AlphaFold Server was developed as a solution to the impossibility of creating a database with all the necessary capabilities. This free tool enables scientists to input their own sequences, which AlphaFold then uses to generate molecular complexes. Since its launch in May, researchers have utilized the tool to produce over 1 million structures.

Lindsay Willmore, a research engineer at Google DeepMind, describes AlphaFold Server as the “Google Maps for molecular complexes.” Users with no coding experience can simply input their protein, DNA, RNA sequences, or the small molecule name, click a button, and within minutes receive their structure and confidence metrics for evaluation.

To accommodate a broader range of biomolecules, AlphaFold 3’s training data was significantly expanded to include DNA, RNA, small molecules, and more. This expansion allowed the model to make significant advancements in structure prediction for various molecule types.

A key change in AlphaFold 3 is the adoption of a generative model based on diffusion for the final part of the structure generation, moving away from the complex custom geometry-based module used in AlphaFold 2. This shift simplified how the model handles diverse molecule types.

However, this alteration presented a challenge as the model would inaccurately predict “ordered” structures with a spiral shape for disordered regions of proteins, which were not included in the training data. To address this, the team leveraged AlphaFold 2’s proficiency in predicting disordered interactions, providing distillation training to AlphaFold 3 to enhance its ability to predict disorder.

The team coined a new expression: “Trust the fusilli, reject the spaghetti,” emphasizing the importance of accurate disorder predictions in the molecule structures generated by AlphaFold 3.

Source: https://blog.google/technology/ai/how-we-built-alphafold-3/