How AI Is Breaking Down the Bottlenecks in Small Molecule Drug Discovery

Clinical trials for drug development are notoriously slow and expensive, and only a small fraction of drug candidates end up being approved by the regulatory authorities. The extensive bottlenecks in the traditional drug discovery process are all too familiar to those in the pharmaceutical industry: The clinical phase alone takes about a decade and accounts for nearly three-quarters of R&D costs, with the average cost of developing a drug rising to a staggering $2.2 billion in 2024.

Perhaps most critically, the pharmaceutical industry has struggled with accurately predicting successful development outcomes during early-phase work: About 80-90 percent of drug candidates fail to get approved despite the many years and extensive funding put into their development. Drug safety is a big reason for that, with unanticipated toxicity accounting for about 30 percent of drug development failures.

AI-discovered molecules are more successful in Phase I trials

However, emerging data suggests that artificial intelligence and computational approaches are not just addressing these challenges – they’re fundamentally transforming our ability to predict successful outcomes. The early stages of novel drug development are particularly crucial – and this is where AI and computational chemistry can make the most significant impact.

These AI and computational approaches can help find effective novel therapies to target the right proteins to treat disease in the early discovery and optimization phase, rather than later in the process, which has long been the case. Computational toxicity prediction is particularly helpful in the early stages of drug discovery, because it can exclude molecules that are likely to fail in clinical trials. Finding the best drugs earlier – and ruling out those that will not work – can save a decade of expensive research and increase the chance that the drug will pass the clinical trial phase.

Recent research on AI-native biotech companies reveals encouraging trends that suggest computational approaches are indeed beginning to overcome some of the most fundamental challenges in drug discovery. For example, an analysis published in Drug Discovery Today found that AI-discovered molecules are substantially more successful than historic industry averages in Phase I trials – reaching an 80-90 percent success rate in Phase I, compared with a 40-65 percent industry average.

This early-stage success is particularly significant because it indicates that AI-driven approaches are solving one of the most fundamental challenges in drug discovery: designing molecules that possess the multiple characteristics necessary for pharmaceutical development.

The power of multi-parameter optimization

At the heart of this transformation lies a skill that AI-powered computational chemistry is much better and faster at than humans: multi-parameter optimization, the process of balancing multiple properties of a potential drug at the same time – such as potency, safety, specificity, blood-brain barrier permeability and many others. This makes it much more accurate, faster and more efficient to design the most promising candidates, even if those properties conflict with each other.

Traditional approaches to drug discovery can optimize only one parameter at a time, making it difficult to improve one aspect without adversely affecting others. For example, a drug meant to treat brain tumors needs to be able to penetrate the blood-brain barrier so it can reach the brain. But a drug that crosses the barrier efficiently might not be sufficiently selective about its target, which could reduce the drug’s effectiveness or cause unwanted side effects. The traditional approaches might optimize for issues like blood-brain barrier permeability first and address other properties later, potentially pushing problems down the road rather than addressing them at the beginning of the process.

Meanwhile, AI-enhanced computational tools fundamentally change the approach to drug design. Rather than sequential optimization that can lead to later-stage failures, AI enables simultaneous optimization across all critical parameters already in the discovery stage. With AI, researchers can feed in data about multiple constraints and ask the algorithms to find a known molecule that works best with all those constraints – or generate a novel one. Using rapidly advancing generative AI and machine learning tools to develop optimal drug candidates faster and more accurately, by simultaneously analyzing multiple parameters, increases the likelihood of success and is ultimately expected to lead to the development of more effective, reliable and safe treatments for patients.

AI-based computational tools can also learn the unique requirements for different therapeutic areas. AI algorithms can incorporate these nuanced requirements to generate drug candidates that are tailored for specific diseases and target organs rather than simply meeting general criteria for being drug-like molecules. For example, a compound that targets brain tumors faces different optimization challenges than one designed for the chronic inflammation associated with arthritis, diabetes, atherosclerosis and other diseases.

To truly empower these AI-driven approaches, ultra-large datasets of molecules are required, both to screen and, more importantly, to train these models. The larger the dataset, the more chemical space is covered, which also increases the chances of success. Instead of screening tens of thousands (or even a few million) molecules and moving a few dozen into development, computational researchers can screen as many as tens of billions of molecules.

The next step: Integrating AI approaches into development pipelines

With the increased complexity of these approaches, a major challenge is the ability to perform them at scale. Therefore, the next step for leveraging AI in drug discovery is to incorporate tools that allow the discovery process to be scaled through the use of AI agents – autonomous computational systems that can perform complex tasks or processes without constant human intervention.

For example, the agents can be used to collect and analyze the necessary and constantly growing amount of information and rule out the less relevant drug candidates.

Once the agents have been trained on many parameters, chemical constraints and other relevant variables such as toxicity levels and FDA requirements, they will ultimately be able to provide researchers with the leading molecular candidates for any disease.

The pharmaceutical industry’s challenge now is not whether to adopt AI-driven computational drug design, but how quickly and effectively it can be integrated into existing development pipelines. While challenges remain, the early evidence suggests that AI and computational chemistry hold the key to better medicines developed more efficiently and reaching more patients faster than ever.