Proteins perform a large variety of functions within organisms either alone by themselves or in cooperation with their binding partners. Studies on functions of these biomolecules and the relationship between amino acid sequence, three-dimensional structure, and functions have been prominent biological topics in the last several decades. Recently, with the rapid progress in the fields of biophysics and structural biology, ^{[ 1 ]} enthusiasm for protein design has motivated researchers to create proteins with novel structure or functions; the success of such work is also a critical test of our understanding of protein folding. These developments are important for both theoretical studies and practical applications. Proteins with specific functions are urgently needed in many fields, such as health care, agriculture, manufacture, and environmental protection. Computational design of proteins is a route to novel proteins that may be used as protein therapeutics, enzymes with special activities, or de novo proteins in synthetic organisms. ^{[ 2 , 3 ]}

In this review, we briefly discuss the theory of computational protein design and summarize recent progress in this field. The rationale for computational protein design will be introduced first, including the basic methods and strategies. Then, we will describe the research progress on de novo protein design. In the next section, design of functional proteins will be discussed, including enzyme design and design of protein–protein interactions. The last section provides conclusions and perspectives.

The basic assumption in computational protein design is that the structure of a well-folded protein is always in a state of lowest free energy. ^{[ 4 , 5 ]} Therefore, identifying the amino acid sequence with lowest energy for a given target structure with particular functions is the most important aim of protein design. Early attempts to build three-dimensional conformations of proteins for a fixed protein backbone structure, with side chain rotamers in a precomposed library, were made by Ponder and Richard. ^{[ 6 ]} The delicate procedure of using discrete rotamers has effectively reduced the number of conformations to be evaluated for a certain sequence; this development enlarged the search space of sequences greatly. Currently, the basic approach to protein design involves packing of side chains for multiple candidate sequences and evaluating the free energy of each sequence with required structure. Hence, in the past three decades, in order to improve the capabilities and efficiency of computational protein design, most efforts have been devoted to the following two goals: development of sampling methods to explore the sequence space and optimization of free-energy functions to identify the minimum (Fig.  1 ). ^{[ 7 ]}

Fig. 1.

	Figure Option View Download New Window
	Fig. 1. Flow chart of key steps in computational protein design.

As is the case with the evolution of proteins in nature, selection pressures are needed in the process of computational protein design. The function of free energy is the selection pressure that determines the direction of evolution. Therefore, the desired results can be obtained only by using the function of free energy that can correctly describe the nature of interactions within or between proteins. ^{[ 8 , 9 ]} Currently, the function of free energy usually consists of the following three parts: (i) terms that describe packing of atoms, including van der Waals forces and geometric complementarity, (ii) terms that describe the polar interactions, such as hydrogen bonding and electrostatic forces, and (iii) terms that describe hydrophobic effects and desolvation, which mainly affect the buried surface area. ^{[ 10 , 11 ]}

As mentioned above, sampling methods or a sampling strategy are another essential element of computational protein design. How long it takes to obtain a target sequence depends on the sampling strategy. Generally, a sampling strategy consists of two parts: degrees of freedom of conformations and the algorithm of sampling. The degrees of freedom of conformations include the translation and rotations of the whole protein as well as conformations of the backbone and side chain of each residue. ^{[ 12 ]} The degrees of freedom of backbones can be described by dihedral backbone angles, φ and ψ , which should satisfy the Ramachandran plot. ^{[ 13 ]} If we take the flexibility of a backbone into consideration, the side chain-coupled local motion of the backbone, known as backrub, ^{[ 14 ]} is often present. Side chain degrees of freedom are represented by rotamers. ^{[ 15 ]} The degrees of freedom also vary depending on the purposes of design. Once the parameters of the set of degrees of freedom are fixed, the conformation of the protein is predetermined. Sampling methods currently employed in computational protein design can usually be subdivided into two classes: stochastic algorithms and deterministic algorithms. By means of these algorithms, candidate sequences with better properties can be identified. Stochastic methods include the Monte Carlo (MC) methods ^{[ 16 ]} and genetic algorithms. ^{[ 17 ]} Exhaustive computations are not required when a problem is solved by stochastic algorithms. The deterministic methods, in contrast, can identify global minima of a system, for example, dead-end elimination ^{[ 18 ]} and mean-field theory. ^{[ 19 ]}

There are two major challenges for computational protein design. First, as we mentioned above, the sequence space is enormous (20 ^N for an N -residue protein) ^{[ 20 ]} and makes it impossible to evaluate all sequences. Thus, the features of the sequence space cannot be thoroughly analyzed by the sampling strategies that we introduced above. Rather than trying to find all possible solutions in the sequence space, researchers are now paying more attention to the question whether proteins with a particular function or structure can be successfully designed. Nevertheless, access to all possible solutions in the sequence space would surely result in more knowledge about the relation among sequence, structure, and function of a protein. With the increasing number of computational approaches and experimental techniques, more custom-designed proteins are expected to be discovered and tested empirically. On the other hand, more accurate and reasonable functions for scoring are needed, ^{[ 21 ]} because in the huge sequence space, there is a need for accurate calculations regarding the candidates. Given the large number of possible results with different sequences, the accuracy of functions for scoring is key to the efficiency of computational protein design. Commonly used energy functions are based on physical potential, statistical potential, or a hybrid of both. ^{[ 22 – 25 ]} No matter which strategy is employed, as long as the accuracy is improved, computational protein design will most likely be successful.

