Proteins are the molecular machines of life: they carry out the complex molecular processes required by cells with unrivalled accuracy and efficiency. Many of these processes depend on proteins having the ability to bind specifically to a given small molecule. If we could make proteins from scratch to bind any desired target molecule, it would open the door to a wide range of biotechnological applications that are not currently possible using natural proteins. Writing in Nature, Dou et al.1 describe a computational method for designing proteins tailored to bind a small molecule of interest, and use it to make ‘fluorescence-activating’ proteins — biotechnological tools that have potential applications in biomedical research.
The functions of proteins are dictated by the specific three-dimensional structures into which they fold. One such structure is the β-barrel: a pleated β-sheet rolled into a cylindrical structure. This is frequently found in natural proteins whose function is to bind small ligand molecules2. So far, several challenges have prevented the de novo design of β-barrel proteins that can bind target small molecules. One issue is that the cylindrical shape of β-barrels requires the two edges of the original β-sheet to come together to form the base of the cylinder, a process that is complicated by the tendency of β-sheets to form intermolecular interactions that generate protein aggregates. Another is that these cylinders must be open at one end to allow the binding of a small molecule inside the barrel, which can cause substantial destabilization of the protein’s folded structure. Thus, the design of ligand-binding β-barrel proteins requires a delicate balance between stability and ligand-binding activity.
Designing a cavity that is geometrically and chemically complementary to the small molecule of interest is just as challenging. This objective is complicated by the huge number of positions and orientations in space that the small molecule could adopt, and by the difficulty of identifying the exact combination of amino-acid residues that will make the necessary interactions with the ligand for specific binding. The latter point is particularly challenging given that there are some 10143 possible amino-acid sequences for proteins made up of 110 amino acids (as Dou and colleagues’ β-barrel is). Thus, accurate and efficient computational procedures are required to find optimal solutions from such an astronomically large number of possibilities.
Dou et al. provide a solution to these problems. They started by devising general principles for designing stably folded β-barrel proteins of a predetermined size. They discovered that structural irregularities known as glycine kinks and β-bulges must be introduced into β-barrels to alleviate molecular strain and to maintain the continuous pattern of hydrogen bonds needed to form the cylindrical structure. Using their approach, they built computational models of 500 possible β-barrel ‘backbones’ and performed calculations to identify amino-acid sequences that would stabilize each backbone. The four designs that were predicted to be most stable were synthesized, and the authors observed that one of these folded into a monomeric β-barrel. Strikingly, the computationally predicted model and the experimentally determined structure were very similar, and the designed protein was highly stable.
The authors went on to design a β-barrel protein that binds DFHBI, a small molecule that fluoresces on binding. They used an innovative computational procedure that models a large number of poses (positions and orientations in space) of a small molecule and how they dock into a protein binding site while simultaneously optimizing protein sequences for binding activity. Through this approach, the authors generated 56 β-barrel designs that were predicted to fold stably and to bind tightly to DFHBI. Of these, 20 were found to be monomeric and soluble when tested experimentally, and exhibited some of the characteristic properties of β-sheets — indicating that they were potentially β-barrels. Two of those proteins bound DFHBI with moderate binding affinities (in the range of 13–50 micromolar).
Dou and colleagues next used an iterative procedure to further improve the binding affinity of their proteins. This included steps in which the X-ray crystal structures of two protein variants were used to guide additional rounds of computational design, and in which every amino-acid residue of another variant was systematically mutated to find changes that improved binding affinity. In this way, they identified three “mini-fluorescence-activating proteins” (Fig. 1) that bind DFHBI with submicromolar affinity and enhance its fluorescence both in vitro and in vivo. These designer proteins could be used to monitor gene expression and to track proteins in cells, as well as in biosensors that detect the presence of chemicals.
The development and application of this computational method for designing β-barrel proteins that bind small molecules is the first demonstration of the de novo design of both protein fold and function, a milestone in the field. Previous computational designs of ligand-binding proteins relied on building a binding cavity into a protein template found in nature3, or one that had previously been created in the laboratory4. By contrast, Dou and co-workers have designed a β-barrel protein that has a shape distinct from those found in nature, and constructed a binding pocket that is specifically tailored to a target small molecule.
As noted earlier, the authors’ initial designs needed further optimization to identify proteins that have sufficiently high binding affinities for potential applications. More-accurate predictions of protein structures are needed to eliminate the need for such fine-tuning. One way of achieving this might come from recognizing that proteins are not rigid molecules that adopt a single predominant structure — like all machines, proteins need to move to accomplish their tasks with high efficiency5,6. Indeed, ligand binding is often the trigger that causes a protein receptor to undergo a structural change enabling the transmission of a biological signal7. Computational methods for the rational design of proteins that undergo particular structural changes have recently been developed8. If these could be combined with Dou and colleagues’ approach, it might be possible to access more-complex protein functions than were previously possible, opening the door to the on-demand creation of protein-based molecular machines.
Nature 561, 471-472 (2018)