;)
The traditional 3+3 dose escalation design has dominated Phase I oncology trials for over two decades (Le Tourneau et al., 2009, Kurzrock, 2021). Three patients per cohort, fixed escalation increments, predefined maximum tolerated dose (MTD) stopping rules. Simplicity enabled adoption. Inefficiency was tolerated because Phase I was viewed as a critical but resource-intensive step in drug development.
Machine learning models now enable dose-finding approaches—Bayesian optimal interval (BOIN) design, adaptive dosing guided by real-time pharmacokinetic/pharmacodynamic data, reinforcement learning optimization—that generate substantially more information per patient, reduce unnecessary dose escalation, and identify MTDs more precisely (Yuan et al., 2016). Yet adoption remains slow. Regulatory complexity and operational inertia create parallel barriers: regulatory guidance remains incomplete, while institutional workflows developed over decades around 3+3 designs resist change
Evidence supporting AI-driven dose-finding is robust. AI-developed compounds (where AI has been applied across target discovery, lead optimization, and/or preclinical development) that completed Phase I trials as of December 2023 showed 80-90% Phase II progression rates (Ardigen, 2025; Jayatunga et al., 2024), substantially exceeding the ~40% progression rates for traditionally developed compounds (Jayatunga et al., 2024). While Phase I progression isn’t deterministic (Phase II enrolls different populations and assesses different efficacy endpoints), this data suggests that improved dose selection and/or optimized compound characteristics translate to better Phase II outcomes.
Why hasn’t adoption accelerated?
Regulatory clarity remains incomplete. FDA guidance broadly endorses adaptive designs (U.S. Food and Drug Administration, 2019), defining adaptive designs as clinical trials employing prospectively planned opportunities to modify trial design and/or statistical procedures based on interim data, while maintaining type I error control specific statistical guidance on AI/ML methods for dose-finding remains in development. This guidance applies across Phase I through Phase III studies, with specific Phase I applications including adaptive dose escalation, enrichment strategies based on emerging biomarker data, and response-adaptive randomization. Sponsors proposing novel dose-finding approaches encounter additional scrutiny, extended review timelines, and requests for simulation studies demonstrating operational characteristics. This regulatory friction discourages innovation even when scientific rationale is compelling.
Operational inertia is powerful. Organizational change requires institutional commitment, investment in retraining, and tolerance for process uncertainty. When novel methods are scientifically compelling but operationally unfamiliar, resistance emerges from multiple stakeholders simultaneously. Clinical research organizations, contract research sites, and sponsorship teams have built workflows, training programs, and institutional knowledge around 3+3 designs. Transitioning to Bayesian methods requires retraining, protocol template updates, and uncertainty about outcomes. The switching cost—even when novel methods are objectively superior—creates resistance (Kurzrock, 2021).
Data requirements are often underestimated. AI-driven dose-finding improves efficiency when historical data and mechanistic understanding enable credible prior distributions (Yuan et al., 2016). Compounds with limited preclinical characterization, novel mechanisms, or sparse historical data cannot leverage AI effectively; the AI system lacks training data to generate reliable predictions. Sponsors often discover early in the development process that their compounds fall into this category necessitating reversion to more conservative, empirical dose-finding approaches (Yuan et al., 2016).
The practical path forward.
Current evidence indicates adoption is accelerating through pragmatic hybrid approaches. Hybrid approaches are gaining traction (Yuan et al., 2016; Precision for Medicine, 2023). Bayesian methods for dose escalation paired with 3+3-like cohort structures provide efficiency gains while maintaining operational familiarity. Real-time PK/PD monitoring enables dynamic protocol adjustments without requiring full ML infrastructure. These pragmatic adaptations balance maximum theoretical efficiency against practical adoptability.
Emerging momentum is genuine. Insilico Medicine’s demonstration of target discovery to Phase I initiation in 30 months—substantially compressing traditional timelines—used AI throughout the development pipeline. The IPF program (INS018_055) exemplifies this acceleration: target discovery and drug candidate identification required 18 months (compared to typical 3–6 years); preclinical development and Phase 0 required an additional 12 months, reaching Phase I in under 30 months total (Insilico Medicine, 2022). As similar case studies accumulate, competitive pressure will accelerate adoption (Ardigen, 2025).
The Phase I landscape is transitioning, but adoption lag across organizations, regulatory bodies, and industry partnerships still exceeds scientific capability (Kurzrock, 2021).
References:
Ardigen. (2025, December 16). AI CROs & pharma shift: The next wave of drug discovery innovation. Retrieved from https://ardigen.com/the-next-wave-of-drug-discovery-innovation/ Last accessed 15JAN26
Insilico Medicine. (2022, February 23). From start to phase 1 in 30 months: AI-discovered and AI-designed anti-fibrotic drug enters phase I clinical trial. Retrieved from https://insilico.com/phase1 Last accessed 15JAN26
Jayatunga, M. K. P., Ayers, M., Bruens, L., Jayanth, D., & Meier, C. (2024). How successful are AI-discovered drugs in clinical trials? A first analysis and emerging lessons. Drug Discovery Today, 29(6), 104009. https://doi.org/10.1016/j.drudis.2024.104009
Kurzrock, R., Lin, C.-C., Wu, T.-C., Hobbs, B. P., Carmagnani Pestana, R., & Hong, D. S. (2021). Moving beyond 3+3: The future of clinical trial design. American Society of Clinical Oncology Educational Book, 41, e133–e144. https://doi.org/10.1200/EDBK_319783
Le Tourneau, C., Lee, J. J., & Siu, L. L. (2009). Dose escalation methods in phase I cancer clinical trials. Journal of the National Cancer Institute, 101(10), 708–720. https://doi.org/10.1093/jnci/djp079
Yuan, Y., Hess, K. R., Hilsenbeck, S. G., & Gilbert, M. R. (2016). Bayesian optimal interval design: A simple and well-performing design for phase I oncology trials. Clinical Cancer Research, 22(17), 4291–4301. https://doi.org/10.1158/1078-0432.CCR-16-0592
