When AI predicts what works, but not why: Three leading scientists on AI's future in drug discovery

The application of AI in drug discovery poses a significant challenge: identifying promising molecules without a clear explanation of the AI's reasoning. Unlike minor AI errors, such as a chatbot recommending non-existent restaurants, the stakes in medical applications are much higher, demanding rigorous validation. Scientists are navigating how to trust predictions from systems like Google DeepMind's AlphaFold, which relies on decades of accumulated experimental knowledge to predict protein structures.

Nobel laureate Michael Levitt suggests that the "black box" nature of AI is not unique, as biology itself contains many mysteries. He notes that drug regulators, such as the U.S. Food and Drug Administration, prioritize evidence of a medicine's safety and effectiveness over a complete explanation of every underlying mechanism. For Levitt, the crucial step is to test AI predictions against reliable experimental data and ensure findings can be independently reproduced.

Daniela Rhodes, an emeritus scientist, supports this perspective, highlighting that medicine has often advanced through experimentation before full understanding, citing Chinese herbal medicine as an example. While AI tools, including machine learning with Nvidia GPUs, can accelerate research stages like identifying millions of particles for molecular structure determination, Rhodes emphasizes that AI's outputs, even from AlphaFold, are "really good prediction[s]" but "not the absolute truth."

This indicates that AI can significantly aid in identifying potential drug candidates and accelerating initial research. However, laboratory experiments and comprehensive clinical trials remain essential. Researchers must still establish how AI-identified molecules behave in different human bodies, especially given that many drugs have historically been tested mainly on men within a limited age range.