Abstract
AlphaFold2 has demonstrated a remarkable success in predicting the structures of globular proteins and folded domains with near-experimental accuracy. However, it typically represents intrinsically disordered regions (IDRs), protein segments that lack a stable 3D structure under physiological conditions, as long extended loops that appear to float around the structured core. While AlphaFold2's static prediction cannot capture the conformational heterogeneity and the dynamic nature of IDRs, it performs well in predicting IDRs from sequence. AlphaFold3 introduces significant architectural and training modifications over its predecessor, including the use of cross-distillation aimed at reducing structural hallucinations in disordered regions. In this study, we look into how these models differ in representing IDRs. We evaluate the performance of AlphaFold3 and AlphaFold2 on disorder prediction, using the CAID3 benchmark. Our analysis shows that AlphaFold3 does not outperform AlphaFold2 in this benchmark. We observe that solvent accessibility remains a robust and consistent proxy for predicting intrinsic disorder across both models. However, changes in the predicted secondary structure content and pLDDT scores lead to different interpretations of disorder. Overall, our findings suggest that AlphaFold2 remains the preferred choice for identifying intrinsically disordered regions, as it avoids structural hallucinations while providing predictions comparable to those of AlphaFold3.