What Impact Will Artificial Intelligence (AI) Have on Clinical Trials?

Timothy Hare

VP, Head of Data Science, 81qd

New York, NY

Artificial intelligence (AI) will have a direct and marked impact on the future of clinical trials by increased incorporation into the clinical trial process itself, as well as an indirect impact via improvements in the preclinical drug discovery process, including the target identification (TID), lead identification (LID), and lead optimization (LO) processes that supply the pool of preclinical candidates (PCCs) for clinical trials. 

The most discernible direct impact of AI on the evolution of clinical trials will likely be seen in recruitment, participation, and data collection, as well as the emergence of new clinical biomarkers (indicators of the severity or presence of some disease state) and surrogate end points (observed effects of treatment that may correlate with a clinical end point). 

The US Food and Drug Administration (FDA) now supports the use of AI to improve patient diagnosis and prognosis, identify early response to treatment, and develop novel imaging biomarkers that can be used to categorize and triage patients. Integrated understanding across a broad range of complex patient data types (imaging, ECG, genomics, proteomics, human biome genetics) is increasingly facilitated by AI as part of biomedical investigations into human health and disease, all of which sets the stage for more efficient and effective clinical trials. 

AI is increasingly facilitating better real-time data capture for more desirable and demographically diverse clinical trial enrollment. Patients are increasingly comfortable with AI as part of the health care ecosystem, as seen in a recent survey that indicated 42% of respondents are comfortable with doctors using AI to make decisions about their care and treatment. Increasing integration of social media data, the Internet of Things data, laboratory data, and electronic medical records will enhance AI effectiveness in this regard. 

Additionally, there is the prospect of virtualization of parts of the clinical trial process, allowing patients to participate more easily and irrespective of their geographic region. This is helpful since approximately 70% of potential participants live more than two hours from a clinical trial site and traveling is considered a significant burden. Clinical trials will increasingly utilize social media, telemedicine, and biosensors as part of the clinical trial process, facilitating recruitment and simplified real-time data collection. 

In late-stage clinical trials, with the majority of recruitment sites having difficulty meeting minimum and demographically diverse enrollment levels, AI supports increased levels and diversity of patient recruitment by shortening the time to diagnosis for patients who have been misdiagnosed or are under-diagnosed – expanding the pool of candidate clinical trial patients. AI is able to predict the risk of non-adherence, so that clinical trial patient selection can be refined to reduce the risk of falling below enrollment levels needed to achieve statistically significant results. 

AI will improve clinical trial patient qualification and monitoring by being able to identify more prognostic surrogate end points and by discovering new biomarkers for disease type, state, and potential for progression to different outcomes. This will allow more nuanced and earlier detection of efficacy, as well as better constellations of assay readouts indicative of emergent toxicity. 

“Clinical trial designers are increasingly positioned to make use of all of this information, leading to better-informed strategic decisions about clinical trial structure and feasibility.” 

In addition to the marked improvements that AI will undoubtedly directly effect on the clinical trial landscape, there are a number of indirect impacts that will also contribute to reshaping research and development. This is an AI high-leverage point for clinical trial impact. Better preclinical decisions will result in safer and more efficacious PCCs, reducing the escalating costs of bringing new drugs to market, and allowing drug development in formerly refractory therapeutic areas. 

AI will improve our knowledge of TID networks (maps identifying biological pathways for a given therapeutic area), improving preclinical research and development’s (R&D’s) ability to focus efforts on more efficacious points of therapeutic intervention. AI will reduce error rates in LID that would otherwise inflate R&D costs and introduce unacceptably high late-stage clinical trial risks. 

AI will also improve the complex drug candidate structure-activity relationship (SAR) maps that medicinal chemists rely on for reducing toxicity, enhancing potency, increasing target specificity to avoid side effects, increasing oral bioavailability (the fraction of drug in circulation after oral dosing), and optimizing the half-life in circulation of drug candidates, as part of the synthesis of chemical variations on a theme during LO. LO is the stage where medicinal chemists begin borrowing information from different parts of the SAR map, across the full-range of active molecules coming out of LID, to produce improved PCCs. 

Drug discovery assays (laboratory biochemistry that attempts to mimic a particular therapeutic area of human biology in a meaningful way) often test millions of candidate molecules, all of which need to be assessed in a fully automated and cost-effective manner. The increasingly complex processes that support this goal produces a subset of false-positive (FP) and false-negative (FN) results that inflate follow-up costs in the first instance by sending too many candidates into expensive (materials, time, labor) secondary confirmatory assays, and also degrade the body of information (the SAR map) needed by medicinal chemists for effective LO in the second instance. This puts clinical trials at a severe risk of bringing suboptimal drugs to market and also an unacceptable risk of program failure, where the clinical trial fails to support bringing the new drug to market, and the program is subsequently cancelled. 

