The Illustrated AlphaFold

The article provides a detailed visual walkthrough of the AlphaFold3 architecture, focusing on how the model works to predict protein structures from sequences alone. Unlike previous versions, AlphaFold3 can predict the structure of proteins complexed with other molecules. The model consists of three main sections: Input Preparation, Representation Learning, and Structure Prediction. Input Preparation involves converting user-provided sequences into numerical tensors and retrieving similar molecules for embedding. Representation Learning updates these representations using various forms of attention. Structure Prediction utilizes these representations to predict the structure using conditional diffusion. The article also explains tokenization, retrieval of Multiple Sequence Alignments (MSA) and templates, and the representation of templates using distograms. The inclusion of MSA and templates enriches protein representations and informs structure predictions. The retrieval process involves searching for similar sequences and structures without training, enhancing the model's accuracy. Overall, the article offers a comprehensive understanding of AlphaFold3's complex architecture for protein structure prediction.