EP 58

Do We Need Mavericks in AI for Drug Discovery

The podcast explores the role of mavericks in AI for drug discovery, discussing the challenges and opportunities in this field. With a 90% failure rate of drugs in clinical trials, the industry is looking for innovative solutions. The episode delves into the stories of mavericks who are trying to revolutionize AI drug discovery, including Alex Zavornikov, Demis Hassabis, and Aviv Rejev. The conversation also touches on the importance of open-source architectures and the need for a networked coalition to drive progress in this field.

AIdrug discoverymavericksopen-source architectureshuman biology

play_arrowListen to episode

Key Concepts

AI in Drug Discovery

AI can be used to design drugs, but human biology is the main challenge
AI models can optimize for proxy metrics, but this may not translate to real-world success
End-to-end AI approaches are being explored, where AI is used to choose the target and design the drug

Mavericks in AI for Drug Discovery

Mavericks like Alex Zavornikov and Demis Hassabis are trying to revolutionize AI drug discovery
Different approaches are being taken, including foundation models and open-source architectures
The future of drug discovery may belong to a networked coalition of mavericks

Open-Source Architectures

Open-source tools like RF diffusion are being used by commercial competitors
The goal of open-source architectures is to build a broad institute of AI drug discovery
Open-source tools can constrain and accelerate commercial actors, forcing them to compete on clinical outcomes

Human Biology and Drug Discovery

Human biology is the main challenge in drug discovery, not algorithmic limitations
Drugs often fail due to human biology, not because of design flaws
Understanding human biology is essential for success in drug discovery

Networked Coalition

The future of drug discovery may belong to a networked coalition of mavericks
Biological humility, closed-loop data, and cross-cultural fluency are essential for success
A networked coalition can combine different approaches and expertise to drive progress

Episode Summary

check_circle90% of drugs that enter human clinical trials fail before they get approved
check_circleThe hype around AI in drug discovery is deafening, but the reality is more complex
check_circleMavericks like Alex Zavornikov and Demis Hassabis are trying to revolutionize AI drug discovery
check_circleThe biggest drop-off in clinical trials happens in phase two, where two-thirds of programs fail
check_circleDrugs often fail due to human biology, not because of algorithmic limitations
check_circleGoodhart's law states that when a measure becomes a target, it ceases to be a good measure
check_circleAI models can optimize for proxy metrics, but this may not translate to real-world success
check_circleThe finish line in drug discovery is not just designing a drug, but understanding human biology
check_circleMavericks are taking different approaches, including end-to-end AI, foundation models, and open-source architectures
check_circleThe future of drug discovery may belong to a networked coalition of mavericks
check_circleBiological humility, closed-loop data, and cross-cultural fluency are essential for success
check_circleThe negative results curator is a crucial gap in the current landscape, as AI systems need to learn from failures

Full Transcript

Discussion

Join the discussion — sign in to leave a comment.