AlphaFold 3 publishes its architecture and pseudocode but keeps code and weights behind a non-commercial server

2024-05-08Industry

Josh Abramson et al., "Accurate structure prediction of biomolecular interactions with AlphaFold 3," Nature (published online May 8, 2024), doi:10.1038/s41586-024-07487-w; Google DeepMind AlphaFold team and Isomorphic Labs, "AlphaFold 3 predicts the structure and interactions of all of life's molecules," blog.google (May 8, 2024)

Google DeepMind and Isomorphic Labs published AlphaFold 3 in Nature, received December 19, 2023 and accepted April 29, 2024. The paper describes a model that predicts the "joint structure of complexes including proteins, nucleic acids, small molecules, ions and modified residues." It replaces the Evoformer of AlphaFold 2 with a Pairformer that operates on pair and single representations with reduced multiple-sequence-alignment processing, and replaces the structure module with a diffusion module that predicts raw atom coordinates directly. The abstract reports "far greater accuracy for protein–ligand interactions compared with state-of-the-art docking tools," "much higher accuracy for protein–nucleic acid interactions compared with nucleic-acid-specific predictors" and "substantially higher antibody–antigen prediction accuracy compared with AlphaFold-Multimer v.2.3." The methods give full algorithm details with pseudocode in Supplementary Algorithms 1 to 31, and the training and evaluation data come from public sources, principally the Protein Data Bank. AlphaFold 2's source code and trained weights were released under an open-source license with its July 2021 paper; AlphaFold 3 is released without code or weights. Access is through AlphaFold Server, a free platform for non-commercial research that models proteins, DNA, RNA, a selection of ligands, ions and chemical modifications, while Isomorphic Labs applies the model with pharmaceutical partners to drug design.

What it changesAlphaFold 3 splits a model between publication and secrecy. The architecture, the training approach and the pseudocode are now prior art against every filing made after May 8, 2024 on diffusion-based structure prediction, by competitors and by the authors themselves outside the United States grace period, while the trained weights, the product of training runs the paper describes but does not hand over, stay behind server terms. Companies that publish model research can make the same split deliberately: file on the architecture and training method before publication, publish what builds standing in the field, and keep weights and data pipelines as trade secrets, with access controls and use restrictions as the reasonable measures that secrecy requires. For pharmaceutical and chemistry teams, the server's non-commercial limit means commercial discovery work cannot run on it, and compounds designed from any structure prediction still need an inventorship record showing the chemist's contribution beyond selecting the predicted structure.

More on Trade secrets and open release · Prior art, public disclosure and obviousness

← All recent developments · Search all developments

Burbage Law is the practice of Jon Burbage, a registered patent attorney. 342 United States patents have issued from applications he wrote; the full list is on the Patents page. More about Jon Burbage.

This is general information about a published decision, agency guidance or piece of legislation. It is not legal advice and does not create an attorney-client relationship. Items concerning offices outside the United States are reported for information only and are not counsel on the law of those countries.