AI Drug Discovery: Closing the Data Loop for Better Hits

Eroom’s Law describes a brutal reality in pharma: the cost of developing new drugs has roughly doubled every nine years since the 1950s. We are looking at a cycle where bringing a single drug to market takes over a decade and costs up to $2.5 billion, all while facing a failure rate higher than 90%. This is why the industry is pivoting so hard toward an AI workflow to compress these timelines.

The real value of AI isn't just "speed"—it's about increasing the quality of candidates that actually make it to the clinical phase. If you can filter out the garbage early, you save billions in failed trials.

From Empirical Screening to Predictive Design

The biggest shift is happening in hit identification. Traditionally, this was a numbers game: screen millions of molecular entities against a protein target and hope something sticks. Now, we're seeing a move toward predictive design. Instead of physically testing a library, researchers use LLM agents and specialized models to design candidates from scratch.

This removes the physical ceiling on how many starting points a company can explore. AI can effectively kill off low-quality candidates before a single pipette is touched in the lab. However, there is a ceiling here too. AI still struggles to reliably predict kinetics or the "developability" of a compound. Every AI-generated lead still requires physical validation.

The Lab Bottleneck and the "Data Wall"

Here is the paradox: AI is great at generating hits, but our lab infrastructure wasn't built to characterize them at this scale. Traditional screening produced "binary" data—essentially a yes/no response on whether a compound bound to a target. AI-driven discovery demands high-fidelity, information-rich data to validate these complex, diverse candidates.

More concerning is the "data wall" many models are hitting. A lot of early AI drug discovery was built on public datasets. The problem?

• Homogeneity: Everyone is training on the same data, leading to the same conclusions and diminishing returns.
• Lack of Structure: Public data wasn't designed for machine learning; it lacks the rigorous labeling and diversity needed for high accuracy.
• Publication Bias: This is the silent killer. Scientific papers almost exclusively report successes. No one publishes their failures, but for an AI model to actually learn, it needs to know what doesn't work just as much as what does.

Building a Real-World Feedback Loop

To move past this, the industry needs a closed-loop system where lab results—including the failures—feed directly back into the model. This is essentially a deep dive into a new kind of R&D where the AI proposes a molecule, the lab tests it, and the raw, unbiased data is used to refine the next iteration of the model.

Moving from a "predict-then-test" linear flow to a continuous loop is the only way to actually break Eroom's Law. We need a complete guide to integrating lab automation with model training if we want to see the 90% failure rate actually drop.