≥G3 AE to control drug), and worsen group (study drug had ≥5% more ≥G3 AE than control drug).

We classified the exploratory variables into six groups representing distinct facets of drug development: target disease, study drug, study timing, study design, safety outcome, and efficacy outcome. At first, we performed logistic regression analysis using each group of exploratory variables and estimated the association of each facet to FDA approval. Lastly, we integrated all exploratory variables into one model (configuration 7) to explore the significant variables for FDA approval considering all facets of drug development. A P ‐value of less than 10% was considered statistically significant. Pseudo R 2 and Akaike information criteria (AIC) were used to evaluate the model's fitness. The association between HR for OS and the probability of FDA approval was depicted using the estimated parameters in the regression model.

To understand the applicability and limitation of the suggested logistic regression model (configuration 7), we also explored the impact of study robustness and time factors on the model. Using configuration 7, we analyzed the association between FDA approval and the factors related to the study's robustness (sample size, 95% confidence interval [CI] on HR for OS, P ‐value for the OS test, and the OS test results). In addition, we examined the influence of the time factor (year of study initiation) on the logistic regression model by incorporating the actual year of study initiation. Furthermore, we performed a correlative analysis between the year of study initiation and study characteristics, which could change from 1999 to 2018 (i.e., drug type of study drug, class of novelty, type of control drug, age of control drug in the market, primary endpoint type, arm design, and safety profile of study drug characterized by ≥grade 3 adverse events).

Results

Two hundred and eight trials and 158,250 participants were included in this exploratory study (the data source shown in Table S2 ). Among them, 79 (38.0%) trials are FDA‐approved groups, and 129 (62.0%) are FDA‐non‐approved groups. Table 1 shows the FDA approval rate by the exploratory variables. Among the factors of therapeutic context, the follow‐on drugs and the year of study initiation showed different ratios by the FDA approval status ( P = .003 and .061, respectively, Table 1 ). The follow‐on drugs had a higher approval rate (50.6%) compared to the first‐in‐class drugs (29.9%). The studies initiated in 2014‐2018 had a relatively higher FDA approval rate (49.3%) than the studies initiated in 1999‐2008 (32.6%) and 2009‐2013 (32.3%). Among the study design factors, the type of primary endpoint and the age of control drug in the market had a different ratio by the FDA approval status ( P = .057 and P = 0.054, respectively). The studies with coprimary endpoints with OS and other time‐to‐event endpoints had higher FDA approval rates (48.3%) than those with OS as the single primary endpoint. The studies with 16‐year‐old or less control drugs showed a relatively higher approval ratio (44.3%) than the studies with more than 16‐year‐old control drugs (31.4%).

The Efficacy Boundary Between Approval and Non‐Approval

Figure 1 shows the histograms of HR for OS by the subgroup defined by exploratory variables, which was significantly associated with the FDA approval. The red‐colored bins represent the number of FDA‐approved drugs, whereas the gray‐colored bins indicate FDA‐non‐approved drugs. The red and gray overlap suggests the boundary separating the FDA‐approved and FDA‐non‐approved drugs. In addition, the ratio between the red and gray areas in the histogram visually represents the FDA approval ratio in the group. Although this visualization did not provide statistical precision, the histogram showed the consistent tendency on the FDA approval ratio observed in Pearson's chi‐squared test. In particular, the follow‐on drugs showed a greater red area than the first‐in‐class drugs. The year of study initiation showed a greater red area between 2014 and 2018 compared to the studies initiated between 1999 and 2008 and 2009 and 2013. If the control drug was older than 16 years in the market, the red area appeared much less than in the studies with the control drug 16 years or less old.

The histogram by safety profile showed a unique trend among the improved, similar, and worsened groups in the frequency of CTCAE grade 3 or more adverse events. The safety‐profile‐improved group showed a large distribution in smaller HR for OS, which meant this group had greater efficacy and an improved safety profile. The mean HR of this group was 0.807. The FDA approval ratio in this group was more significant than in the other groups. In contrast, the safety‐profile‐worsened group showed a mean HR of 0.914 and a more significant number of FDA‐non‐approved drugs with a peak around HR 0.90‐1.0. It implied that the safety‐profile‐worsen group had marginal or no efficacy compared to the control drugs. Thus, it was reasonable that the FDA did not approve them. Reflecting this distribution, the FDA approval ratio in the safety‐profile worsen group was the lowest among the three groups. The safety‐profile‐similar group showed a mean HR of 0.878, which is in the middle of the mean efficacy of the improved and worsened groups. Interestingly, all groups showed a similar overlap area around HR 0.80, which suggested that the threshold between FDA approval and non‐approval remained similar among the three groups.

