Can Machine Learning Overcome the 95% Failure Rate and Reality that Only 30% of Approved Cancer Drugs Meaningfully Extend Patient Survival?

Duxin Sun*, Christian Macedonia, Zhigang Chen, Sriram Chandrasekaran, Kayvan Najarian, Simon Zhou, Timothy A. Cernak, Vicki L. Ellingrod, Hosagrahar V. Jagadish, Bernard Marini, Manjunath P. Pai, Angela Violi, Jason C. Rech, Shaomeng Wang, Yan Li, Brian Athey, Gilbert S. Omenn

Dr. Duxin Sun et al from College of Pharmacy, College of Engineering, School Medicine, Michigan Institute of Data Sciences at University of Michigan, together with colleagues from Lancaster Life Science Group, Aurinia Pharmaceuticals, and Bristol-Meyer Squib Company, jointly published a perspective on J Med Chem, 2024, xxx, entitled: “Can machine learning overcome the 95% failure rate and reality that only 30% of approved cancer drugs meaningfully extend patient survival?”

Cancer care has significantly advanced over the past 30 years. The overall cancer mortality rate in the U.S. has dropped by 33% from 1991-2020, and  the FDA has approved 250 cancer drugs targeting 120 molecular targets from 2000-2022. Despite this progress, however, the U.S. still had over 610,000 cancer-related deaths in 2023. This equates to 1,670 deaths per day, a figure comparable to the U.S. daily COVID-19 death rate during the peak of the pandemic in 2020. Globally, there were 10 million cancer-related deaths in 2022, which is three times more than the peak COVID-19 deaths during the worldwide lockdown in 2020.

Various global research initiatives have been launched to fight the war on cancer. The U.S. reignited the Cancer Moonshot, aiming to reduce the cancer death rate by 50% over the next 25 years. However, the financial burden in cancer care has posed a challenge to achieve the goal of global research initiatives. For instance, the average cost of a course of cancer drug regimen per patient in the U.S. ranges from $170,000 to $277,000. The U.S. spends $200B annually on cancer care, including $60B on cancer drugs, translating to daily costs of $540M and $165M, respectively.

In this perspective, the authors focus on discussion of the following questions:

• Why does 95% of cancer drug development fail despite significant improvement at each step of the process using hundreds of helpful strategies in the past 30 years?

• Why do only 30% approved cancer drugs meaningfully extend patient survival by more than 2.5 months, while the average cost for a cancer drug regimen per patient in the U.S. ranges from $170,000 to $277,000?

• Are current strategies to improve each step of drug development process, including AI-driven approaches, falling into “survivorship bias” trap by overly focusing on many less critical issues but overlooking root causes of key aspects?

• Is it practical to add more criteria to the already lengthy and costly drug development process, which takes 10-15 years and costs $1-2 billions per one approved drug?

• Can application of AI and machine learning (ML) methodology in the current drug development process improve efficiency, boost success rates, and enhance drug’s efficacy?

• What are the root causes of these challenges, and what should be the future research priorities to find potential solutions?

(1). The 95% failure rate of cancer drug development has not improved over the past 30 years.

Hundreds of helpful strategies and criteria have been implemented in the drug development process in the past three decades, which includes target validation, high throughput screening, lead compound optimization, preclinical testing, drug-like property optimization, GLP toxicity testing, GMP manufacturer, and clinical phase I-III trials. Despite the improvements in each step of the process, 95% of small molecule cancer drug development fails from clinical phase I to phase III trials. (Figure 1)

Given the lengthy and costly nature of cancer drug development, it is not practical to continually add more criteria to the process, without eliminating non-essential ones. This raises the question: which criteria are nonessential? This high and persistent failure rate, seen in the in the pharmaceutical industry but not others, may suggest that the existing criteria may have overlooked key aspects of the drug development process (Figure 1). More importantly, do these current strategies fall into the “survivorship bias” trap by overly focusing on many unimportant factors but overlooking major deficiencies? This analogy compares the focus on fixing the damage on the wings of the surviving/returned aircrafts where they can sustain damage and still return home, while overlooking the damage to the critical location, such as engines and cockpits, of the non-surviving/un-returned aircrafts during WWII (Figure 2). While many of current criteria in the drug development process are indeed important, many others may just fix non-essential problems and miss the critical aspects in the process similar to the scenario in survivorship bias.

(2). Only 20-42% of newly approved cancer drugs meaningfully extend patient overall survival beyond 2.5 months.

Due to unmet clinical needs, many cancer drugs are approved based on clinical phase II trials demonstrating improvements in surrogate endpoints, such as progression free survival (PFS), delaying the time to tumor progression. These drugs are then required to be tested in clinical phase III trials to validate their effectiveness in prolonging patient overall survival (OS). The early access to such drugs may benefit cancer patients who have no other treatment options. However, it is disappointing that only 20-42% of these newly approved cancer treatments meet the clinical benefit threshold in prolonging overall survival beyond 2.5 months, a criterion set by the American Society of Clinical Oncology (ASCO) and the European Society for Medical Oncology (ESMO). Approximately 30% of approved cancer treatments did not improve OS, while 40% improved OS by 2 weeks to 2.5 months.