In this field, early famous studies were focused on design of a novel amino acid sequence for desired structure. Many of these proteins were designed rationally, based on the empirical principles of how three-dimensional structures are determined by the amino acid sequence that researchers have deduced. For instance, DeGrado and coworkers designed three-helix and four-helix bundle proteins (Fig.  2(b) ). ^{[ 26 , 27 ]} Later on, possible sequences can be produced with the help of software that can assemble the side chains of residues automatically. The candidate proteins predicted by the software were then tested empirically. In 1997, Mayo’s group successfully designed a β β α protein motif with a novel sequence, which was designed automatically and validated by empirical experiments (Fig.  2(a) ). ^{[ 28 ]} In 2003, Baker’s group designed a 97-residue α β protein motif called Top7 (Fig.  2(c) ), which had not been found in nature, using the RosettaDesign software. ^{[ 29 ]} This is the first successful attempt to design proteins with novel folds, and researchers are inspired to explore the large regions of protein structures that have not been found in nature. Our group has designed a stable β α β motif consisting of two parallel β -sheets connected by an α -helix (Fig.  2(d) ). This β α β motif has also not been found in nature as a stand-alone protein. ^{[ 30 ]} In 2012, based on the rules of how secondary-structure patterns translate into a tertiary motif, Bakers’ group proposed a set of principles for designing “ideal” proteins. ^{[ 31 ]} In accordance with these principles, they successfully designed novel proteins with five different folds. These studies not only explore the infinite possibilities of protein structures, but should also help to reveal the process of protein folding.

Fig. 2.

	Figure Option View Download New Window
	Fig. 2. De novo designed proteins. (a) De novo designed three-helix bundle protein. [ 27 ] (b) De novo designed β β α protein motif. [ 28 ] (c) Design of novel β α protein motif, TOP7. [ 29 ] (d) Design of novel β α β protein motif. [ 30 ]

Recently, novel proteins with more complicated tertiary structures and folds have been designed, which possess some basic functions or potential applications in relevant fields. For instance, Baker’s group has designed protein complexes with complicated structures formed by self-assembling. To be more specific, they designed a 13-nm-diameter complex composed of 24 subunits and an 11-nm-diameter complex composed of 12 subunits, using trimeric proteins as building blocks. The elementary subunits were firstly selected by docking them into the target architecture. Then, the researchers designed the self-assembly interface. ^{[ 32 ]} In addition, right-handed and left-handed antiparallel bundles and a pentameric bundle were designed in 2014. ^{[ 33 ]} Barrels consisting of an α helix, which are a novel structure, have been designed by Woolfson et al. , on the basis of geometric properties and knowledge-based scoring of the helices. ^{[ 34 ]} Hydrophilic channels are observed in these barrels consisting of five, six, or seven helices. Similar studies were performed by DeGrado’s group. They designed dimetal binding sites in a four-helix bundle. They observed a flux of Zn ²⁺ and Co ²⁺ through this designed structure. ^{[ 35 ]} In summary, recently, de novo designed proteins went beyond simple structures; complicated structures are now obtained using basic structural motifs as building blocks. More proteins with more complicated structure and versatile functions are expected to be explored by means of de novo design.

Many successful studies have been conducted in the field of de novo protein design, enzyme design, and protein–protein interaction design, but strategies and tools for computational protein design are far from perfect. In studies involving protein design, sequences with high activity can hardly be obtained after only a single round of design. A survey of relevant publications shows that only proteins with low activity can be obtained in the first round of design in most cases. ^{[ 32 , 42 , 69 ]} Thus, the initial sequences from computational design provide only a starting point. In order to obtain proteins with a desired function, empirical studies are needed to optimize the design. How natural proteins evolve into the ones with high affinity or catalytic activity has yet to be explored. Moreover, binding or catalysis that involves dramatic conformational changes still cannot be taken into consideration during computational protein design. Nor can the binding free energy be accurately calculated. To solve these problems, the existing theories and computational approaches need to be improved. With the development of new strategies and methods, together with the maturity of experimental techniques of protein engineering, computational protein design is attracting more and more interest and yielding promising results for practical applications in biomedical research and for development of novel therapeutics.