The paper presents a method for accurately predicting protein structures and interactions using a 3-track network. The method involves processing multiple sequence alignments (MSA), template information, and pair features derived from MSA. The MSA is first tokenized and embedded into a matrix, with sinusoidal positional encoding added to capture positional relationships. Template information is used to generate initial pair features, which are then processed using axial attention and merged into a 2D feature matrix. The MSA features are processed using self-attention, with soft-tied attention improving long-range contact prediction. Pair features are further refined using coevolution signals derived from MSA features, and the final pair features are updated based on structural information. The 3D structure is predicted using a Graph Transformer-based architecture, and refined using SE(3)-Transformer. The model is trained on a large dataset and evaluated on benchmark sets, showing improved accuracy compared to previous methods. The model is also used for molecular replacement calculations and structure prediction of various proteins, including GPCRs and bacterial oxidoreductases. The results demonstrate the effectiveness of the 3-track network in predicting protein structures and interactions with high accuracy.The paper presents a method for accurately predicting protein structures and interactions using a 3-track network. The method involves processing multiple sequence alignments (MSA), template information, and pair features derived from MSA. The MSA is first tokenized and embedded into a matrix, with sinusoidal positional encoding added to capture positional relationships. Template information is used to generate initial pair features, which are then processed using axial attention and merged into a 2D feature matrix. The MSA features are processed using self-attention, with soft-tied attention improving long-range contact prediction. Pair features are further refined using coevolution signals derived from MSA features, and the final pair features are updated based on structural information. The 3D structure is predicted using a Graph Transformer-based architecture, and refined using SE(3)-Transformer. The model is trained on a large dataset and evaluated on benchmark sets, showing improved accuracy compared to previous methods. The model is also used for molecular replacement calculations and structure prediction of various proteins, including GPCRs and bacterial oxidoreductases. The results demonstrate the effectiveness of the 3-track network in predicting protein structures and interactions with high accuracy.