The Reality of AI in Drug Discovery: Hype vs. Clinical Impact

AI in drug discovery has generated significant hype, but evidence of its clinically relevant impact remains disappointingly limited. While many AI methods have been developed and benchmarked, the translation of these tools into successful clinical outcomes has not yet occurred. The field is currently in a state of "absence of evidence," meaning that while AI has not yet proven its value in the clinic, it has not been definitively proven that it cannot work.

The Gap Between Early Discovery and Clinical Success

Clinical success is measured by the impact on Phase II trial success rates, as early-stage discovery work represents a small fraction of the time and money invested in drug development. Historically, drug discovery has suffered from a lack of transparency regarding failures; successful projects are often attributed to insight and hard work, while failures are rarely discussed. This creates a survival bias where any project that succeeds with an AI component is credited to the AI, while failures are ignored.

The transition from a ligand (a molecule that binds to a target) to a drug (a molecule that works in humans) is a critical hurdle. AI systems must move beyond simply identifying compounds that work in isolated proteins or cells, and instead focus on the processes that translate into clinical success.

Data Quality and the 'Honey Trap' of Applied ML

A primary obstacle to AI's success in drug discovery is the nature of the data. Biological systems are characterized by confounding factors, conditionality, and epistemic opacity. Much of the existing data consists of piles of assays that are difficult to clean, categorize, and for the use in machine learning.

Industry practitioners have noted that applied ML has become a "honey trap" in academia and industry. Researchers often propose ML-guided approaches, gather existing data, and discover that there are too few true data points on the outputs of interest to build a reliable model. This often leads to a researchers pivoting to simulation work or high-throughput systems that are far removed from the original problem to maintain the appearance of using AI.

AI as a Productivity Tool vs. Novel Discovery

While AI has not yet shifted the needle on clinical outcomes, it has provided significant productivity gains for scientists. Structural biologists, for example, use AI tools like AlphaFold to create starting models for chimeric fusions, reducing hours of manual work to a quick prompt. Other uses include debugging software, writing scripts for large datasets, and use of LLMs for drafting experiments.

In this case, AI acts as a tool for efficiency rather than a tool for novel discovery. It is making the same tasks easier and faster, but it is not yet magically accomplishing new things that were previously impossible.

The Path Forward: From 'What Can Be Done' to 'What Should Be Done'

To move the field forward, the focus of AI in drug discovery must shift from modeling data that is readily available to generating the new, high-quality data required to actually improve clinical success rates. This requires a commitment to substantial data generation, which is expensive and expensive and time-consuming, and may be slower than the most optimistic press releases suggest.

Synthesis of Expert and Community Insights

Community discussion highlights several additional bottlenecks:

• Manufacturing and Scaling: There is a growing number of AI-assisted drug candidates, but no corresponding improvement in the capacity to manufacture and scale them.
• The Need for Automation: Some argue that the real win will come when AI agents can run automated loops in the real world via robotics, closing the loop between prediction and experimental validation.
• The Need for Moderation: Experts emphasize that the approach to new AI technologies should be moderate, focusing on the why of the technique rather than the using it because it is new and shiny.