Exploration of Factors Associated with the Probability of FDA Approval

Table 2 shows the coefficients and P ‐value for the multivariate logistic regression analysis with the variables in target disease, study drug, study timing, study design, safety outcomes, efficacy outcomes, and all areas. In the target disease‐related variables, the cancer type and treatment line did not significantly correlate with FDA approval. Of the characteristics of the study drug, the follow‐on drug showed a positive association with the FDA approval (0.816, P = .007). For the study timing, the studies initiated in 2014‐2018 showed a positive coefficient (0.697, P = .078), which may suggest that the later years may have a higher FDA approval rate than the base year 1999‐2008.

In the study design factors, the age of the control drug in the market showed a negative association with FDA approval (−0.911, P = .018). This meant that if the control drugs were older than 16 years, the probability of FDA approval would be lower. In addition, the study arm design to replace SOC with the study drug showed a positive association with FDA approval compared to the base case: adding a study drug to the SOC (0.966, P = .006). Of the safety outcomes, the safety‐profile‐improved group in CTCAE grade 3 or greater AE showed a positive coefficient compared with the safety‐profile‐equivalent group (0.845, P = .038). This meant that the safety‐profile‐improved group had a higher likelihood of approval from the FDA. In the efficacy outcomes, HR for OS showed a negative association with significance (−23.955, P < .001). The higher HR for OS (i.e., less survival benefit) significantly reduced the probability of FDA approval.

With all the exploratory variables in a model, the HR for OS continued to show a similar negative association with the FDA approval (−29.907, P < .001). In addition, the age of the control drug in the market (−2.581, P = 0.008) showed significant association with FDA approval. In addition, the partial replacement of the SOC regimen was positively associated with FDA approval (4.012, P = .079). Among them, the HR for OS was the most influential factor in the probability of FDA approval. The relationship between HR for OS and the probability of FDA approval using the model with all exploratory variables (configuration 7) and efficacy variables (configuration 6) are shown in Figure 2 . The reverse sigmoidal curve with configuration 7 showed a steep slope around HR 0.75 to HR 0.84, with 80% and 20% estimated probability of FDA approval, respectively. A 50% probability of FDA approval was estimated at HR 0.795. The curve based on configuration 6 showed a similar slope and largely overlapped with configuration 7. The scatter plot of the actual probability of FDA approval was slightly higher than the curves estimated by these models; however, it was within a 95% upper confidence interval.

Explore the Characteristics and the Limitations of the Suggested Model

To further understand the characteristics and the limitations of the suggested model configuration 7, we assessed the potential impact of the factors related to the study's robustness and timing. Configuration S1 in Table S3 incorporated the sample size of the studies, 95% confidence intervals on HR for OS, and the P ‐value (0: P ≥ .05, 1: .005 ≤ P < .05, 2: P < .005) for the OS test, which is potentially related to the robustness of the study outcomes. In this configuration, HR for OS and the age of control drugs in the market showed significant association with FDA approval (−32.346, P = .023 and −7.145, P = .015, respectively). In addition, the P ‐value (.005 ≤ P < .05 and P < .005) exhibited a significant positive association with FDA approval (5.995, P = .002 and 3.725, P = .087, respectively). This result was consistent with the high agreement between P ‐value and the FDA approval status shown in Table S4 .

Additionally, we incorporated the OS test results (0: negative, 1: positive) instead of the P ‐value. However, the model did not provide convergence, and no explanatory variables showed significant association. This instability may be caused by the high agreement between the OS test results and FDA approval status shown in Table 3 . Only 3.1% (four studies) of FDA‐approved drugs had negative results for the OS tests, and 5.1% (four studies) of FDA‐non‐approved drugs had positive results for the OS tests. Regarding the association level, the OS test results appear to have a higher magnitude than the HR for OS.

