Computational Methods for De novo Protein Design and its Applications to the Human Immunodeficiency Virus 1, Purine Nucleoside Phosphorylase, Ubiquitin Specific Protease 7, and Histone Demethylases

Abstract

This paper provides an overview of computational de novo protein design methods, highlighting recent advances and successes. Four protein systems are described that are important targets for drug design: human immunodeficiency virus 1, purine nucleoside phosphorylase, ubiquitin specific protease 7, and histone demethylases. Target areas for drug design for each protein are described, along with known inhibitors, focusing on peptidic inhibitors, but also describing some small-molecule inhibitors. Computational design methods that have been employed in elucidating these inhibitors for each protein are outlined, along with steps that can be taken in order to apply computational protein design to a system that has mainly used experimental methods to date.

1 INTRODUCTION

De novo protein design, also referred to as the inverse protein folding problem, is the determination of an amino acid sequence, or set of sequences, that will fold into a given 3-dimensional (3D) protein template, which may be fixed or flexible. While the protein folding problem seeks to identify the one lowest energy conformation for a given amino acid sequence, the protein design problem exhibits degeneracy, in that many amino acid sequences fold into a given template, with the different sequences giving different properties (activity, specificity) to the protein. It therefore has a wide range of applications, from improved design of inhibitors and new sequences with increased stability to the design of catalytic sites of enzymes and drug discovery [1–3].

Until recently, protein design consisted primarily of experimental techniques such as rational design, mutagenesis, and directed evolution. Although these methods produce good results, they are restrictive because of the limited sequence search space (estimated to be only 10³ – 10⁶). Computational methods, on the other hand, can increase this search space to 10¹²⁸, making computational protein design more popular.

Several successes in protein design include increasing the stability and specificity of a target protein [4–6] to locking proteins into useful conformations [7]. Computational methods aid the protein design process by determining folding kinetics [4, 8] and protein-ligand interactions [9]. They help with protein docking [10–12] and assist peptide and protein drug discovery [13–15].

Despite these successes, there are limitations. Currently, it is very difficult to design a protein consisting of 100 or more amino acids. If one assumes an average of 100 rotamers for all 20 amino acids at each position, this problem reaches a complexity of 100¹⁰⁰ = 10²⁰⁰. Coupled with the NP-hard nature [16, 17] of the problem, designing larger proteins (> 100 amino acids) proves a great challenge. In addition to improving the computational efficiency of de novo design algorithms, another challenge is to incorporate true backbone flexibility. These two challenges are interrelated, as incorporating backbone flexibility increases the computational complexity of an algorithm.

The next few sections outline the methodologies and recent advances in computational protein design, using both fixed and flexible backbone templates and describing both deterministic methods and stochastic methods.

2 COMPUTATIONAL METHODS

The various computational methods employed for de novo protein design belong to two classes: those that use fixed backbone templates and those that use flexible backbone templates. A fixed backbone template consists of fixed backbone atom coordinates and fixed rotamer conformations. This was first proposed by Ponder and Richards [18]. This is usually the case when only an X-ray crystal structure of the design template is known. Flexible backbone templates, on the other hand, are more true to nature, as protein structures are inherently flexible. Flexible templates can be a set of fixed backbone atom coordinates, such as the set of structure models obtained from NMR structure determination. Instead of a set of fixed atoms coordinates, the backbone atoms can take on a range of values between specified bounds. The rotamers can also consist of a set of discrete rotamers for each residue or the rotamer angles can be allowed to vary between a specified range.

2.1 Fixed Backbone Templates

2.1.1 Deterministic Methods

Deterministic algorithms include those that use (a) dead end elimination (DEE) methods, (b) self-consistent mean field (SCMF) methods, (c) power law (PL) methods or (d) those that utilize quadratic assignment-like models coupled with deterministic global optimization. The deterministic methods (a), (b), and (c) use a discrete set of rotamers, which are used for tractability of the search problem, while methods (d) can use either a discrete or a continuous set of rotamers.

