AI-powered Drug Discovery lecture by Dr. Michael Levitt, 2013 Nobel Laureate in Chemistry

Michael Levitt explains how decades of computational biology—from early protein‑folding simulations to AlphaFold and AI‑native startups—are converging to make fully AI‑powered, end‑to‑end drug discovery realistically achievable and dramatically faster.



Secret of life and protein folding

He frames the “secret of life” as learning plus self‑assembly: DNA encodes information, proteins self‑assemble into precise 3D structures, and function emerges from physical interactions such as drug binding. Proteins are presented as long amino‑acid chains that spontaneously fold into compact, highly specific shapes, making protein folding like a 3D jigsaw puzzle that biology solves without any external assembler—a key inspiration for future manufacturing.



Early computational modeling and the role of computing power

Levitt reviews his early work from the 1970s on multi‑scale modeling and simplified bead‑chain simulations that could fold small proteins using energy minimization and normal‑mode “thermalization,” albeit with limited accuracy. He emphasizes that raw computing power has since increased by about

10

9

10

9

, and argues that this massive speed‑up, more than brand‑new algorithms, is what made modern AI and large‑scale simulations so pervasive.



Modern AI for protein structure prediction (OpusX and AlphaFold)

He highlights OpusX, a neural‑network approach that predicts backbone torsion angles using many features to produce accurate folds for larger proteins, illustrating how small academic teams can now do powerful AI‑based structure prediction. He then discusses DeepMind’s AlphaFold: a large, mostly AI‑expert team that entered CASP, outperformed others especially on the hardest targets, and used convolutional, recurrent, graph, and attention networks while conceptually treating proteins as rigid pieces in a jigsaw—built on top of 60 years of structural biology and hundreds of thousands of known structures.



From structure to AI‑driven drug pipelines (Insilico Medicine)

Levitt stresses that protein structure prediction is like playing Go—scientifically impressive but only indirectly helpful—whereas the real impact comes from using structures to design better drugs. As an example, he presents Insilico Medicine, which applies AI across the whole pipeline—target discovery, chemistry, virtual screening, toxicity, and clinical trial outcome prediction—aiming to cut costs by orders of magnitude and shrink timelines from many years and hundreds of millions of dollars to far faster, cheaper development.



Aging, targets, and optimism for AI‑powered biomedicine

He describes one clever Insilico strategy: using differences between young and old people’s molecular profiles (e.g., epigenetics, expression) to infer disease‑relevant pathways and targets, analogous to how old cars accumulate more faults than new ones. He concludes that AI is uniquely good at integrating huge, noisy datasets across the entire drug‑discovery chain, turning uncertainty into workable options, and that with such end‑to‑end AI‑driven platforms already reaching human micro‑dosing trials in fibrosis and other indications, the future of AI‑powered biomedicine look