Several different scenarios may exist for the lack of meaningful improvement in overall survival (OS>2.5 months). If newly approved cancer drugs show less than 2.5-month OS improvement compared with standard care, they may offer similar effectiveness to the standard care but may have advantages, such as reduced toxicity, which is still meaningful. However, If these newly approved drugs show a less than 2.5-month OS improvement in comparison with an inappropriate control arm, caution should be exercised regarding their efficacy. In many cases, such suboptimal drug control arms with minimal clinical benefit have remained on the market and have even been listed on cancer treatment guidelines. It is very challenging to discern if these control arms would provide meaningful clinical benefits to cancer patients. Further, If the newly approved cancer drugs (or in combination with standard care) show no OS improvement when compared with placebo (or in combination with standard care), their efficacy is questionable.  Unfortunately, many drugs, which fail to improve patient overall survival, still remain on the market for clinical use based on their original “positive” phase II trials, while effective communication with physicians and patients is often lacking

The three overlooked deficiencies, including insufficient on-/off-target validation, incorrect application of drug-like property evaluation, and inability of clinical dose optimization,  contribute to the 95% cancer drug development failure rate or the low efficacy/safety of approved cancer drugs.

(1). Current drug candidate selection from in vitro target binding IC50s overlooks potency/specificity (PS) against on-/off-target in tumors at relevant clinical doses

(2). Current drug candidate selection from drug-like property criteria overlooks on-/off-target-driven tissue/cell selectivity (TS) influencing adverse effects (on/off-target toxicity) in normal organs at relevant clinical doses.

(3). Clinical doses cannot be optimized to balance clinical efficacy/safety due to the overlooked optimization of in vivo potency/specificity and tissue/cell selectivity, while extensive GLP animal toxicity testing is not predictive for clinical adverse effects (on/off-target toxicity) at therapeutic dose ranges.

We proposed a  “STAR System” (structure-tissue/cell selectivity-activity relationship) to describe how to use these three factors to classify drugs into STAR class I – IV drugs. The STAR system could guide drug development strategies by streamlining the process to improve success rate and efficiency. The goal is to select STAR class I drugs, rather than focusing on STAR class II/IV compounds as seen in the current process.

3. How can machine learning overcome the 95% failure rate and reality that only 30% of approved cancer drugs meaningfully extend patient overall survival?

Emerging technologies like artificial intelligence (AI) and machine learning (ML) may improve the efficiency of each step of the drug development process by reducing cost/time if the ML models are correct and validated. However, it is an oversimplification to assume that applying AI and ML techniques could overcome the 95% cancer drug development failure rates if we still rely on the current and same process in the drug development.

(1). Apply ML models to improve efficiency of the drug development:

ML models have been used to improve the criteria at each step of the current drug development process. A well validated ML model can improve efficiency by saving time and cost, without the need for tedious wet lab experiments, which is very meaningful. Indeed, ML models have been applied at every step in the current drug development process (Figure 1), which includes disease target identification/validation in various cancer types, high throughput screening, drug design and optimization, drug-target interaction (DTI) prediction, pharmacokinetics and drug-like properties prediction, toxicity prediction, and clinical trial designs and pharmacovigilance. Examples of how ML models have impacted each step of the drug development process to improve efficiency by saving time and cost.

(2). Develop STAR-guided ML models to improve success rate and efficiency of the drug development:

The STAR-guided ML models need to be developed to directly predict the delicate balance of clinical dose/efficacy/safety as defined in the STAR system. The STAR-guided ML models can help to design STAR class I drugs, which  improve success rate and efficiency. Based on our new hypothesis, the STAR-guided ML models should be built to predict clinical adverse effects (AE) (on/off-target toxicity) or clinical efficacy (E_ff) by forging links among the following five features with a dynamic weighting (w) based on in vitro/ex vivo, preclinical and clinical data at relevant clinical doses (Figure 5):

Clinical adverse effects at relevant clinical doses ŷ(AE) = ⌠(PS ˟ TS_N ˟ EX_N ˟ C_P ˟ S)_w (1)

Clinical efficacy at relevant clinical doses ŷ(E_ff) =  ⌠(PS ˟ TS_T ˟ EX_T ˟ C_P ˟ S)_w (2)

The STAR-guided ML models can be built in five steps, described below, to predict clinical adverse effects (on and off-target toxicity) ŷ(AE) or clinical efficacy (E_ff) at relevant clinical doses (Figure 5). Among five features, the ML models can use drug structure information (S) to predict three features (PS, TS_N or TS_T, C_P) under relevant clinical doses. The expression feature (EX_T or EX_N) is independent of other four features, but dependent on pathophysiology of tumor or normal organs.