As our dataset covered a long duration ranging from 1999 to 2018 for the study initiation and 2006 to 2021 for the publication, the study characteristics may have significantly changed over time. To assess such sensitivity in our dataset, we performed correlative analysis. Table 4 shows the distribution of the study characteristics by year of study initiation. As reflected the history of drug development over two decades, several characteristics showed significant correlations with the time factor: the ratio of the biologics/protein/hormone of study drugs ( P < .001), the ratio of the follow‐on drugs ( P < .001), the ratio of the study design with coprimary endpoints using OS and other time‐to‐event endpoints ( P < .001), the ratio of the study arm design for replacement of SOC ( P = .007), and the ratio of the safety‐profile‐improved group ( P < .001) exhibited significant and consistent increase over the time from 1999 to 2018. In contrast, placebo/best supportive care/steroid for the control arm significantly decreased over time. Interestingly, the distribution of the age of the control drug in the market showed no significant changes. Our dataset and the suggested model were considered impacted by these time‐sensitive factors. In addition, the number of publications and the ratio of negative studies for OS tests exhibited certain changes over these decades. Figure 3 shows the time course of the number of publications by year and the breakdown of the OS test results. The sample size until 2012 was relatively smaller than that after 2013. In addition to the potential publication bias in the early days, the smaller sample size may have been reflected in the higher volatility of the negative trial ratio during 2006‐2012 (0% to 67%). After 2013, the negative trial ratio appeared more stable, around 60% to 70%, which may represent the higher attention in the industry and academic community to prevent publication bias in recent years. This tendency to report may potentially impact our dataset. To examine such time factor effect in the regression model, we introduced the year of study initiation (i.e., actual year, instead of categorical value for study start year) into configuration 7. However, this configuration did not provide a significant association between the year of study initiation and the probability of FDA approval (−0.137, P = .255, the data shown in Table S5 ). Therefore, the influence of the time factor was considered limited to the explanatory variables. Even in this model, HR for OS and the age of the control drug in the market exhibited significant association (−29.964, P < .001 and −2.252, P = 0.012, respectively), consistent with configuration 7.

Discussion

Considering the therapeutic effect of anticancer drugs, it is reasonable that the probability of FDA approval is associated with efficacy outcomes. This cross‐sectional study explored the extent to which the survival benefit is associated with FDA approval and whether the other background factors influence the relationship between efficacy outcomes and FDA approval. As expected, the HR for OS was an influential factor in the FDA approval, even in the model that adjusted 20 other factors from the therapeutic context, study design, and safety outcomes. The coefficient of HR for OS (−29.907, P < .001) suggests a sharp correlation between HR for OS and the probability of FDA approval. If the HR for OS changes by 0.1 from HR 0.75 to HR 0.85, the probability of FDA approval drops by 63.2% from 79.6% to 16.4%. This finding underscores the importance of survival benefits in the regulatory decision for anticancer drugs. Interestingly, the model using only efficacy outcomes showed a similar curve for the probability of FDA approval to the model with all factors. The FDA may maintain the required effect size level in survival benefit to a large extent independently from the overall therapeutic context, study design, and other outcomes.

The age of the control drug also consistently showed a negative association with FDA approval in both the model with only study design variables (−0.911, P = .018) and the model with all variables (−2.581, P = .008). If the control drugs were older than 16 years, the probability of FDA approval would decrease by 23% compared to the control drug aged equal to or less than 16 years in the model using all variables (Figure S1 ). The mean HRs for both subgroups by the age of the control drug were similar (HR 0.877 and HR 0.876, respectively). However, the overlap area between FDA‐approved and FDA‐non‐approved drugs on the histograms showed notable differences by the age of the control drug. The overlap was observed from HR 0.80 to HR 0.90 for control drug aged equal to or less than 16 years. Contrastingly, HR 0.70 to HR 0.80 was the overlap area for the control drugs aged more than 16 years. The difference suggests that if the registrational study uses a control drug older than 16 years, the required effect size by the FDA may rise by approximately HR 0.1. In addition, it was shown that the HR for OS and the age of control drug have significant interaction in the model using interaction term (Table S6 ). The FDA may raise the bar by the age of the control drug. However, using old control drugs may also interact with other confounding factors that were not used in our analysis, for example, other available treatments, heterogeneity of the target tumor, or the stage of the disease. The observed higher bar with older control drugs may represent the influence of such multiple confounding factors for FDA approval.

While the HR for OS and age of control drug showed consistent association with the probability of FDA approval across models, the safety profile, the class of novelty, and the arm design showed inconsistent association.

