The article presents a computational approach to design complex protein folds and soluble analogues of integral membrane proteins using deep learning. The authors developed a robust deep learning pipeline that combines AlphaFold2 (AF2) for structure prediction and ProteinMPNN for sequence design. This pipeline was used to design highly stable folds, such as Ig-like folds (IGF), β-barrel folds (BBF), and TIM-barrel folds (TBF), as well as soluble analogues of membrane protein folds like claudin, rhomboid protease, and G-protein-coupled receptors (GPCR). The designs were characterized for thermal stability and functional motifs, demonstrating high experimental success rates and remarkable design accuracy. The soluble analogues retained native functional motifs, such as G-protein-binding interfaces and toxin-receptor interaction sites, suggesting the potential for new approaches in drug discovery. The study highlights the effectiveness of deep learning in expanding the functional soluble fold space and enabling the design of complex protein topologies with preserved functionalities.The article presents a computational approach to design complex protein folds and soluble analogues of integral membrane proteins using deep learning. The authors developed a robust deep learning pipeline that combines AlphaFold2 (AF2) for structure prediction and ProteinMPNN for sequence design. This pipeline was used to design highly stable folds, such as Ig-like folds (IGF), β-barrel folds (BBF), and TIM-barrel folds (TBF), as well as soluble analogues of membrane protein folds like claudin, rhomboid protease, and G-protein-coupled receptors (GPCR). The designs were characterized for thermal stability and functional motifs, demonstrating high experimental success rates and remarkable design accuracy. The soluble analogues retained native functional motifs, such as G-protein-binding interfaces and toxin-receptor interaction sites, suggesting the potential for new approaches in drug discovery. The study highlights the effectiveness of deep learning in expanding the functional soluble fold space and enabling the design of complex protein topologies with preserved functionalities.