🔬Generating Molecules, Not Just Models

This episode traces the remarkable journey from AlphaFold2’s landmark achievement in protein structure prediction to the broader landscape of molecular interaction modeling and protein design. The problem AlphaFold2 addressed—predicting the structure of single-chain proteins—was long considered intractable due to its perceived NP-hard nature. The breakthrough came not only from advances in machine learning but also from leveraging evolutionary data to infer co-evolution of amino acids, providing powerful hints about spatial proximity in protein structures. Yet, as the guests explain, the field quickly moved beyond this milestone toward more complex questions, like how proteins interact, how they fold dynamically, and how to model these interactions with small molecules, RNA, and DNA. AlphaFold3 marks a critical shift in this evolution, moving from static structure prediction to modeling heterogeneous molecular interactions. Rather than treating these interactions as isolated problems, AlphaFold3 unifies them within a single model trained across modalities. This progress also reflects a broader trend in machine learning: the shift from regression-style prediction to generative models capable of expressing uncertainty and capturing system dynamics. By sampling from a distribution of plausible structures and interactions, these models allow researchers to better understand the flexibility and variability of biological systems. However, such models also introduce new challenges, particularly around validation and ranking of generated outputs. Enter Boltz and its suite of tools, which aim to democratize access to these cutting-edge capabilities. Boltz builds on open-source principles and a strong community foundation to deliver models that are both state-of-the-art and accessible, with a focus on usability, extensibility, and real-world validation. Boltz2 and BoltzGen combine structure prediction, affinity estimation, and generative design in one pipeline, enabling use