Progress, Pitfalls, and Impact of AI‐Driven Clinical Trials

Abstract

Since the deep learning revolution of the early 2010s, significant efforts and billions of dollars have been invested in applying artificial intelligence (AI) to drug discovery and development (AIDD). However, despite high expectations, few AI‐discovered or AI‐designed drugs have entered human clinical trials, and none have achieved clinical approval. In this perspective, we examine the challenges impeding progress and highlight opportunities to harness AI's potential in transforming drug discovery and development.

BACKGROUND

The integration of computational techniques into drug discovery dates back to the early development of computational devices. The deep learning revolution of 2013–2014, followed by advancements in generative AI in 2017, catalyzed a surge of companies leveraging AI across various domains of drug discovery. These domains include disease modeling, target identification, de novo design of small molecules and biologics, preclinical and clinical pharmacology, clinical trial optimization, and the analysis of real‐world evidence for personalized medicine. ¹ Despite widespread adoption of AI by biotechnology and pharmaceutical companies, the overall number of drug approvals has increased only marginally. High expectations and substantial investments have yet to translate into a significant number of AI‐discovered therapeutics reaching clinical trials, let alone achieving clinical approval. ²

In the early days of deep learning, it was impossible to evaluate the effectiveness of specific approaches startups employed in drug discovery as no time lines were available. Schuhmacher et al. ³ of the St Gallen consortium developed a set of benchmarks for comparing startups and big pharmaceutical companies. In 2024, a decade since the deep learning revolution, with multiple AIDD companies reaching clinical trials, it is possible for AIDD productivity benchmarks to be developed. Several recent industry reports reveal the heterogeneous nature of the biopharmaceutical sector and the diverse applications of AI within it. According to a BiopharmaTrend report published in April 2024, eight leading AI drug discovery companies had 31 drugs in human clinical trials: 17 in Phase I (including one terminated), five in Phase I/II (including one discontinued), and nine in Phase II/III (including one with non‐significant results). ² Other reports, employing broader definitions of AI drug discovery companies, estimated 67 molecules in clinical trials, with one repurposed generic molecule launched. ⁴ These discrepancies underscore the challenges in assessing the true impact of AI in drug discovery.

MODELS OF AI‐DRIVEN DRUG DISCOVERY COMPANIES

Companies utilizing AI in drug discovery generally follow one of three models:

1. Repurposing or In‐Licensing Based on AI‐Derived Hypotheses: These companies employ AI to generate disease and target hypotheses, then in‐license known drugs, repurpose generics, or even utilize common reagents. This approach allows them to bypass years of hit identification, optimization, and early trials. While it enables rapid initiation of Phase II studies, several programs have failed to demonstrate efficacy, leading to substantial declines in market value. This model carries high target choice risk but low chemistry risk.

2. Designing New Entities for Established Targets: Focusing on validated, high‐value targets, companies in this category aim to develop best‐in‐class, differentiated molecules using AI‐driven design. By avoiding the risks of target discovery and validation, they concentrate on achieving superior chemistry—a challenging endeavor due to competition from established players with significant resources. This model presents low target choice risk but high chemistry risk.

3. Designing Novel Molecules for Novel Targets: Utilizing integrated, end‐to‐end AI platforms, these companies select high‐novelty targets without clinical‐stage competitors and design first‐in‐class molecules. One such company completed a Phase IIa study demonstrating safety and efficacy. This model involves high target choice risk and moderate chemistry risk.

These models are illustrated in Figure 1 , which visualizes the therapeutic pipeline stages and how different AI‐driven models fit into the drug discovery and development process.

Figure 1.

Three models of advancing drug candidates to clinical trials commonly adopted by AIDD companies.

Despite thousands of AIDD companies emerging globally between 2014 and 2019, only a few have reached the clinical stage, and no novel AI‐discovered drugs have attained clinical approval. ⁵ Furthermore, from 2012 to 2024, platform partnerships between AIDD and pharmaceutical companies have not resulted in AI‐discovered targets or AI‐designed molecules reaching Phase II studies, casting doubt on the industry's promise to significantly accelerate drug discovery.

