Research Notes

AlphaFold 3 and Modeling the Molecules of Life

By Satwik · May 13, 2026

Google DeepMind and Isomorphic Labs published AlphaFold 3 in May 2024, extending structure prediction beyond single proteins to the interactions that drive biology: proteins with DNA, RNA, ligands, ions, and other proteins. Architecturally it moved to a diffusion-based module that predicts atomic coordinates directly, and it reported substantial accuracy gains on protein-ligand and protein-nucleic-acid complexes over prior specialized tools.

Why it mattered: much of biology is about complexes, not isolated chains. Predicting how a drug candidate binds a target, or how a protein reads DNA, is central to drug discovery and molecular biology. AlphaFold 3 brought accurate modeling of these interactions within reach, with clear implications for accelerating therapeutic design.

Two governance points sit in our notes. First, the release model changed: rather than fully open code, AlphaFold 3 launched via a web server with usage restrictions, prompting debate in the scientific community about reproducibility and access versus controlled release. DeepMind later broadened access to the code for academic use under conditions. That tension, openness for science versus control for safety, is now a recurring feature of frontier bio-AI.

Second is dual-use. Tools that speed beneficial drug discovery also, in principle, inform the design of harmful molecules, which is why controlled release and access conditions were defended on biosecurity grounds. We do not overstate this; structure prediction is far from a weapon, and the beneficial uses are enormous. But AlphaFold 3 is a clean case study in how capability advances in AI-for-science now come bundled with deliberate release-strategy decisions, and why "just open-source everything" is no longer the reflexive default in this domain.