DEE methods historically use fixed-backbone templates and a discrete set of rotamers [19–23]. DEE works by systematically eliminating rotamers that cannot be part of the sequence with the lowest free energy. The energy function used in DEE is a combination of individual terms (rotamer to template) and pairwise terms (rotamer to rotamer) and is given by Eq. 1, where E(i_a) is the rotamer-template energy for rotamer i_a of amino acid i. Likewise, E(i_a,j_b) is the pairwise rotamer-rotamer energy between rotamer i_a of amino acid i and rotamer j_b of amino acid j. Finally, N is the total number of amino acid positions.

$$E=\sum _{i=1}^{N}E\left(i_a\right)+\sum _{i=1}^{N-1}\sum _{j>i}^{N}E(i_a,j_b)$$ (1)

The general idea behind DEE is that if the pairwise energy, E(i_a,j_b), between rotamer i_a and j_b is greater than that between i_c and j_b, for all j_b in some set {B}, then the rotamer i_a can be removed, or pruned, from the set because it cannot be part of the global energy minimum conformation. This relationship is expressed by Eq. 2. This equation can be made more tractable by the bounds implied by Eq. 2, given in Eq. 3.

$$E\left(i_a\right)+\sum _{j\ne i}^{N}E(i_a,j_b)>E\left(i_c\right)+\sum _{j\ne i}^{N}E(i_c,j_b),\forall \left\{B\right\}$$ (2)

$$E\left(i_a\right)+\sum _{j\ne i}^N\underset{b}{\min }E(i_a,j_b)>E\left(i_c\right)+\sum _{j\ne i}^N\underset{b}{\max }E(i_c,j_b)$$ (3)

Recent advances to the DEE algorithm allow faster computation of larger protein design problems. BroMAP (branch-and-bound rotamer optimization using MAP estimation) expands smaller search trees while performing moderate computation in each node, reducing the total run time [24]. The BroMAP method generates lower bounds using approximate maximum-a-posteriori (MAP) estimation. Another advance, called MinDEE (for minimized-DEE) not only prunes rigid rotamers as described by the original DEE algorithm, but guarantees that rotamers belonging to the energy minimized global minimum energy conformation will not be pruned [25]. This is important because conformations that survive the original DEE algorithm are often subjected to further rotamer energy minimizations. When this occurs, the combined process is heuristic and the provable guarantee is lost. Other modifications to DEE include X-DEE (extended DEE), which gives gap-free lists of low-energy states for a given energy range and was applied to the determination of protonation states of a protein [26], and type-variant DEE, which can be used in multistate protein design [27]. Further descriptions of DEE modifications and successes can be found in Fung et al. [3].

SCMF methods involve predicting the values of elements of a conformational matrix P(i,a) to determine the probability that amino acid i will adopt a rotamer conformation a. Koehl and Delarue [28] were among the first to describe such a method in protein design. This iterative process begins with an initial guess for the conformational matrix, giving equal probability to all rotamers (Eq. 4).

$$P(i,a)=\frac{1}{A},a=1,2,\dots ,A$$ (4)

The energy, based upon the mean field potential E(i,a), is then calculated, which is dependent upon the conformational matrix.

$$E(i,a)=U\left(x_{i_a}\right)+U(x_{i_a},x_0)+\sum _{j=1,j\ne i}^N\sum _{b=1}^{B}P(j,b)U(x_{i_a},x_{j_b})$$ (5)

In equation 5, x₀ is the coordinates of atoms in the fixed template, while x_ia and x_jb are the coordinates of atoms of position i with rotamer conformation a and position j with rotamer conformation b, respectively. The potential energy, U, can described by classical Lennard-Jones potential. Once the energy has been calculated, the conformational matrix is then updated (Eq. 6) and this is repeated until convergence is obtained. It should be noted, however, that although this method is deterministic, it does not guarantee convergence to the global minimum solution, as DEE does.

$$P(i,a)=\frac{e^{\frac{E(i,a)}{RT}}}{\sum _{a=1}^A{e}^{\frac{E(i,a)}{RT}}}$$ (6)

The SCMF method was extended by the Saven group [29–32], introducing information theory concepts and formulating it as an optimization problem. They maximized the sequence entropy subject to composition and mean-field energy constraints. Bhattacherjee and Biswas [33] recently developed a SCMF based theory in order to consider the effect of correlated mutations by evaluating site-specific amino acid pair probabilities in a library of sequences.