CHALLENGES AND OPPORTUNITIES IN AI‐POWERED DRUG DISCOVERY

The limited success of AI in advancing drugs to clinical approval highlights several challenges inherent in the biopharmaceutical industry. These challenges, often non‐technical, relate to the industry's process‐oriented nature and the absence of clear benchmarks for time, cost, and success probability at each stage. Table  1 summarizes these challenges and corresponding opportunities for AIDD companies.

Table 1.

Common challenges and opportunities in AI‐powered drug discovery and development

Challenges	Opportunities
Lack of Clear Industry Benchmarks: AIDD companies lack of clear industry standards and benchmarks to compare AIDD approaches. AIDD companies often focus on newsworthy proofs of concept instead of focusing on beating the time, cost, and success rates of traditional methods at the same or higher level of quality. There is a lack of qualified industry analysts who evaluate the novelty of the programs and approaches and track the benchmarks in AIDD companies or in big pharma	Establish Transparent Benchmarks: Tracking and publishing key metrics such as time and cost to discover and validate targets can set industry standards. Time, cost, novelty assessment, and preclinical parameters of every preclinical candidate (PCC) should be disclosed alongside the total numbers. Time and success rates at every stage that follows PCC should be published or announced. AIDD companies should aim to exceed these benchmarks, demonstrating clear contributions of AI in delivering better drugs efficiently
Focus on Single Capability: Many companies concentrate on either biology or chemistry without integrating the multitude of steps required for high‐quality clinical programs	Develop End‐to‐End Capabilities: Understanding the entire drug discovery process and integrating AI capabilities with traditional approaches can produce superior therapeutic assets. Comprehensive platforms can address complex challenges more effectively than single‐focus solutions
Imbalance Between AI and Traditional Methods: Companies led by expert drug hunters do not value or trust AI and require pilots and proofs of concept instead of immediate deployment. Companies led by AI experts often underestimate the standards of high‐quality therapeutic assets and the level of validation required	Frontier AI Integration With Traditional Methods: AIDD companies must understand that their purpose is to speed up, cost‐down, and increase the number of high‐quality therapeutic assets while adhering to the highest quality standards. Splitting AI teams into frontier research and applied research can integrate innovative tools into established processes led by experienced drug hunters and drug developers
Reliance on Partnerships with Low Success Rates: Historically, platform collaborations between AI companies and pharmaceutical firms have low probabilities of success	Build Internal Expertise: Acquiring robust drug discovery and development capabilities by developing internal assets before partnering increases the likelihood of success. Starting collaborations from validated, high‐quality therapeutic assets and platforms can enhance confidence and outcomes
Underestimating Biological Complexity: Pursuing complex, unsolved problems without outperforming existing benchmarks can lead to unrealistic expectations and failures	Balance Novelty with Feasibility: Focusing on areas with established workflows where AI can outperform human intelligence increases success probability. Once AI proves effective in these domains, companies can explore novel hypotheses with greater confidence
Insufficient AI Validation: AI models often lack extensive validation in active drug discovery programs, leading to poor performance when applied beyond proofs of concept	Validate Extensively: Testing AI platforms in real‐world programs and integrating them into production only after thorough validation ensures reliability. Continuous feedback from multi‐year studies can refine models and enhance their applicability
Disconnected Public Datasets: Massive amounts of publicly funded data remain unintegrated, limiting their utility in AI models	Collaborate on Data Integration: Governments, companies, and technology providers can develop unified data lakes, maximizing the value of public data and accelerating drug discovery efforts that benefit global health
Regulatory Obstacles: Anticipated regulations and public concerns can deter companies from adopting frontier technologies, hindering innovation	Maximize Transparently and Publish: Publishing interim and final results at every stage and demonstrating safety can build public trust and inform regulatory frameworks that support innovation while ensuring patient safety
Geopolitical Challenges: Decoupling and deglobalization impede cross‐border collaborations, data sharing, and investments crucial for AI and biotechnology advancements	Demonstrate Global Benefits: AIDD companies should rapidly execute on the promise of AI and deliver effective affordable medicines to demonstrate the value of global collaborations. Demonstration of significant public benefit and significant improvement in efficacy and cost of innovative therapeutics may help defuse the geopolitical tensions
Inefficient Funding Allocation: Large investments by technology giants into platform capabilities often overlook the necessity of experimental validation, resulting in AI initiatives that lack real‐world applicability and trust of the pharmaceutical industry	Prioritize Experimental and Clinical Validation: Technology providers should invest in or collaborate with companies that have demonstrated experimental success, ensuring that AI initiatives are grounded in practical effectiveness and trusted by pharmaceutical partners
Mismatch in Development Paces: AI models evolve rapidly, but drug validation processes are lengthy, leading to potential obsolescence of models by the time drugs are validated	Synchronize Efforts: Dividing AI teams into frontier research and applied research can balance innovation with practicality, allowing new models to be tested while validated methods continue to drive drug development forward

