AI Was Supposed To Discover New Drugs. Where Are They?

Key Takeaways Copied to clipboard!

• Despite high expectations, no drugs currently on the market can be directly attributed to AI, indicating that the technology has not yet delivered on its promise to revolutionize drug discovery speed.  Copied to clipboard!
• AI is currently most effective at the initial stage of drug discovery—searching vast molecular libraries for candidates that geometrically fit a target protein—but it does not solve the subsequent, lengthy, and failure-prone stages like optimization and clinical trials.  Copied to clipboard!
• The future of AI in medicine may lie in creating 'digital twins' for in-silico clinical trials, which could replace unreliable animal models and account for human diversity, though current clinical trial data used for training is heavily biased toward Western populations.  Copied to clipboard!

Segments Expand All Collapse All

Traditional Drug Discovery Process

Copied to clipboard!

(00:02:58)

• Key Takeaway: The conventional drug discovery approach focuses on finding a single ‘magic bullet’ molecule that fits a target protein, often leading to late-stage failures due to unforeseen side effects or ineffectiveness in human trials.
• Summary: The traditional method aims for a one-size-fits-all drug, which is unlikely to work well because individuals react differently. This sequential process is expensive and time-consuming, often taking decades, with less than 10% of attempts reaching the market. Failures late in clinical trials result in the loss of billions of dollars and years of development.

AI’s Role in Drug Development

Copied to clipboard!

(00:05:35)

• Key Takeaway: AI is hoped to accelerate drug discovery by replacing lengthy, labor-intensive experimental efforts with reliable in-silico predictions, though its power is currently limited to the initial screening phase.
• Summary: The goal of using AI is to short-circuit development time by making reliable computer-based predictions before physical experiments begin. AI excels at quickly searching massive molecular libraries to find candidates with good geometric fits to target proteins. However, finding this initial ‘hit’ is only the very first step in a much longer process.

Failure of Early AI Drug Startups

Copied to clipboard!

(00:08:30)

• Key Takeaway: The collapse of companies like Benevolent AI demonstrates that overblown expectations, predicated on AI solving the entire workflow rather than just one part, led to massive valuation drops.
• Summary: The first wave of AI drug discovery claims around the mid-2010s failed to produce any market-ready drugs attributable to AI within the expected timeframe. This failure stemmed from the belief that AI was a singular solution, ignoring that no single part of the workflow can be managed exclusively by AI. Benevolent AI’s valuation plummeted after its initial flotation due to these unmet promises.

Generative AI Potential and Limitations

Copied to clipboard!

(00:10:04)

• Key Takeaway: Generative AI offers the potential to invent genuinely new molecules, but this innovation still requires subsequent optimization and rigorous testing through traditional clinical trial stages.
• Summary: Generative AI is a more sophisticated method that creates potentially novel molecules that have never existed before. This capability offers an opportunity to invent new ‘magic bullets.’ However, even these AI-generated candidates must still undergo multiple stages of optimization, laboratory testing, and clinical trials, areas where AI’s contribution is less clear.

Rethinking Clinical Trials with AI

Copied to clipboard!

(00:11:46)

• Key Takeaway: Replacing animal models with patient-specific data to build ‘digital twins’ for in-silico clinical trials is seen as a necessary evolution for more reliable drug testing.
• Summary: Experts agree that current mouse models are often not accurate predictors of human success, and human responses to drugs are highly diverse. Building reliable computer models of relevant human body parts allows for virtual clinical trials using patient-specific data, eliminating animal testing and reducing risk to human volunteers initially.

Data Bias and Global Health Equity

Copied to clipboard!

(00:14:48)

• Key Takeaway: The reliance on decades of Western clinical trial data creates significant bias, making predictions unreliable for diverse global populations, such as those in India.
• Summary: Current databases used to train AI models are dominated by data from white males, leading to inherent biases that affect reliability, especially for women. Clinicians in heterogeneous countries like India often rely on this Western data, which is a ‘flight of fantasy’ for predicting effective treatments for their local patient populations.

Future Trajectory and Medical Adoption

Copied to clipboard!

(00:16:25)

• Key Takeaway: The path forward involves AI contributing incrementally across the entire workflow, moving from short-term emergency response modeling (weather forecasting analogy) toward long-term wellness prediction (climate forecasting analogy).
• Summary: AI’s role is best seen as a contributor throughout the workflow, potentially shaving years off the decade-long process rather than reducing it to a couple of years instantly. While the medical establishment is slow to adopt this radical transformation, younger medical students are increasingly pushing for the integration of predictive software and coding skills.