The power law (PL) method is based on the similarity of a topology optimization problem and a protein design problem [34]. Instead of solving the topology optimization problem in discrete space, however, it is solved in continuous space in order to reduce the intensive computation required. The problem introduced by this method is the artifact of obtaining non-discrete optimal solutions. This is solved by introducing solid isotropic material with penalization (SIMP). The rigidity of a material is expressed as an artificial power law in continuous space. Using this formulation, the interaction energy between the known amino acid at position j and the unknown amino acid at position i is given by Eq. 7.

$$E_{ij}\left(x_j\right)=-E_{ij}^0{x}_i^p$$ (7)

The energy value between the amino acids at sites i and j is given by E⁰_ij. This is also referred to as the stiffness of the material, in terms of the SIMP formulation. The state of the undetermined amino acid at position i is represented by the variable x^p_i. Equation 7 is used as the energy function in the quadratic programming problem given by Eq. 8, where B_ij is constructed in order to obtain equality constraints that limit the selection of one amino acid per position, N is the number of residue positions, m is the number of amino acid types available, and Q_ij contains the interaction energy between residues.

$$\min x_i^p{Q}_{ij}{x}_j^p$$

$$\begin{array}{cc}\hfill \text{subject to}\sum _{j=1}^{Nm}& B_{ij}{x}_j=c_i\text{for}i=1,2,\dots ,N+m\hfill \\ \hfill & 0\le x_i\le 1\text{for}i=1,2,\dots ,Nm\hfill \end{array}$$ (8)

Klepeis et al. [5, 6] proposed a novel approach to de novo protein design using a quadratic assignment-like global optimization model to select sequences. The original integer linear programming (ILP) model uses a single fixed backbone template structure. Further advances by Fung et al. [16, 35, 36], incorporating backbone flexibility into the model, are presented in subsequent sections.

2.1.2 Stochastic Methods

Because the protein design problem is NP-hard [16, 17], deterministic methods may reach a limit and stochastic methods are employed to find locally optimal solutions. These methods include genetic algorithms, Monte Carlo sampling, and combinatorial library methods.

Monte Carlo (MC) methods work by performing a mutation of a residue at a certain position and calculating the energies of the sequence before and after the mutation. The mutation will be accepted if it lowers the energy, or, if the energy is higher, it will accept the mutation based upon the Metropolis criterion [37] (Eq. 9).

$$P_{\mathit{accept}}=\mathit{min}\{1,e^{-\beta \Delta E}\},\beta =\frac{1}{kT}$$ (9)

In Eq. 9, P_accept is the probability that a move will be accepted, ΔE is the change in energy before and after adding a mutation, T is the temperature, and k is the Boltzmann constant. Rosetta Design [38] is a MC based algorithm. This algorithm has options for both fixed-backbone template design and flexible-backbone template design (considering an ensemble of fixed backbones), however, only local minima are obtained, not necessarily the global minimum. Two complementary Markov chain Monte Carlo (MCMC) protein design methods were developed by Thomas et al. [39]. One uses constrained shuffling which generates a small set of high-likelihood new sequences. The other uses component sampling, which generates all the sequences that meet given constraints. Zou and Saven [40] combine MC methods with SCMF methods to bias the sequence selection using predetermined sequence identity probabilities. The probabilities are determined using SCMF methods, while the sequence selection is performed using MC methods. Successes of MC methods consist of new sequences for the fixed backbones of the β1 domain of protein G, λ repressor, and sperm whale myoglobin [41, 42], and the redesign of the src SH3 domain, the λ repressor, U₁A, protein L, tenascin, procarboxypeptidase, acylphosphatase, S6, and FKBP12 [4].