In this context, AI will enhance our understanding of biological networks and help identify the subset of truly viable drug targets they contain. This TID network knowledge map is essential to evaluate all PCC on-target effects (the effect on the intended target and known biological pathways), and also off-target effects (the effects on unintended targets and associated biological pathways, known and unknown), as the latter can lead to side effects that put clinical trials at risk of failure for lack of efficacy and/or toxicity. AI will help biopharmaceutical companies to realize the holy grail of “fail early” in the preclinical drug development pipeline. 

Failing early is good as it minimizes R&D investment in the wrong PCCs destined to fail late, otherwise inflating the cost of preclinical development and/or late-stage clinical trials. Late-stage clinical trial failure is of particular concern as it has a larger adverse impact on the average cost of bringing drugs to market, which has been steadily rising for decades. R&D costs need to be recovered. Without AI, the only option is to continue to increase the selling price of the drugs that get approved, in order to recoup the losses for all of the failed R&D programs. 

AI-based virtual screening, the process of searching for new drugs via computer simulation (eg, without actually running physical assays in the laboratory), can predict which candidate drugs are good leads, reducing LID error rates that otherwise inflate R&D costs, or increase the probability of late-stage clinical trial failure. Often physical assays are run and compared with AI predicted values, such that bad assay results (assays have artifacts, they are not perfect) that would have incorrectly eliminated a PCC that would have otherwise contributed information to the SAR map or even been a clinical trial success on its own (FN) are flagged and rescued during follow-up testing. Thus, reducing FNs reduces program risk. 

As well, those candidate drugs that are truly inactive and would otherwise have made it farther along in the process (FP) can also be eliminated. Allowing FPs to linger in the R&D development pipeline raises costs, increases development time, and contaminates the SAR map that medicinal chemists rely on to make improvements to PCCs. Thus, reducing FPs reduces program costs. 


Overall, this process of winnowing out FNs and FPs earlier improves the efficiency and effectiveness of the R&D process. 


A purely AI-driven design of drug candidates is coming into its own, as scientists discover better AI model inputs for drug candidate features (molecular “feature encodings” that AI consumes), as well as improved physical molecular library design. Molecular library design involves the purchase and/or chemical synthesis of a broad range of molecular structures. This inventory is then stored in very large automated warehouses for testing across all future drug programs. 

Strategies coming from AI recommendations, where AI designs truly novel drug candidates that medicinal chemists might not otherwise have created, represents a strategic and tactical paradigm shift. Participation of AI in this way will provide higher quality candidate drugs for validation, as well as expand diversity in the molecular libraries used for other therapeutic area opportunities. 

Preclinical LID assay development has been made more robust by AI, which is able to interpret a broad array of complex data. Large, complex mixtures of data such as candidate drug biophysical attributes (solubility, size, etc.) as well as protein and gene network expression data are increasingly fed into AI to provide a richer constellation of information to model. 

Also, live cell assays where the size, shape, and health of the cell are observed after exposure to candidate drugs, can be interpreted by AI. The capacity to quickly process this vast and diverse body of information means that AI is now able to contribute new information in areas like drug-target interaction theory, rational drug design, biological pathway analysis, and the planning that goes into new drug synthesis by medicinal chemists. 

AI is also being used to improve early warning of diminished LID assay quality by monitoring the inputs, the outputs, the robotics, and the influence of uncontrolled factors in the environment in which these complex automated processes run. The cost of complexity can be instability, and complex systems are prone to failure in ways that can’t be anticipated by humans, but can easily be predicted by AI. Round-the-clock system monitoring can dramatically reduce the cost of R&D attributable to materials waste, the challenge of scheduling the use of limited fixed resources such as shared robotic platforms, and the cost of rework when systems degrade and the product becomes unacceptable. 

With the advent of new FDA standards and guidelines, the direct impact of AI on clinical trials will be manifested by the increased diffusion of innovative therapies into the design and execution of clinical trials. In addition to the marked improvements that AI will undoubtedly directly effect on the clinical trial landscape, there are a number of indirect impacts that will also contribute to reshaping R&D. 

When you consider the fact that only about one out of eight PCCs will ultimately make it to market, the total R&D burden in modern biopharmaceutical drug discovery becomes clear. Studies estimate that the full life cycle development cost of bringing a single new drug to market may average almost $3bn. This is up from $802m in 2003 or approximately 275% in 15 years, unadjusted for inflation.

Studies also focus on industry productivity and profitability. Productivity and profitability in the biopharmaceutical industry have been declining as well, and some studies indicate that the industry may soon reach a crisis point if this trend is not reversed. A recent study indicates that R&D returns in drug development currently stand at 3.2% and could reach 0 in 2020, meaning that a dollar would return a dollar— ie, no profit. 

AI will be the catalyst to slow down and eventually reverse this trend in the biopharmaceutical industry. AI has the potential to dramatically improve clinical trial efficiency and effectiveness, directly and indirectly, ensuring that the volume of new drugs for unmet medical needs coming to market continues and even accelerates.

If you found this content valuable, please consider subscribing to Rx Data News to receive monthly updates on the state of data analytics and machine learning in the pharmaceutical industry.

© 2019 PNG Publishing

Follow us on twitter:

  • Black Twitter Icon