The study drug's safety profile based on the CTCAE grade 3 or more AE rates showed an interesting tendency across the regression models. The safety‐improved group positively associated with FDA approval (0.845, P = .038) in the model factoring only safety variables. However, this association became non‐significant (0.925, P = .479) in the model with all variables, including HR for OS. The histogram by safety subgroup helps discuss this difference (Figure 1 ). The histogram indicated that the safety‐improved group had higher efficacy (i.e., mean HR 0.807 for OS) than the safety‐worsen group (i.e., mean HR 0.914 for OS). This collinearity between the safety profile and HR for OS may play a role in changing the association of the safety‐improved group to the FDA approval in the model using HR for OS.

Moreover, in the closer histogram observation, the overlap areas between the FDA‐approved and FDA‐non‐approved groups were similar among the safety‐improved, safety‐equal, and safety‐worsened groups. It implied that the FDA may have retained its standard for HR for OS regardless of the safety profile in CTCAE grade 3 or higher AE rates. It is often said that “there is a trade‐off relationship where a drug with high safety is approved even with low efficacy.” Conversely, “a drug with high efficacy is approved even if there are some safety issues.” However, the present results did not support such a plausible trade‐off concept.

The follow‐on drug showed a positive association with FDA approval in the model with only study drug variables (0.816, P = .007), meaning the follow‐on drug would have a higher likelihood of obtaining FDA approval than the first‐in‐class drugs. It may be counterintuitive since the first‐in‐class drug is generally demanded in the oncology community, and the “me‐too” drugs are criticized. However, as shown in the histogram (Figure 1 ), the follow‐on drugs had higher efficacy (i.e., mean HR 0.811 for OS) than the first‐in‐class drugs (i.e., mean HR 0.919 for OS) as a whole tendency. This tendency was likely translated to the higher approval rate in the follow‐on drugs than the first‐in‐class drugs. Similar to the safety profile, this positive association between the follow‐on drugs and FDA approval was lost in the model using 21 variables from all areas. It may also be attributed to the collinearity of the follow‐on drugs to the HR for OS.

In the study design aspect, the arm design with intent to replace the SOC by the study drug, that is, head‐to‐head comparison, showed a positive association to FDA approval in the model factoring only study design variables (0.966, P = .006); however, lost association in the model factoring all variables (−1.891, P = .158). Conversely, the arm design to partially replace the SOC regimen may have positively affected the FDA approval in this model (4.012, P = 0.079). In the actual regulatory review, the arm design may have some meaning in the interpretation of data. However, the contribution may be situational.

Notably, the other factors, including year of study initiation, cancer types, treatment lines, type of study drug, type of control drug, primary endpoint, and median OS of control drug, did not show significant association with FDA approval in our models. Our dataset included clinical studies over the past decades since 1999; therefore, no significant association between study timing and FDA approval implied that the FDA had maintained its bar for regulatory approval over the past decades.

The additional analyses to assess the potential impact of the factors related to the study's robustness to the model have indicated that several statistical parameter estimates resulting from trials (e.g., the width of the confidence interval on HR for OS and P ‐value) may have influenced the approval statuses under some complex relationships. These additional analyses also have indicated that the HR for OS and the age of the control drugs consistently showed a strong relationship with approval even after taking these factors into account.

This study also revealed that several factors in the statistical results may be involved in the actual (i.e., value) judgment in the FDA's evaluation (Tables S3 and S4 ); however, the causal relationships between them are complex, and it is difficult to investigate them causally using a normal regression model. For example, if multiple estimates of statistical parameters are included as variables, there is a possibility of a collider bias occurring in analysis.

Overall, our analysis suggests that the efficacy outcome, that is, survival benefit measured by HR for OS, is an influential factor in FDA approval among other faucets of clinical development. Consistently, several previous studies reported a positive association between efficacy outcomes and the probability of FDA approval. Malik et al. (2014) investigated the efficacy and regulatory outcomes of 88 anticancer agents and reported objective responses in Phase 1 studies related to FDA approval. 9 Similarly, Malek et al. (2017) explored the association between the efficacy signals in Phase 1 and FDA approval status using anticancer drugs for multiple myeloma. This study reported that study drugs with higher overall response rates (ORR) in Phase 1 trials were more likely to advance to Phase 2 trials and achieve FDA approval. 10 Despite the primary objective of Phase 1 being to evaluate the safety profile and determine the dose for the next phase, these studies consistently reported that the efficacy outcomes exhibited a more significant correlation with the ultimate regulatory approval. Our study employed the registrational Phase 3 trials that differ from these previous studies; however, we commonly found a significant correlation between the efficacy outcomes and the FDA approval status.

