By Ryan Neely, Ph.D.
New FDA Guidance for Digitizing Preclinical Testing
On April 10th, the FDA announced a plan to phase out animal testing requirements for investigational new drug (IND) applications1. This move underscores a growing trend to tackle the high cost and low success rates that plague drug development by improving in vitro and in silico approaches to predicting safety and efficacy in humans. In many ways this makes sense – the predictive validity of animal models of disease has proven to be low or at best very narrow. In other words, selecting drugs that are safe and effective in rodents doesn’t always translate well to humans. This may in part explain the abysmal failure rates of clinical trials – oncology is historically bad at 97%2,3. The hope of many researchers and biotech companies is that AI will unlock better tools to predict drug safety efficacy before moving to human testing.
The Problem with a Singular Focus on AI-Powered Preclinical Testing
There is some promising early evidence that computational models of human physiology or human organoid platforms might better predict drug effects in real humans4,5. However, it’s not clear that focusing on the preclinical stage alone will make it substantially easier and cheaper to bring a new drug to market. Although it’s probably too early to say, the use of AI in drug discovery so far hasn’t obviously improved success rates for drug approval6. Undoubtedly, this is in large part due to the challenge of recapitulating the incredible complexity and inter-personal heterogeneity of human physiology in silico or in vitro. It’s certainly possible that an accurate simulation of human biology may be possible one day. But we are far from the level of understanding needed to accurately model the complex interactions between molecules, cells, and organ systems (we’ve “discovered” the existence of entire ligaments in the knee as recently as 20137). Organoid models can be useful approximations of complex tissue types, but short of growing entire humans in a dish, they won’t ever capture whole-body dynamics.
Real Humans – A Better Focus for AI?
One often overlooked approach to improving the efficiency of clinical programs is to point AI models at data collected from human participants during drug trials. By predicting safety and efficacy outcomes much earlier than traditional endpoints would, this approach could triage assets much more quickly at lower cost by reducing the length and participant size of trials. Beyond the power of AI models themselves, other technical advancements make this approach more feasible than ever. The proliferation of consumer wearables and low-power IoT sensors makes collecting longitudinal data convenient and cost-effective, lowering the bar for collecting the kinds of large datasets needed to train powerful models. To its credit, the FDA maintains a qualification program for new biomarkers or surrogate endpoints that can be used to accelerate trials9. Qualified biomarkers currently range from questionnaires for Major Depressive Disorder to molecular assays for renal injury. However, to date, the database lists only 13 fully qualified biomarkers.
Accelerating drug development with wearable sensors
Continuous or semi-continuous monitoring of trial participants could be the missing puzzle piece in AI-powered approaches to improving clinical success rates. For illustrative purposes, it helps to think beyond the wrist-worn consumer devices we know today. Instead, imagine a future where detailed physiological data can be collected from a bandaid-like sensor placed on the skin above an organ of interest (a future that we happen to be building here at Skribe Medical). Collecting data on an hourly basis from a network of these sensors across the body surface could capture a longitudinal picture of health or disease for each trial participant. Combining this rich dataset with AI models could improve the efficiency of clinical trials in multiple ways:
- Pre-trial enrichment or segmentation of participants into functional subgroups. Before a trial even begins, collecting a few days of baseline physiology from trial participants could be used in conjunction with existing datasets to identify patients with different risk profiles or potential to benefit from treatment. After trial completion, a post-hoc analysis can be used to identify patient characteristics that predicted success. In some cases, biomarkers or characteristics that identify subgroups of patients for whom the treatment works well could rescue a drug that may otherwise fail when measuring efficacy across a more diverse population.
- Identifying safety issues as early as possible. Currently, significant cardiotoxicity is identified in 9% of Phase I trials compared to 36% of Phase II/III trials, and accounts for 45% of postmarket drug withdrawals10. In an ideal world, all of these issues would be detected in Phase I, before hundreds of millions are spent on larger and more elaborate trials. Identifying early subclinical signs that predict future toxicity would be a powerful way to maximize the value of Phase I trials and de-risk later phases of development.
- Development of new biomarkers indicative of long-term efficacy. Determining patient outcomes for many diseases can require many months or even years of observation. Identifying prognostic markers that predict long-term outcomes can significantly accelerate drug approvals. Capturing longitudinal multimodal data from trial participants can be a powerful way to identify trends that predict future success.
- Collecting data to support indication expansion. Identifying new indications for approved drugs is a significant way to reach more patients and expand revenue. Convenient, low-profile wearable sensors worn for the duration of clinical trials can generate data that can be easily mined to identify promising and even surprising use cases for a new drug. Collecting this data early can de-risk future trials and provide early evidence to secure regulatory approval.
A Human-First Approach
Improving preclinical testing with new AI-driven approaches may yield better success rates in the clinic. However, decades of drug development history have shown that predicting drug effects in humans is notoriously difficult. The convergence of ubiquitous wearable devices combined with low-power sensors and data-hungry AI models suggests that a focus on collecting high-quality human data during trials may be a shorter path to better clinical success. Building lightweight, wearable sensors that facilitate compliance and maximize data quality can lay the foundation for powerful predictive models that identify safety and efficacy early in development – saving time and cost while bringing effective treatments to patients sooner.
References
- Chabner, B., Roberts, T. Chemotherapy and the war on cancer. Nat Rev Cancer 5, 65–72 (2005). https://doi.org/10.1038/nrc1529
- DiMasi, J.A. (2001), Risks in new drug development: Approval success rates for investigational drugs. Clinical Pharmacology & Therapeutics, 69: 297-307. https://doi.org/10.1067/mcp.2001.115446
- Brogi, S., Ramalho, T.C., Kuca, K., Medina-Franco, J.L. and Valko, M., 2020. In silico methods for drug design and discovery. Frontiers in chemistry, 8, p.612.
- Matsui, T. and Shinozawa, T., 2021. Human organoids for predictive toxicology research and drug development. Frontiers in Genetics, 12, p.767621.
- Claes, S., Vereecke, E., Maes, M., Victor, J., Verdonk, P. and Bellemans, J., 2013. Anatomy of the anterolateral ligament of the knee. Journal of anatomy, 223(4), pp.321-328.
- Laverty, H.G., Benson, C., Cartwright, E.J., Cross, M.J., Garland, C., Hammond, T., Holloway, C., McMahon, N., Milligan, J., Park, B.K. and Pirmohamed, M., 2011. How can we improve our understanding of cardiovascular safety liabilities to develop safer medicines?. British journal of pharmacology, 163(4), pp.675-693.
Leave a Reply