The advent of generative AI signifies a significant shift in the field of artificial intelligence, akin to how AlphaGo and AlphaFold revolutionized AI research. As highlighted in the ThinkCast podcast episode, “When Not to Use Generative AI” (https://podcasts.apple.com/kr/podcast/gartner-thinkcast/id1144653856?i=1000658557982), the strengths and limitations of generative AI are explored. In summary, generative AI excels in content generation, conversational interfaces, and knowledge discovery, while it faces challenges in planning and optimization, forecasting, decision engineering, and autonomous systems.

In the context of AI drug discovery, generative AI holds immense potential by assisting medicinal chemists in generating diverse hypotheses. This capability aligns with the core objective of medicinal chemistry—to understand and manipulate molecular structures to achieve desired properties. While the relationship between molecular structure and properties is well-established, its intricate nature and dependence on numerous factors make data interpretation and predictive modeling a formidable task. This challenge is further amplified in drug discovery settings where novel structures with limited data are often the focus. Generative AI offers a transformative approach by enabling the visualization of novel structures and their predicted properties, sparking innovative ideas that might otherwise remain elusive.

Despite its strengths, the podcast aptly points out the limitations of generative AI in optimization and decision-making, particularly in science-driven research. Relying solely on generative AI for such tasks would be counterproductive. The mention of autonomous systems in the podcast underscores the inherent challenges in trusting generative AI as a core component of large-scale automation systems. The stochastic nature of random number-based tasks, coupled with the limitations of generative AI in guaranteeing predictive accuracy, necessitates a cautious approach. This aligns with my advocacy for human-in-the-loop AI, where the overall system architecture involving task-level AI models is carefully analyzed, evaluated, and adjusted by multiple experts.

Drug discovery demands exploration into unexplored chemical (and biological) space and collaboration among experts from diverse disciplines. This necessitates a well-defined strategy for integrating generative AI with other established techniques, such as theory-based simulations and mathematical modeling. The scarcity of data sufficient to train AI models poses a unique challenge in drug discovery, where the very nature of the work involves a continuous exploration of novel molecules. This data scarcity further underscores the importance of human-in-the-loop AI, as human expertise remains crucial in guiding the AI model throughout the drug discovery process.

The assertion that AI will not replace experts but rather empower AI-savvy experts to surpass their non-AI-equipped counterparts holds profound implications for medicinal chemists. Embracing human-in-the-loop AI systems should be considered an essential component of career development for medicinal chemists moving forward.