The vast majority of drug development programs fail during clinical testing. Focusing on flaws in preclinical models and trial design, this article offers practical improvements through predictive modeling, biomarker validation, and data-driven innovation.
Image credit: Oleksiy Mark/Shutterstock.com
The Long Path to Approval
Drug discovery is a long and challenging journey from target idea to market approval. Across a 10- to 15-year development pipeline, most drug candidates fail, and the probability of approval from Phase I remains low, highly dependent on the therapeutic area. Despite advancements in technology and methods, research and development (R&D) productivity has largely stagnated, reflecting the trend described by Eroom’s Law.
Analyses across therapeutic areas indicate that roughly 90% of candidates entering human studies ultimately fail, most often due to inadequate efficacy, unacceptable toxicity, or poor pharmacokinetics despite robust preclinical optimization.1 The underlying challenges are multifactorial, including biological heterogeneity and imperfect target selections, preclinical models that poorly represent human biology, complex trial design and operations, stringent safety and regulatory safeguards, and escalating economic and manufacturing pressures.
This article explains why drug discovery is high-risk and slow, and demonstrates how Artificial Intelligence (AI), automation, advanced laboratory models, genomics, and precision medicine can shorten timelines and enhance success rates.
From Biology to Valid Targets
Turning biology into valid targets begins with acknowledging heterogeneity: disease mechanisms vary across cell types, tissues, and microenvironments, so targets must be defined in the right context. Druggability is approached pragmatically by selecting an appropriate modality, such as a small molecule, monoclonal antibody, ribonucleic acid interference, or clustered regularly interspaced short palindromic repeats gene editing, while balancing trade-offs in selectivity, exposure, delivery, and safety.2
Causality is strengthened with human evidence. Genome-wide association studies, whole-exome sequencing, and whole-genome sequencing identify protective loss-of-function variants that link targets to clinical phenotypes and are associated with faster progression to Phase II/III or approval. Recent reviews confirm that targets supported by robust human genetic validation, such as PCSK9, ANGPTL3, or GPR75, progress through clinical development with a higher probability of success than targets discovered solely through phenotypic screening.2
Proprotein Convertase Subtilisin/Kexin Type 9 (PCSK9) illustrates this path: natural inactivating variants lowered low-density lipoprotein cholesterol and guided inhibitor development. From the outset, teams plan translational read-through by nominating pharmacodynamic biomarkers and employing minimally invasive sampling to demonstrate the mechanism early, then align experiments with a clear Target Product Profile that specifies the population, effect size, safety bounds, and dosing. Such disciplined alignment between genetic validation, biomarker-based proof of mechanism, and Target Product Profile has been shown to reduce late-stage failures and improve return on R&D investment.2
Preclinical Models That Predict Humans
Traditional animal models and two-dimensional (2D) cultures often fail to accurately reproduce human physiology, leading to preclinical signals that do not translate effectively to people. A meta-analysis of 108 oncology agents showed only a modest relationship between animal toxicities and human adverse events, highlighting gaps in translation. The study found median positive predictive values of approximately 0.65 and negative predictive values of around 0.50, indicating that animal toxicity data have only moderate reliability in predicting human outcomes.
These gaps reflect the absence of human-like tissue architecture, immune context, biomechanics, and metabolic programs in legacy systems. Stem-cell–derived organoids and microphysiological systems (MPS) now capture key organ functions, gene–environment interactions, and patient heterogeneity with improved accuracy. Newer organoids and MPS (organ-on-chip, vascularized co-cultures) better mimic organ structure–function and patient heterogeneity, improving prediction of efficacy and safety and supporting reduction of animal use.3,4
In silico approaches now complement wet models: Absorption, Distribution, Metabolism, and Excretion (ADME) and toxicity simulators, paired with AI analysis of multi-omics and high-content imaging, help triage weak candidates before costly trials. However, reproducibility remains limited by batch variability, protocol divergence, and the absence of regulatory qualification; therefore, continued standardization and benchmarking are essential for these systems to achieve translational validation.