Genetic algorithms, in comparison, generate a number of random amino acid sequences and exchange them for a fixed template [43–47]. This method originated in genetics and evolution [48]. High-energy sequences are eliminated, while low energy ones form hybrids. The algorithm terminates when the solution converges. Genetic algorithms have been used to design ligands for prolyl oligopeptidase, p53, and DNA gyrase [49] and to design short peptides with high stability that resemble the antibody epitopes of thrombin and blood coagulation factor VIII [50].

Finally, combinatorial library methods attempt to maximize the entropy subject to a set of constraints [30, 32, 51–57].

2.2 Flexible Backbone Templates

2.2.1 Deterministic Methods

Georgiev and Donald [58] incorporated backbone flexibility into the DEE method. As previously mentioned, the original DEE algorithm does not take backbone flexibility into account. Modifications to DEE have attempted to do so, but the modified DEE algorithms don't have the provable guarantee of eliminating rotamers that are not part of the global minimum energy conformation. Georgiev and Donald [58] provided modifications in which this provable guarantee is kept intact and allows for backbone flexibility. Equation 3 is expressed equivalently as

$$\begin{array}{cc}\hfill & E_{\Theta }\left(i_a\right)+\sum _{j\ne i}^N\underset{b}{\min }{E}_{\Theta }(i_a,j_b)\hfill \\ \hfill -& E_t-\sum _{j\ne i}^N\underset{b}{\max }{E}_{\Delta }\left(j_b\right)-\sum _{j\ne i}^N\sum _{k\ne i,k>j}^N\underset{b,d}{\max }{E}_{\Delta }(j_b,k_d)\hfill \\ \hfill >& E_{\Phi }\left(i_c\right)+\sum _{j\ne i}^N\underset{b}{\max }{E}_{\Phi }(i_c,j_b),\forall \left\{B\right\},\left\{D\right\},\hfill \end{array}$$ (10)

where E_t represents the difference between lower and upper bounds on the template energy. E_t is the interval of possible template energies. E_Θ(i_a) represents a lower bound on the sum of the energy interactions between the atoms of rotamer i_a and the energy interactions between the atoms of rotamer i_a and the template atoms. Similarly, E_Θ(i_a,j_b) is defined as the lower bound pairwise energy between rotamers i_a and j_b. E_Θ(i_a) and E_Θ(i_c,j_b) are defined as the maximum bounds, similar to the descriptions for E_Θ. Finally, the interval terms, E_Δ, are defined as

$$\begin{array}{cc}\hfill E_{\Delta }\left(i_a\right)& =E_{\Phi }\left(i_a\right)-E_{\Theta }\left(i_a\right)\hfill \\ \hfill E_{\Delta }(i_a,j_b)& =E_{\Phi }(i_a,j_b)-E_{\Theta }(i_a,j_b).\hfill \end{array}$$ (11)

What Eq. 10 essentially does is, given a compact space of backbone conformations, compare a lower bound on the energy achievable when amino acid i has a rotamer identity of i_a to an upper bound on the energy achievable with a competing rotamer identity say i_e, of the same amino acid. Georgiev et al. [59] also introduced backbone flexibility into the DEE algorithm by incorporating back rub motions. Back rub motions are side-chain coupled local backbone motion and is commonly observed in proteins.

