AlphaFold2 has shown promising results in predicting protein-protein interaction interfaces, particularly for interactions mediated by intrinsically disordered regions (IDRs). Using a dataset of 42 protein-peptide complexes not redundant with AlphaFold2 training data, the study demonstrates that AlphaFold2-Multimer achieves a 40% success rate in identifying the correct interface when using full-length protein sequences. By fragmenting the proteins and combining different evolutionary information integration strategies, this success rate increases to 90%. Similar results are obtained using a larger dataset from the ELM database. The study also explores the specificity of AlphaFold2 in distinguishing between alternative binding partners, highlighting the utility of AlphaFold2 confidence scores in this task. Protein interactions are crucial for many biological processes, with IDRs playing a significant role in regulation, transport, and signaling. AlphaFold2 has been used to predict protein-protein interactions, outperforming traditional methods in terms of success rate and model quality. However, its performance is sensitive to input parameters, MSA quality, and protein delimitations. The study shows that fragment-based scanning, with fragments of 100 amino acids, significantly improves success rates, with 89% of cases correctly identifying the binding region. Combining different MSA strategies further increases the success rate to 90.5%. The study also highlights the importance of considering the biological context and the potential for intramolecular contacts to mask the binding region. For short binding motifs, the success rate may be lower, and the use of evolutionary information may be less effective. The analysis of a larger dataset from the ELM database confirms the effectiveness of the fragment scanning approach, with success rates reaching up to 87.3%. The study underscores the value of using AlphaFold2 confidence scores to assess the reliability of predictions and the importance of considering the evolutionary context when predicting interactions involving IDRs. Overall, the study demonstrates that AlphaFold2 can be a powerful tool for predicting protein-protein interactions, particularly when combined with fragment-based scanning and careful consideration of evolutionary information.AlphaFold2 has shown promising results in predicting protein-protein interaction interfaces, particularly for interactions mediated by intrinsically disordered regions (IDRs). Using a dataset of 42 protein-peptide complexes not redundant with AlphaFold2 training data, the study demonstrates that AlphaFold2-Multimer achieves a 40% success rate in identifying the correct interface when using full-length protein sequences. By fragmenting the proteins and combining different evolutionary information integration strategies, this success rate increases to 90%. Similar results are obtained using a larger dataset from the ELM database. The study also explores the specificity of AlphaFold2 in distinguishing between alternative binding partners, highlighting the utility of AlphaFold2 confidence scores in this task. Protein interactions are crucial for many biological processes, with IDRs playing a significant role in regulation, transport, and signaling. AlphaFold2 has been used to predict protein-protein interactions, outperforming traditional methods in terms of success rate and model quality. However, its performance is sensitive to input parameters, MSA quality, and protein delimitations. The study shows that fragment-based scanning, with fragments of 100 amino acids, significantly improves success rates, with 89% of cases correctly identifying the binding region. Combining different MSA strategies further increases the success rate to 90.5%. The study also highlights the importance of considering the biological context and the potential for intramolecular contacts to mask the binding region. For short binding motifs, the success rate may be lower, and the use of evolutionary information may be less effective. The analysis of a larger dataset from the ELM database confirms the effectiveness of the fragment scanning approach, with success rates reaching up to 87.3%. The study underscores the value of using AlphaFold2 confidence scores to assess the reliability of predictions and the importance of considering the evolutionary context when predicting interactions involving IDRs. Overall, the study demonstrates that AlphaFold2 can be a powerful tool for predicting protein-protein interactions, particularly when combined with fragment-based scanning and careful consideration of evolutionary information.