To realize clinical impact at scale, protocols must be standardized, variability reduced, and assays prospectively benchmarked against outcomes in early trials and registries, enabling validated, human-relevant preclinical workflows.3,4
Clinical Trials and Evidence Generation
Phase II is the dominant failure point and a poor predictor of Phase III success, largely due to small samples, surrogate endpoints, and multiplicity that inflate false positives. Mechanism-aligned endpoints and enriched populations can increase signal-to-noise by selecting patients with the relevant biomarker, phenotype, or genetic trait, while acknowledging that many biomarkers remain unvalidated and may not translate to clinical outcomes. Analysis across therapeutic classes reveals that only about one-quarter of drugs entering Phase II progress to Phase III, and fewer than half of those ultimately get approved.5,6
Adaptive designs, such as group-sequential rules, sample-size re-estimation, drop-the-loser, and seamless Phase II/III, enable teams to stop early for futility or toxicity, refine doses, and maintain type I error control with pre-specified plans. These designs are increasingly accepted by regulators and can reduce total sample size and time-to-decision by up to 30% when implemented with rigorous statistical control.
Execution still decides outcomes: proven, motivated sites; realistic eligibility; patient-friendly scheduling; strong recruitment/retention; rapid electronic capture; and vigilant, real-time monitoring protect power and data quality. Ultimately, real-world evidence can complement trials in characterizing safety, effectiveness, and generalizability when integrated with rigorous design and analysis.5,6
Safety, Ethics, Regulation, And Cost
Safety spans Nonclinical (Good Laboratory Practice, GLP) toxicology, First-in-Human (FIH) safeguards, and long-term pharmacovigilance, informed by past signals such as TGN1412 and BIA-10-2474. Recent analyses emphasize that early, harmonized safety surveillance, using systematic collection, expedited reporting, and aggregate analyses, remains inconsistent across regions and can delay identification of cross-sponsor safety trends.
Ethical oversight by Institutional Review Boards (IRBs) or Independent Ethics Committees (IECs), guided by the principles of the Belmont Report and the Declaration of Helsinki, shapes study design, pace, and ongoing risk-benefit review. Evolving guidance favors prespecification and audit-ready data through a Development Safety Update Report (DSUR) and independent Data Safety Monitoring Boards (DSMB), working with Safety Management Teams (SMT), to analyze aggregate signals and decide on adaptive or stopping actions.7
Emerging therapeutic technologies increase Chemistry, Manufacturing, and Controls (CMC) readiness requirements, making early manufacturability, quality, and release testing integral to determining whether to proceed or not. Highly capitalized costs drive the “kill-earlier” discipline, concentrating resources on programs with credible exposure, response, and translational biomarkers.
Risk-sharing partnerships, such as Pharmacoepidemiological Research on Outcomes of Therapeutics by a European Consortium (PROTECT), the Innovative Medicines Initiative (IMI), and TransCelerate BioPharma (TransCelerate), help standardize methods, align with International Council for Harmonisation (ICH) expectations, and reduce duplicative work across sponsors.
Global harmonization of safety methodologies, including the development of formal Safety Surveillance Plans, has been proposed as a way to improve data comparability and accelerate approvals.7
Building Equitable Innovation
How AI is accelerating drug discovery - Nature's Building Blocks | BBC StoryWorks
Video credit: bbcstoryworks/Youtube.com
Future drug discovery efforts should embed AI and automation across design, triage, site selection, and data quality to shorten cycles and reduce errors. Current evidence suggests that AI is highly effective in accelerating virtual screening and early-stage discovery; however, its overall impact remains limited by the quality of training data, model interpretability, and regulatory acceptance.
Explainable and hybrid AI–experimental workflows are emerging as best practices. Human-relevant models, including organoids and MPS, must be scaled with regulatory qualification and shared reference benchmarks to ensure their reliability. Genomics should inform target selection and patient enrichment, paired with companion diagnostics (CDx) and adaptive platform trials (APT). A robust data infrastructure is essential, encompassing interoperable FAIR datasets (Findable, Accessible, Interoperable, and Reusable), privacy-preserving linkages, and credible external and synthetic controls.8
Chemistry, Manufacturing, and Controls (CMC) should be modernized through continuous and modular manufacturing, real-time release, automated Quality Control (QC), and “CMC by design.” Incentive structures should be aligned through milestone-based collaborations, shared-risk agreements, and disciplined early-termination governance.