A novel modeling and computational framework introduces a quadratic assignment-like global optimization model [5, 6, 16, 35, 36, 60] to select sequences. It incorporates backbone flexibility by describing the template in terms of distance bins, with each bin given the same energy. A single backbone template or a set of multiple templates obtained from NMR experiments, molecular dynamics simulations, or docking calculations can be used. From this, a rank ordered list of sequences is obtained. Equation 12 shows the basic mixed-integer linear programming (MILP) model developed by Fung et al. [16].

$$\underset{y_i^j{y}_k^l}{\min }\sum _{i=1}^{n-1}\sum _{j=1}^{m_i}\sum _{k=i+1}^n\sum _{l=1}^{m_k}{E}_{ik}^{jl}(x_i,x_k)w_{ik}^{jl}$$

$$\begin{array}{cc}\hfill \text{subject to}& \sum _{j=1}^{m_i}{y}_i^j=1\forall i\hfill \\ \hfill & \sum _{j=1}^{m_i}{w}_{ik}^{jl}=y_k^l\forall i,k>i,l\hfill \\ \hfill & \sum _{l=1}^{m_k}{w}_{ik}^{jl}=y_i^j\forall i,k>i,j\hfill \\ \hfill y_i^j,& y_k^l,w_{ik}^{jl}\in \{0,1\}\forall i,j,k>i,l\hfill \end{array}$$ (12)

The model minimizes the sum of all pairwise energy interactions. The energy term, E^jl_ik, is dependent upon the distance between residue i and residue k. The distances can either be the C^α–C^α distances or the centroid-centroid distances. Backbone flexibility is introduced in two ways, the first is by discretizing the energy function into bins based upon the distance between a pair of residues. Therefore, a pair of residues may take on a range of distances that fall within a particular distance bin and still retain the same energy value. The second way flexibility is introduced is by expanding the model depicted in Eq. 12 to allow for the distances spanned by two resides to fall across multiple bins. This is achieved in two ways: by introducing a weighted average energy term or by introducing a binary variable that selects a particular distance bin for a pair of residues depending upon all the bins the pair spans. Details of these model formulations can be found in Fung et al. [36].

The sequences selected using the MILP models are subsequently re-ranked in a validation stage that consists of fold specificity calculations, approximate binding affinity calculations, or both. Both validation methods incorporate true backbone flexibility. A fold specificity for each sequence is calculated by generating an ensemble of structures of the mutant sequence and of the native sequence. The ensemble incorporates backbone flexibility by allowing a continuous range between upper and lower bounds on the C^α-C^α distances and ϕ and ψ angles. The fold specificity is determined according to Eq. 13, where β = 1/k_BT.

$$f_{\mathit{specificity}}=\frac{\sum _{i\in \mathit{novel}}{\exp }^{-\beta E_i}}{\sum _{i\in \mathit{native}}{\exp }^{-\beta E_i}}$$ (13)

The set novel contains the ensemble of structures of the mutant sequence, while the set native contains the ensemble of structures of the native sequence. The fold specificity is a measure of how likely a mutant sequence will fold into the native structure. A higher fold specificity indicates a greater likelihood. Fold specificities are calculated based on (a) the full atomistic forcefield either through the ASTRO-FOLD approach [61–70] and deterministic global optimization [71–77] or (b) the AMBER forcefield via a novel NMR structure refinement method [3, 35].

The second validation method is most applicable when designing proteins or peptides to bind to another target protein. This method re-ranks the sequences from the selection stage by calculating an approximate binding affinity (Eq. 14).

$$K\ast =\frac{q_{PL}}{q_P{q}_L}$$ (14)

The approximate binding affinity (K*) is a ratio of the partition functions of the complex (PL), target protein (P), and peptide or ligand (L). The partition functions are calculated using rotamerically-based conformation ensembles for each species. Backbone flexibility is introduced in this validation method by the way in which the ensembles are constructed, allowing the peptide ensemble to take on a number of discrete backbone and rotamer configurations and allowing the complex ensemble to take on a number of discrete peptide backbone and rotamer configurations and docked configurations. Details of the approximate binding affinity calculation can be found in Bellows et al. [78].