Limitations

This study has several intrinsic limitations. First, we focused on only the registrational Phase 3 studies measuring OS as a primary or coprimary endpoint. Therefore, our finding should not be extrapolated to the clinical study with other primary endpoints, for example, progression‐free survival (PFS) or event‐free survival (EFS). Second, there may be publication bias due to the retrospective approach and high reliance on the published data. The ratio of negative studies for the OS test showed higher volatility, especially for the early period from 2006 to 2012. Although the year of study initiation did not show a significant association on the probability of FDA approval, the explanatory variables may have been affected by the time factor. Third, although we tried to capture critical facets of clinical development into the model, the 21 exploratory variables may only represent part of the development program. For instance, the dataset did not capture information regarding the balance in baseline characteristics, post‐study treatment, and regulatory decisions not directly related to the clinical study outcomes, for example, product quality issues. Our findings do not apply to cases where these other factors primarily determine the regulatory outcomes. In addition, particular AEs critically determine the risk and benefit balance in drug development. For example, QT prolongation, interstitial lung disease, hepatotoxicity, and central nervous system toxicities are known as critical AEs. 14 , 15 , 16 , 17 , 18 The safety profile variables in this study did not capture such individual critical AEs. Fourth, although the HR for OS showed a significant association with FDA approval, the statistical test results for OS have a higher association with the FDA approval status. Therefore, the correlation between HR for OS and FDA approval was not observed if the model incorporated OS test results. Regarding the prediction for FDA approval, the statistical test results have the highest predictability among any factors in this study. Therefore, our finding regarding the association between HR for OS and the FDA approval should be considered to provide a complementary perspective to the OS test results, particularly in the case the magnitude of the efficacy matters for the regulatory decision. Finally, although the study implied that several statistical parameter estimates resulting from trials may have influenced the approval statuses, normal regression analysis is limited in investigating such complex relationships. It is necessary to adopt analyses incorporating these complex components and relationships to investigate how these factors affect the FDA's evaluation.

For the extrinsic limitation of the applicability of this study, it is essential to note that this study did not cover the post‐Project Optimus era. The FDA initiated the Project Optimus in 2021 to change the dose selection paradigm for anticancer drugs. Essentially, the FDA required the pharmaceutical industry and the investigators to shift from the conventional maximum tolerated dose (MTD)‐based approach, which is basically for cytotoxic chemotherapies, to the modernized approach using exposure–response analysis, PK/PD modeling, and randomized study design to compare multiple target doses. The Project Optimus and the subsequent FDA's guidance for the industry regarding dose optimization (2024) emphasizes the importance of exploring a safer profile, which the FDA expects to result in better tolerability, longer duration of treatment, and thereby, maximal clinical benefit for the patients. 19 Since 2021, the Project Optimus‐recommended approaches have been gradually assimilated in the development programs. However, our dataset only covers the clinical trials until 2021, namely the pre‐Project Optimus era. Therefore, our finding, especially for the association between the safety profile and the FDA approval status, should have limitations to the post‐Project Optimus era. When we consider the application of our findings to future trials, it should be noted that such limitations exist. Further research is warranted for the registrational trials initiated after the Project Optimus.

Conclusion

The HR for OS exhibited a remarkable association with the probability of FDA approval. This study visualized the probability of FDA approval over the HR for OS and suggested the threshold with a 50% success rate of FDA approval at HR 0.795 for OS. The threshold was stable even though the other factors from the therapeutic context, study design, and safety outcomes were incorporated into the model. Furthermore, the P ‐value for the OS test and the width of the confidence interval on HR for OS showed a significant association with the probability of FDA approval. These findings underscore the rigorous standard required for new anticancer drugs to obtain regulatory approval from the FDA.

Author Contributions

Koji Ishizuka and Shunsuke Ono designed the research and wrote the manuscript. Koji Ishizuka performed and analyzed the data. Shunsuke Ono verified the data and analysis. Koji Ishizuka and Shunsuke Ono gave final approval for the publication of the version. Koji Ishizuka and Shunsuke Ono agreed to be accountable for all aspects of the work and to ensure that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Conflicts of Interest

Koji Ishizuka is employed by Daiichi Sankyo, Inc. Shunsuke Ono declares no competing interests for this work.

