AlphaFold-Multimer is a model trained to predict the structures of multi-chain protein complexes with higher accuracy than previous methods. While AlphaFold has been successful in predicting single-chain protein structures, predicting multi-chain complexes remains challenging. AlphaFold-Multimer is specifically trained on multimeric inputs with known stoichiometry, allowing it to predict interfaces between chains more accurately. On a benchmark dataset of 17 heterodimer proteins without templates, AlphaFold-Multimer achieves at least medium accuracy (DockQ ≥ 0.49) on 13 targets and high accuracy (DockQ ≥ 0.8) on 7 targets, compared to 9 targets of at least medium accuracy and 4 of high accuracy for the previous state-of-the-art system. It also predicts structures for a large dataset of 4,446 recent protein complexes, achieving high accuracy on 26% of cases for heteromeric interfaces and 36% for homomeric interfaces. The model improves upon previous methods by incorporating multi-chain features, handling symmetry, and using a modified loss function to better capture interface interactions. The model is trained on a dataset including multimers and is capable of handling multi-chain inputs during both training and inference. The model's performance is evaluated using DockQ scores, which measure the quality of predicted interfaces. AlphaFold-Multimer outperforms existing methods, including those based on AlphaFold, in predicting protein complex structures. The model's confidence metric combines predicted interface TM-score (ipTM) and single-chain TM-score (pTM) to provide a more accurate assessment of model performance. The model is implemented in Python and is available on GitHub.AlphaFold-Multimer is a model trained to predict the structures of multi-chain protein complexes with higher accuracy than previous methods. While AlphaFold has been successful in predicting single-chain protein structures, predicting multi-chain complexes remains challenging. AlphaFold-Multimer is specifically trained on multimeric inputs with known stoichiometry, allowing it to predict interfaces between chains more accurately. On a benchmark dataset of 17 heterodimer proteins without templates, AlphaFold-Multimer achieves at least medium accuracy (DockQ ≥ 0.49) on 13 targets and high accuracy (DockQ ≥ 0.8) on 7 targets, compared to 9 targets of at least medium accuracy and 4 of high accuracy for the previous state-of-the-art system. It also predicts structures for a large dataset of 4,446 recent protein complexes, achieving high accuracy on 26% of cases for heteromeric interfaces and 36% for homomeric interfaces. The model improves upon previous methods by incorporating multi-chain features, handling symmetry, and using a modified loss function to better capture interface interactions. The model is trained on a dataset including multimers and is capable of handling multi-chain inputs during both training and inference. The model's performance is evaluated using DockQ scores, which measure the quality of predicted interfaces. AlphaFold-Multimer outperforms existing methods, including those based on AlphaFold, in predicting protein complex structures. The model's confidence metric combines predicted interface TM-score (ipTM) and single-chain TM-score (pTM) to provide a more accurate assessment of model performance. The model is implemented in Python and is available on GitHub.