In addition to the two methods mentioned above, groups have applied the methodologies pertaining to fixed backbone templates to multiple discrete templates in order to add some flexibility to their design. Su and Mayo [79] and Ross et al. [80] used the DEE algorithm with fixed backbone template for each of their backbone templates in order to generate several sets of perturbed backbones from the native structure of their design. Kono and Saven [30] used their SCMF method on a set of fixed backbone templates in order to design sequences that would be robust to distance changes in the template.

2.2.2 Stochastic Methods

Monte Carlo simulations have been used to generate multiple discrete templates based upon an initial fixed-backbone template [7, 81, 82]. Using this, the Pande group [81, 82] used genetic algorithms with the discrete templates to design sequences that exhibited higher diversity than the corresponding natural sequence alignments. Kraemer-Pecore et al. [7] also used genetic algorithms with multiple discrete templates to identity a sequence that folded into the WW domain.

Another example of using genetic algorithms with Monte Carlo simulations involves iterating between sequence space and structure space. Desjarlais and Handel [83] generated multiple discrete backbone templates and used a genetic algorithm to exchange not only rotamers but also backbone torsional information in recombination. The backbone structures were then refined using Mote Carlo. Using this method, they were able to design three new core variants of the protein 434 cro.

Monte Carlo methods have also been used by Kuhlman et al. [84] and Saunders and Baker [85] where the method again uses a set of discrete backbone templates. A Monte Carlo search gives the lowest energy sequence for each template, then performs structure prediction of the sequences to allow for shifts in structure space, and continues to iterate between sequence space and structure space for a desired number of iterations. They have successfully used this method to design a new sequence for Top 7, a 93 residue α/β protein with a novel fold.

Finally, Harbury et al. [86–88] have used a continuum backbone template with discrete rotamers to design a family of α-helical bundle proteins with a right-handed superhelical twist. Their method works by creating a continuum template using backbone parameterization and then performing a sequence search from rotamer libraries. Backbone movements are introduced by treating the parameters as variables during the sequence search for energy minimization.

3 APPLICATION DOMAINS

The computational protein design methods described can be applied to a number of systems. Four systems are outlined below. All involve using de novo protein design methods to design inhibitors that target the human immunodeficiency virus 1, purine nucleoside phosphorylase, ubiquitin specific protease 7, and histone demethylases. Some systems have a long history of using computational de novo design to discover inhibitors, while others have been treated with only experimental techniques. The following sections review the work that has been done in each of the four areas with regard to target identification, inhibitor design and discovery, and what computational methods, if any, have been applied to the system.

3.1 Human Immunodeficiency Virus 1

The human immunodeficiency virus 1 (HIV-1) is the virus that causes AIDS, the acquired immunodeficiency syndrome. While there is no vaccine for HIV-1, there are treatments for those who have been infected. Called antiretroviral treatment, it reduces the risks associated with HIV if begun right after infection. The treatment was first developed in 1995 and consists of a cocktail of at least three drugs. The cocktail usually includes two nucleoside analogue reverse transcriptase inhibitors and either a protease inhibitor or a non-nucleoside reverse transcriptase inhibitor [89]. These drugs work by targeting different aspects of the HIV replication process, leading to a rise in CD4 T cells (the target of HIV). However this is then followed by an increase in drug-resistant variants of HIV.

Recently, new studies have focused on inhibiting HIV-1 entry into cells using fusion inhibitors. This is promising because the inhibitors can potentially target the numerous intermediates that are formed during viral entry into the cell [90]. HIV-1 infects cells using a multistep process. The process begins when gp160, an envelope glycoprotein, proteolytically cleaves into a surface subunit (gp120) and a transmembrane subunit (gp41) [91]. This is done by a cellular convertase [92]. In the next step, gp120 binds to the host cell receptor CD4, leading to a conformational change in gp120 and causing the extension of the V3 loop of gp120. This extension binds to chemokine receptor CCR5 or CXCR4 or both. The binding of the chemokine receptor to the V3 loop leads to the exposure of gp41. gp41 is divided into three domains: extracellular domain, transmembrane domain, and cytoplasm domain [93]. The extracellular domain can be further divided into four functional domains: fusion peptide, N-terminal heptad repeat (NHR or HR1), C-terminal heptad repeat (CHR or HR2), and a tryptophan rich region [93]. Both NHR and CHR contain a number of 4-3 heptad repeats, and these repeats generally form coiled-coil structures [94]. After gp41 has been exposed, the NHR and CHR come together to form a 6-helix bundle. Once the 6-helix bundle is formed, the viral and cell membranes fuse together and HIV has successfully infected the host cell.

