Google DeepMind has published the source code for the AlphaFold 3 machine learning system, designed for predicting the three-dimensional structure of proteins and modeling protein interactions with other types of molecules. The algorithms for machine learning implemented in the second version of AlphaFold were awarded the Nobel Prize in Chemistry this year. The toolkit associated with AlphaFold 3 is written in Python and C++, and is distributed under the CC BY-NC-SA 4.0 license. Trained models are provided based on a user agreement. Additionally, a tool has been launched server, allowing experimentation with AlphaFold 3 in online mode.
The model, trained on a collection describing the structures of all known proteins and amino acid sequences, addresses the protein folding problem and enables the prediction of new protein three-dimensional structures with accuracy that is at least comparable to laboratory analysis, and significantly exceeds existing prediction methods when evaluating interactions between proteins and other types of molecules. The third version of the AlphaFold model features a new architecture called 'Pairformer', which evolves the idea of the 'transformer' architecture.
Unlike AlphaFold 2, the new version is not limited to proteins consisting of a single polypeptide chain and can be applied to predict protein complexes with DNA and RNA, as well as modeling modified variants of these molecules. AlphaFold takes a list of molecules as input, and produces a joint 3D structure as output, defining the most likely interaction of the specified molecules.
From a practical perspective, AlphaFold 3 can be used for drug development and treatment methods, as well as for creating new proteins. For instance, a protein capable of attaching to specific cancer cells has been designed using AlphaFold, which could be used in next-generation cancer therapies. AlphaFold is also actively utilized in studying interactions between antibodies and proteins to understand the human immune response and to create new antibodies.
Source: opennet.ru
