Artificial intelligence is shifting from accelerating individual steps in drug discovery to predicting which drug candidates will actually succeed in human trials, a capability that could reshape how pharmaceutical companies invest billions in development. Rather than simply generating new molecules faster, the latest AI systems are tackling the harder problem: determining which of those molecules will be safe and effective when tested in patients. This represents a fundamental change in how the industry approaches one of its most expensive and uncertain challenges.

Why Is Predicting Clinical Success So Much Harder Than Finding Drug Candidates?

The pharmaceutical industry faces a brutal statistic: only about 1 in 10 drug candidates that enter clinical trials actually make it to market approval. This means companies invest enormous resources developing molecules that ultimately fail, wasting time and capital. While AI has become excellent at generating new compounds and predicting protein structures, the real bottleneck has shifted. As one expert noted, "the field has become increasingly good at generating molecules, but molecule generation is no longer the primary bottleneck. The hard question is which molecules are going to work safely and effectively in humans".

The challenge lies in understanding causation rather than just spotting patterns. AI systems excel at recognizing correlations in massive datasets, but drug development requires understanding why a molecule will work, not just that it might. A molecule must navigate multiple biological systems, avoid toxicity, reach the right targets in the body, and produce therapeutic effects in diverse patient populations. Each of these factors involves complex, interconnected biological mechanisms that simple pattern recognition cannot fully capture.

How Are Companies Using AI to Predict Which Drugs Will Succeed?

VeriSIM Life, a company founded in 2017, has developed what it calls a "credit-score-like" risk assessment system for investigative drugs. The platform, called BIOiSIM, combines virtual animal-to-human drug simulations with machine learning to predict clinical trial success. In validated studies, the system achieved nearly 90% accuracy in predicting whether drugs would succeed, compared with the historical 10% success rate. The company now has about 20 active partnerships with pharmaceutical companies and is developing its own drug candidates to validate the approach.

"Predictions are made using hybrid AI models, which combine virtual animal-to-human drug simulations with machine learning. The system models how a drug behaves and explains the underlying biological and physical reasons for that behavior," explained Jo Varshney, founder and CEO of VeriSIM Life.

Jo Varshney, DVM, Ph.D., Founder and CEO of VeriSIM Life

The platform works by assigning each drug candidate a dynamic "credit score" that changes based on modifications to its chemical structure, dosage, or targeted disease. Rather than relying on a single prediction method, BIOiSIM blends multiple AI algorithms with mechanistic models that capture biological principles. This hybrid approach helps the system explain not just what will happen, but why, which is critical for building trust with pharmaceutical companies making billion-dollar decisions.

One recent breakthrough involved an AI system called Robin, which was published in Nature in May 2026. Robin demonstrated the first fully autonomous discovery and validation process by analyzing 551 scientific papers in about 30 minutes, a task that would take a human researcher roughly 540 hours of manual work. The system identified ripasudil, a therapy used for glaucoma, as a candidate for repurposing to treat dry age-related macular degeneration, the leading cause of blindness in developed countries. Researchers then confirmed the finding in laboratory experiments.

What Are the Real-World Applications of These AI Systems?

Pharmaceutical companies are already using these AI platforms to accelerate their development timelines. VeriSIM Life has enabled several client programs to enter clinical trials significantly ahead of traditional schedules, and one of its own drug candidates for pulmonary arterial hypertension is less than a year away from entering human trials. A study by the Boston Consulting Group examined 15 AI-assisted drug candidates that advanced to clinical trials and found that 5 of them did so in under four years, compared with the historical average of five to six years.

Beyond speed, these systems address specific bottlenecks in drug development. Pharmaceutical sponsors typically ask AI systems three critical questions: which candidate molecules will work in patients, what adverse effects can be anticipated and reduced during toxicology studies, and how to identify the right patient population for clinical trials to achieve statistical significance. Traditional tools can address one or two of these questions, but they cannot predict overall outcomes because biology is fundamentally interconnected. A molecule's chemistry, how it behaves in animal models, how it metabolizes in human bodies, and how different patient populations respond all influence whether a drug will succeed.

How to Leverage AI Predictions in Drug Development Strategy

• Early Candidate Ranking: Use AI systems to rank multiple drug candidates by predicted success probability before investing heavily in preclinical studies, allowing companies to focus resources on the most promising molecules.
• Toxicity Prediction and Reduction: Deploy AI-driven digital twins of animals and humans to simulate how molecules will behave in the body, potentially reducing or replacing traditional six to nine-month animal toxicity studies.
• Patient Population Identification: Apply AI analysis to identify which patient subgroups are most likely to respond to a drug candidate, improving clinical trial design and statistical power.
• Indication Selection: When a molecule shows promise for multiple diseases, use AI to rank which disease indications offer the best chance of success based on biological mechanisms and patient variability.
• Iterative Refinement: Combine AI predictions with laboratory experiments in a feedback loop, using experimental results to refine the AI model and guide the next round of candidate modifications.

What Challenges Remain for AI-Driven Drug Discovery?

Despite impressive progress, AI systems still face limitations. Robin, for example, scored 47.9% on biostatistics tasks but only 15.3% on bioinformatics tasks, a gap the researchers attributed to bioinformatics requiring multi-step mechanistic reasoning rather than single-step statistical computation. This gap matters because understanding the causal pathway from a drug target to a therapeutic effect is critical for developing effective treatments. AI systems that excel at pattern recognition may struggle when they need to reason through complex biological mechanisms step by step.

The availability of data also constrains these systems. VeriSIM Life addresses this by creating synthetic data through mathematical representations of human and animal systems, allowing the platform to simulate drug interactions even when experimental data is limited. The company has tested its system across 72 disease areas and thousands of different targets, but the quality and breadth of available data still influence prediction confidence.

What Policy Changes Could Accelerate AI-Driven Drug Discovery?

Realizing the full potential of AI systems like Robin and BIOiSIM will require support across several policy areas. First, shared data infrastructure needs sustained public investment. Federal initiatives including the National Institutes of Health's Bridge2AI program and the All of Us Research Program, alongside public-private partnerships such as Open Targets, represent productive frameworks that merit continued support. Privacy-enhancing technologies could expand access to datasets without compromising individual privacy.

Second, a risk-based regulatory framework for AI in drug discovery is essential. Discovery-stage systems like Robin operate upstream of patient care; any candidates they identify must still pass rigorous clinical trials before reaching patients. Policies that clarify and simplify the FDA review process for such tools could encourage wider adoption. The FDA is already transitioning away from animal testing by prioritizing new approach methodologies, including AI and computational modeling, creating an opportunity for these systems to reduce animal testing while improving drug development efficiency.

Third, commercial incentives matter. Drug-pricing policies that reduce expected returns can dampen investment in early-stage, high-uncertainty research that AI systems are designed to accelerate. AI-driven efficiency gains offer a complementary path to addressing rising drug development costs without compressing the returns that motivate private investment. Finally, investment in automated laboratory infrastructure, including cloud labs, biofoundries, and robotic experimentation platforms, will be necessary to fully realize the potential of autonomous discovery systems.

The convergence of AI prediction, automated experimentation, and supportive policy could fundamentally reshape pharmaceutical innovation. Rather than relying on luck and brute-force screening, drug developers could make evidence-based decisions early in development, focusing resources on molecules with the highest predicted probability of success. For patients waiting for treatments to rare diseases and conditions, this shift could mean faster access to new therapies. For pharmaceutical companies, it could mean lower development costs and higher success rates, making investment in innovation more attractive even as drug prices face regulatory pressure.