This study presents a computational approach for designing soluble and functional membrane protein analogues using deep learning. The researchers developed a robust pipeline based on the inversion of the AlphaFold2 (AF2) structure prediction network and sequence design using ProteinMPNN. This method enables the design of complex protein folds and soluble analogues of integral membrane proteins, including G-protein-coupled receptors (GPCRs), claudins, and rhomboid proteases. The designs were shown to be highly stable and accurate, with experimental structures demonstrating remarkable design precision. The soluble analogues were functionalized with native structural motifs, demonstrating the potential to bring membrane protein functions to the soluble proteome, which could enable new approaches in drug discovery. The study highlights the challenges of designing complex protein folds, particularly those with non-local topologies and large sizes. However, the developed approach allows for the design of such folds without the need for parametric or symmetric restraints, leading to a de facto expansion of the functional soluble fold space. The computational framework based on AF2-seq-MPNN is flexible and generalizable, avoiding the need for fold-specific retraining or parametric design constraints. The researchers successfully designed and characterized several folds that have been challenging to engineer with previous methods, achieving high experimental success rates in terms of soluble and folded designs. Structural characterization of the designs showed very high accuracy, both in terms of overall fold and fine details of side-chain conformations, which are critical for the design of function. The study also demonstrated the ability to expand the soluble fold space and enable the design of analogues of protein topologies only found in membrane environments. The study further shows that membrane protein folds generally follow the same design principles as soluble protein folds, and that many such folds can be readily designed in soluble form. The researchers propose that this could promote the designability of functional proteins by enabling access to a plethora of folds not present in the soluble fold space. Additionally, the creation of soluble analogues of membrane proteins that retain many of the native features of the original membrane proteins, such as enzymatic or ligand-binding functions, could greatly accelerate the study of these proteins in more biochemically accessible soluble formats. The study demonstrates the potential of the method by incorporating native structural motifs into designed soluble analogues. By designing soluble analogues in the context of the natural functional site, the researchers preserved even complex structural features of the sites, such as the extracellular β-sheeted domains of claudins. The precision of the design approach enabled conformational-specific design for the active and inactive GPCR states, differentiated by subtle conformational changes. Consequently, the designs harbored identical G-protein-binding sites, yet they uniquely either constitutively facilitated or precluded G protein binding in solution. The computational design of specific conformational states that can mediate biological function remains an outstanding problem for which the researchers provide a flexible and broadly applicable methodological workflow. Such an approach could constitute aThis study presents a computational approach for designing soluble and functional membrane protein analogues using deep learning. The researchers developed a robust pipeline based on the inversion of the AlphaFold2 (AF2) structure prediction network and sequence design using ProteinMPNN. This method enables the design of complex protein folds and soluble analogues of integral membrane proteins, including G-protein-coupled receptors (GPCRs), claudins, and rhomboid proteases. The designs were shown to be highly stable and accurate, with experimental structures demonstrating remarkable design precision. The soluble analogues were functionalized with native structural motifs, demonstrating the potential to bring membrane protein functions to the soluble proteome, which could enable new approaches in drug discovery. The study highlights the challenges of designing complex protein folds, particularly those with non-local topologies and large sizes. However, the developed approach allows for the design of such folds without the need for parametric or symmetric restraints, leading to a de facto expansion of the functional soluble fold space. The computational framework based on AF2-seq-MPNN is flexible and generalizable, avoiding the need for fold-specific retraining or parametric design constraints. The researchers successfully designed and characterized several folds that have been challenging to engineer with previous methods, achieving high experimental success rates in terms of soluble and folded designs. Structural characterization of the designs showed very high accuracy, both in terms of overall fold and fine details of side-chain conformations, which are critical for the design of function. The study also demonstrated the ability to expand the soluble fold space and enable the design of analogues of protein topologies only found in membrane environments. The study further shows that membrane protein folds generally follow the same design principles as soluble protein folds, and that many such folds can be readily designed in soluble form. The researchers propose that this could promote the designability of functional proteins by enabling access to a plethora of folds not present in the soluble fold space. Additionally, the creation of soluble analogues of membrane proteins that retain many of the native features of the original membrane proteins, such as enzymatic or ligand-binding functions, could greatly accelerate the study of these proteins in more biochemically accessible soluble formats. The study demonstrates the potential of the method by incorporating native structural motifs into designed soluble analogues. By designing soluble analogues in the context of the natural functional site, the researchers preserved even complex structural features of the sites, such as the extracellular β-sheeted domains of claudins. The precision of the design approach enabled conformational-specific design for the active and inactive GPCR states, differentiated by subtle conformational changes. Consequently, the designs harbored identical G-protein-binding sites, yet they uniquely either constitutively facilitated or precluded G protein binding in solution. The computational design of specific conformational states that can mediate biological function remains an outstanding problem for which the researchers provide a flexible and broadly applicable methodological workflow. Such an approach could constitute a