Funding

Funding from Ministry of Education, Culture, Sports, Science and Technology (MEXT): Shunsuke Ono KAKEN‐HI:19K07215.

Supporting information

Supporting Information

Data Availability Statement

Data available from the corresponding author upon request.

References

1. Global Oncology Trends 2023 Outlook to 2027 . IQVIA Institute. 2023. Accessed February 11, 2024. https://www.iqvia.com/insights/the‐iqvia‐institute/reports‐and‐publications/reports/global‐oncology‐trends‐2023

2. Paul SM, Mytelka DS, Dunwiddie CT, et al. How to improve R&D productivity: the pharmaceutical industry's grand challenge. Nat Rev Drug Discov. 2010;9:203‐214.

3. Wong CH, Siah KW, Lo AW. Estimation of clinical trial success rates and related parameters. Biostatistics. 2019;20:273‐286.

4. Hay M, Thomas DW, Craighead JL, Economides C, Rosenthal J. Clinical development success rates for investigational drugs. Nat Biotechnol. 2014;32:40‐51.

5. Thomas DV, Burns J, Audette J, et al. Clinical Development Success Rates 2006–2015. BIO Industry Analysis; 2016.

6. Thomas DW, Chancellor D, Micklus A, et al. Clinical Development Success Rates and Contributing Factors 2011–2020. BIO/QLS Advisors/Informa UK Ltd.; 2021.

7. Ladanie A, Schmitt AM, Speich B, et al. Clinical trial evidence supporting US Food and Drug Administration approval of novel cancer therapies between 2000 and 2016. JAMA Netw Open. 2020;3:e2024406.

8. Gyawali B, Hey SP, Kesselheim AS. Assessment of the clinical benefit of cancer drugs receiving accelerated approval. JAMA Intern Med. 2019;179:906‐913.

9. Malik L, Mejia A, Parsons H, et al. Predicting success in regulatory approval from Phase I results. Cancer Chemother Pharmacol. 2014;74:1099‐1103.

10. Malek E, Saygin C, Ye R, et al. Predicting successful phase advancement and regulatory approval in multiple myeloma from phase I overall response rates. JCO Clin Cancer Inform. 2017;1:1‐14.

11. Beinse G, Tellier V, Charvet V, et al. Prediction of drug approval after phase I clinical trials in oncology: RESOLVED2. JCO Clin Cancer Inform. 2019;3:1‐10.

12. US Food and Drug Administration . Benefit‐risk assessment for new drug and biological products guidance for industry. 2023. Accessed February 11, 2024. https://www.fda.gov/regulatory‐information/search‐fda‐guidance‐documents/benefit‐risk‐assessment‐new‐drug‐and‐biological‐products

13. von Elm E, Altman DG, Egger M, Pocock SJ, Gotzsche PC, Vandenbroucke JP; STROBE Initiative . Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement: guidelines for reporting observational studies. Lancet. 2007;370:1453‐1457.

14. Roden DM. A current understanding of drug‐induced QT prolongation and its implications for anticancer therapy. Cardiovasc Res. 2019;115:895‐903.

15. Gemma A. Drug‐induced interstitial lung diseases associated with molecular‐targeted anticancer agents. J Nippon Med Sch. 2009;76:4‐8.

16. Min JH, Lee HY, Lim H, et al. Drug‐induced interstitial lung disease in tyrosine kinase inhibitor therapy for non‐small cell lung cancer: a review on current insight. Cancer Chemother Pharmacol. 2011;68:1099‐1109.

17. Vincenzi B, Armento G, Spalato Ceruso M, et al. Drug‐induced hepatotoxicity in cancer patients ‐ implication for treatment. Expert Opin Drug Saf. 2016;15:1219‐1238.

18. Peddi PF, Peddi S, Santos ES, Morgenstern D. Central nervous system toxicities of chemotherapeutic agents. Expert Rev Anticancer Ther. 2014;14:857‐863.

19. US Food and Drug Administration . Optimizing the dosage of human prescription drugs and biological products for the treatment of oncologic diseases. Guidance for industry. 2024. Accessed October 6, 2024. https://www.fda.gov/regulatory‐information/search‐fda‐guidance‐documents/optimizing‐dosage‐human‐prescription‐drugs‐and‐biological‐products‐treatment‐oncologic‐diseases

Associated Data

Supplementary Materials

Supporting Information

Data Availability Statement

Data available from the corresponding author upon request.