3.1.1 Targets

There are two main targets of HIV-1 for fusion inhibition: gp41 and gp120. gp41 is targeted in order to prevent the formation of the 6-helix bundle, thereby preventing membrane fusion. gp120 is targeted in order to prevent the binding of the chemokine receptor. This prevents the exposure of gp41, thereby stopping membrane fusion. Targeting gp120 is more upstream in the infection process than targeting gp41.

A number of research groups have elucidated the 3D structure of gp41, which is necessary for any computational protein design. Weissenhorn et al. [95], Chan et al. [96], and Tan et al. [97] used X-ray crystallography to determine the structure of gp41, confirming that it forms a coiled coil structure and consists of a 6-helix bundle. Zhou et al. [98] have also reported the structure of gp41 in complex with a small molecule inhibitor, providing further information as to the precise location of the binding site of gp41. The small molecule inhibitor binds to the hydrophobic pocket of the NHR of gp41. This prevents the formation of the interface between the NHR and CHR. The identification of a hydrophobic pocket of the NHR of gp41 as a potential drug target is further supported by Chan et al. [99]. They show that the inhibitory activity of the peptides they examined depend upon their ability to bind to the hydrophobic, coiled-coil cavity. Sia et al. [100] elucidated the crystal structure of a complex of gp41 with a short constrained C-peptide, which also binds to the hydrophobic core of gp41 NHR. In addition, Balogh et al. [101] presented a novel method that gives explicit structural constraints on a small molecule inhibitor bound to the hydrophobic pocket of gp41 for NMR. The functional role of the coiled-coil domain of gp41 was explored by Wild et al. [102]. They compared the effects of amino acid substitutions in the coiled-coil domain on both structure and function and found a correlation between the destabilizing effects of the substitutions on the structure and virus entry.

The 3D structure of gp120, especially the V3 loop, has been more difficult to obtain compared to gp41. This is because of the high flexibility of the V3 loop, and the fact that gp120 needs to bind to CD4 in order to induce the conformational change that extends the V3 loop. Vranken et al. [103] determined the structure of the V3 loop using NMR. This provided 20 backbone structures, giving a flexible backbone template that can be used for computational protein design. Recently, the crystal structure of gp120 bound to CD4 and CCR5 was determined by Huang et al. [104]. This structure shows gp120 with the V3 loop extension and the chemokine receptor bound to the V3 loop. This provides the location of the binding site for CD4 and for the CCR5 chemokine receptor on the V3 loop. Tan and Rader [105] further analyzed the binding regions of gp120 using a flexibility analysis of all known gp120 structures. They identified two rigid regions, which may serve as stable targets for drug design.

Other studies have also attempted to determine conserved regions of gp120 and gp41, which can be targets for broadly neutralizing antibodies [106–111].

3.1.2 Inhibitors

There are two FDA (U. S. Food and Drug Administration) fusion inhibitors on the market. The first, Fuzeon (enfuvirtide), was developed by Roche and Trimeris and has been in use since 2003 [90, 112]. The second, Selzentry (maraviroc), was developed by Pfizer and approved in 2007 [113].

Fuzeon is used in patients that have developed a drug resistance to standard antiretroviral treatments. Its sequence is based on the sequence of gp41 CHR and targets gp41 NHR. Selzentry differs from Fuzeon in that it is a small molecule and targets gp120. Selzentry binds to the chemokine receptor CCR5. One issue with Selzentry is that it is only effective in patients with HIV using the CCR5 receptor. The drug has limited activity in patients with HIV using the CXCR4 receptor or both receptors.