LIMITED MODEL AND PROGRAM BENCHMARKING

There are multiple model‐specific benchmarks that provide the ability to compare different generative models, particularly in the chemistry space, ⁶ as well as summaries of experimental validation studies. For example, a consortium of scientists from Cornell, Cambridge, and EPFL universities recently published a comprehensive overview of molecular generation models with experimental validation. ⁷ However, no benchmark for real therapeutic programs out of generative AI exists. Pharmaceutical and biotechnology companies do not publish the time, cost, and success rates of preclinical and clinical experiments. One recently published example presented the timeline for a therapeutic program targeting Idiopathic Pulmonary Fibrosis (IPF), a complex disease with no clear genetic origin or existing standard of care; the company aimed to discover a novel target and design small molecules using AI. The selected target was highly novel, with no known selective inhibitors or drugs in clinical trials. It was also absolutely novel for the indication and was prioritized using multiple AI methods as well as the hallmarks of aging assessment. The process from program initiation to preclinical candidate nomination took 18 months. ⁸ The program recently completed a Phase IIa study, demonstrating safety, tolerability, and unexpected dose‐dependent efficacy measured by increases in forced vital capacity in idiopathic pulmonary fibrosis. Another published example disclosed 12 months from initiation to completion. ⁹ Non‐peer‐reviewed time line announcements claimed records of 9 months. Publishing time lines, costs, success stories, and failures is important for the industry to create a set of program‐level benchmarks and increase the transparency of the industry. Regardless of what technology is used in drug discovery and development, the ultimate goal is to make the process faster, cheaper, and increase the probability of success.

CHALLENGES AND OPPORTUNITIES FOR AI IN CLINICAL DEVELOPMENT

While AI has markedly accelerated early‐stage drug discovery, progress in drug development has been incremental. Common applications of AI in development involve automating routine tasks such as medical writing, regulatory submissions, and clinical data analysis. These applications improve efficiency and reduce costs but are unlikely to significantly increase success probabilities or approval rates. Automation alone does not equate to scientific advancement.

Drug development remains a highly regulated domain with processes refined over decades, leaving limited room for disruptive innovation. However, AI offers opportunities in predicting and optimizing clinical trial outcomes. Advanced AI systems can integrate preclinical knowledge with historical clinical data, facilitating end‐to‐end reasoning across the drug discovery and development continuum. ¹⁰ By predicting success probabilities at various trial stages, identifying optimal biomarkers, and suggesting combination therapies, AI can enhance decision making and efficiency in clinical development.

Moreover, AI Life Models, AI systems trained on longitudinal, multi‐species data can provide insights into complex biological processes, potentially identifying novel therapeutic avenues. These capabilities represent significant opportunities to bridge the gap between discovery and development, ultimately improving patient outcomes.

CONCLUSION

The promise of AI in revolutionizing drug discovery and development remains largely unfulfilled due to a combination of technical, operational, and strategic challenges. For the AIDD industry to realize its potential, it must demonstrate clear, reproducible successes where AI significantly outperforms traditional methods in delivering safe and effective therapeutics. Establishing clear industry benchmarks starting from time, cost, novelty, and preclinical parameters of each preclinical candidate (PCC) nomination, fostering collaboration between AI experts and traditional drug developers, and engaging transparently with regulators are crucial steps toward this goal. As the industry evolves, accumulating a robust body of evidence through successful AI‐discovered and AI‐designed drug approvals will be essential. Such progress may provide tangible evidence needed by the regulators to pave the way for new regulatory frameworks that further accelerate the translation of AI‐driven discoveries into real patient benefits.

CONFLICT OF INTEREST

Alex Zhavoronkov, PhD is affiliated with Insilico Medicine: Insilico Medicine is a global clinical‐stage commercial generative AI company with several hundred patents and patent applications and commercially available software. All other authors declared no competing interests for this work.