The field has progressed through three distinct phases. The first wave focused on large-scale data generation, the second on applying early machine learning to model molecules, and the current third wave leverages modern foundation models (like transformers and diffusion models) for highly effective generative design.
A fundamental shift is occurring from predictive AI, which acts like a 'microscope' to understand existing biology (e.g., AlphaFold, Chai-1), to generative AI, which acts like 'Photoshop for molecules' to design novel therapeutics from scratch (e.g., Chai-2). This moves AI from a passive analytical tool to an active creative partner in discovery.
Diffusion models are particularly well-suited for biological design because they can sample a wide range of distinct hypotheses. Unlike older models that might produce a single, averaged-out prediction, diffusion models can generate a diverse 'brainstorm' of potential molecular structures, increasing the odds of finding a novel and effective solution.
The speaker challenges the industry's focus on simply increasing the probability of success to reverse Eroom's Law. He argues that AI's greatest potential is in enabling the pursuit of more ambitious, high-value drugs that can change the standard of care, even if those projects carry higher risk.
Keep pulling the thread on Joshua Meyer.