Building multi-disciplinary expertise that bridges biology, computational science, statistics, and regulatory affairs is essential. Advancing equity requires inclusive eligibility criteria, decentralized and globally distributed trial sites, and culturally and linguistically appropriate engagement, ensuring that innovations reach diverse populations efficiently, ethically, and with verifiable outcomes.8
Download your PDF copy now!
Sustaining Progress Through Validation
Drug discovery remains challenging due to biological complexity, model limitations, clinical execution risks, stringent safety and regulatory requirements, and high capital costs. Evidence across recent literature suggests that improving early predictive models, integrating human genetic validation, and embedding adaptive, data-driven strategies can substantially reduce attrition and cost.
The greatest opportunity lies in strengthening early predictive signals and improving translational alignment. Combining AI, validated human-relevant models such as organoids and MPS, genetics-guided targeting, and smarter evidence strategies, including CDx, APT, and real-world data, can increase success rates and shorten developmental timelines.
Achieving this requires rigorous validation, transparent methodologies, auditable data, and robust quality systems. Meaningful progress will initially be evident in select therapeutic areas and will continue to expand as technical capabilities, organizational culture, and infrastructure evolve across the broader ecosystem.
References
- Sun, D., Gao, W., Hu, H., & Zhou, S. (2022). Why 90% of clinical drug development fails and how to improve it? Acta Pharmaceutica Sinica B. 12(7). 3049-3062. DOI:10.1016/j.apsb.2022.02.002, https://www.sciencedirect.com/science/article/pii/S2211383522000521
- Zhang, X., Yu, W., Li, Y., Wang, A., Cao, H., & Fu, Y. (2024). Drug development advances in human genetics‐based targets. MedComm. 5(2). DOI:10.1002/mco2.481, https://onlinelibrary.wiley.com/doi/10.1002/mco2.481
- Atkins, J. T., George, G. C., Hess, K., Marcelo-Lewis, K. L., Yuan, Y., Borthakur, G., Khozin, S., LoRusso, P., & Hong, D. S. (2020). Pre-clinical animal models are poor predictors of human toxicities in phase 1 oncology clinical trials. British Journal of Cancer. 123(10). 1496-1501. DOI:10.1038/s41416-020-01033-x, https://www.nature.com/articles/s41416-020-01033-x
- Luce, E., & Duclos-Vallee, J.-C. (2025). Stem Cells and Organoids: A Paradigm Shift in Preclinical Models Toward Personalized Medicine. Pharmaceuticals. 18(7). DOI:10.3390/ph18070992, https://www.mdpi.com/1424-8247/18/7/992
- Van Norman, G. (2019). Phase II Trials in Drug Development and Adaptive Trial Design. JACC: Basic to Translational Science. 4(3). 428-437. DOI:10.1016/j.jacbts.2019.02.005, https://www.jacc.org/doi/10.1016/j.jacbts.2019.02.005
- Fogel, D. B. (2018). Factors associated with clinical trials that fail and opportunities for improving the likelihood of success: a review. Contemporary Clinical Trials Communications. 11. 156-164. DOI:10.1016/j.conctc.2018.08.001, https://www.sciencedirect.com/science/article/pii/S2451865418300693
- Samara, C., Garcia, A., Henry, C., Vallotton, L., Cariolato, L., Desmeules, J., & Pinçon, A. (2023). Safety surveillance during drug development: comparative evaluation of existing regulations. Advances in Therapy. 40. 2147-2185. DOI:10.1007/s12325-023-02492-3, https://link.springer.com/article/10.1007/s12325-023-02492-3
- Blanco-González, A., Cabezón, A., Seco-González, A., Conde-Torres, D., Antelo-Riveiro, P., Piñeiro, Á., & Garcia-Fandino, R. (2023). The Role of AI in Drug Discovery: Challenges, Opportunities, and Strategies. Pharmaceuticals. 16(6). DOI:10.3390/ph16060891, https://www.mdpi.com/1424-8247/16/6/891
Further Reading
Last Updated: Nov 13, 2025