AlphaFold3 at CASP16

Abstract The CASP16 experiment provided the first benchmark for AlphaFold3. In contrast to AlphaFold2 and other methods, AlphaFold3 can also predict the structure of non-protein molecules. In this study, we assess the performance of AlphaFold3 using both automatic server submissions and manual predictions from the Elofsson group. All predictions were generated via the AF3 web server, with manual interventions applied to large targets. Our analysis shows that AF3 performs comparably to top predictors for proteins and complexes, with average GDT_TS and DockQ scores indicating high model quality. Manual curation did not significantly improve results over default AF3 server submissions. Compared to AlphaFold2-based methods, we found that AF3 performs slightly better for protein complexes. Still, when massive sampling is applied to AlphaFold2, the difference disappears, using standard measurements. The performance of the AlphaFold3 server is comparable to the best methods when only taking the top-ranked predictions into account, but slightly behind when examining the best out of the five submitted models. Further, there are many targets where one method makes a good prediction while another top-ranked method fails, indicating that a venue for progress could be to develop better methods for identifying the best models. In the official ranking from CASP, AlphaFold3 performs better than AlphaFold2 for easier targets, but not for harder targets. Finally, RNA predictions remain difficult, and the accuracy of stoichiometry predictions is limited, especially for heteromeric targets. Overall, AlphaFold3 provides an easy-to-use method that offers close to state-of-the-art predictions.