This paper presents AlphaFold-Multimer, a model designed to predict the structures of multi-chain protein complexes. The model is specifically trained on known stoichiometry of multimeric inputs, significantly improving the accuracy of predicted multimeric interfaces compared to input-adapted single-chain AlphaFold while maintaining high intra-chain accuracy. On a benchmark dataset of 17 heterodimer proteins without templates, AlphaFold-Multimer achieves medium accuracy (DockQ ≥ 0.49) on 13 targets and high accuracy (DockQ ≥ 0.8) on 7 targets, outperforming the previous state-of-the-art system. For a large dataset of 4,446 recent protein complexes, AlphaFold-Multimer successfully predicts heteromeric interfaces in 70% of cases and produces high accuracy predictions in 26% of cases, an improvement of +27 and +14 percentage points over the flexible linker modification of AlphaFold, respectively. For homomeric interfaces, the success rate is 72% and the high accuracy rate is 36%, an improvement of +8 and +7 percentage points, respectively. The paper details the modifications made to the AlphaFold system to handle multiple chains during training and inference, including multi-chain permutation alignment, multiple sequence alignment construction, cross-chain genetics, and multi-chain cropping. The results demonstrate that AlphaFold-Multimer provides superior performance compared to existing approaches, including those based on using AlphaFold.This paper presents AlphaFold-Multimer, a model designed to predict the structures of multi-chain protein complexes. The model is specifically trained on known stoichiometry of multimeric inputs, significantly improving the accuracy of predicted multimeric interfaces compared to input-adapted single-chain AlphaFold while maintaining high intra-chain accuracy. On a benchmark dataset of 17 heterodimer proteins without templates, AlphaFold-Multimer achieves medium accuracy (DockQ ≥ 0.49) on 13 targets and high accuracy (DockQ ≥ 0.8) on 7 targets, outperforming the previous state-of-the-art system. For a large dataset of 4,446 recent protein complexes, AlphaFold-Multimer successfully predicts heteromeric interfaces in 70% of cases and produces high accuracy predictions in 26% of cases, an improvement of +27 and +14 percentage points over the flexible linker modification of AlphaFold, respectively. For homomeric interfaces, the success rate is 72% and the high accuracy rate is 36%, an improvement of +8 and +7 percentage points, respectively. The paper details the modifications made to the AlphaFold system to handle multiple chains during training and inference, including multi-chain permutation alignment, multiple sequence alignment construction, cross-chain genetics, and multi-chain cropping. The results demonstrate that AlphaFold-Multimer provides superior performance compared to existing approaches, including those based on using AlphaFold.