In addition to the two FDA approved fusion inhibitors, there are many more that have been published in the open literature. Table 1 provides the names, sequences, IC₅₀ values, and references for a number of HIV fusion inhibitors.

Table 1.

HIV Fusion Inhibitors. Amino Acid letter “B” Denotes the Unnatural Amino Acid α- Aminoisobutyric Acid (Aib). Cross-Linked Amino Acids are Denoted by the ^ Character. Upper-Case Letters Indicate L-Amino Acids while Lower-Case Letters Indicate D-Amino Acids

Name	Sequence	Length	IC50(μM)	Refs.
(PIE7)3	PEG-(KGACDYPEWQWLCAA)3	45	0.00025	[114]
C34M2	WMEWDREINNYTSLIHS LIEESQNQQeKNEkELL	34	0.00031	[115]
Selzentry	Small molecule ligand	n/a	0.00040	[113]
C34	WMEWDREINNYTSLIHSLIEESQNQQEKNEQELL	34	0.0006	[116]
Fuzeon	YTSLIHSLIEESQNQQEKNEQELLELDKWASLWNWF	36	0.0037	[114]
Sifuvirtide	SWETWEREIENYTRQIYRILEESQEQQDRNERDLLE	36	0.00394	[117, 118]
P5	WMEWDREINNYTSLIHSL IEESQNQQEKNEQELLEL DKWASLWNWFNITNWLWYIK	56	0.06	[119]
PIE7	KGACDYPEWQWLCAA	15	0.62	[114]
C14linkmid	MTWQ^EWDREIQ^NYT	14	35	[100]
C14Aib	MTWBEWDREIBNYT	14	144	[100]
C14wt	MTWMEWDREINNYT	14	>500	[100]

A few of the peptides contain D-amino acids, which are resistant to proteolytic digestion, thereby increasing their half-life in vivo [114, 115, 120]. Recent work has also focused on finding peptidic inhibitors of enfuvirtide-resistant HIV strains [117, 121–124], which have developed due to the increased use of Fuzeon. The mechanisms of action of second-generation fusion inhibitors are being investigated [94] as well as the minimal sequence and minimal interface for fusion inhibitors [125, 126]. Liang [127] provides a review of inhibitors targeting CXCR4, one of the chemokine receptors that bind to gp120. In addition, various small molecules are being investigated to find new HIV-1 inhibitors that may have increased bioavailability and reduced cost of production over peptidic inhibitors [93, 101, 127–131].

3.1.3 Computational Methods

A number of computational methods have been used to design novel HIV-1 fusion inhibitors. Singh et al. [132] used homology modeling and docking studies to investigate the interactions between the chemokine receptor CXCR4 and flavonoids. They report that the compounds they examined may become important new antiviral drugs. Berchanski and Lapidot [133] designed new conjugates of neomycin with two arginine peptide chains to target CXCR4 using a multistep docking procedure [134]. Imai et al. [135] used their computer program ANTIS to design inhibitors that target CCR5. The Jiang group [136–139] used a virtual screening method based on computer modeling to identify gp41 small-molecule inhibitors. They screened 20,000 compounds and found that one of the compounds, ADS-J1, had an IC₅₀ in the low micromolar range. Recently, Bellows et al. [140] used their de novo design framework to discover novel peptide inhibitors of gp41. The peptides are 12 amino acids in length and the most promising one showed an IC₅₀ of 31μM in inhibition studies of HIV-1 entry into host cells.

In addition to targeting gp41 and gp120, computational methods have elucidated inhibitors that target the protein disulfide isomerase (PDI) of HIV-1 [141].

3.2 Purine Nucleoside Phosphorylase

Purine nucleoside phosphorylase (PNP) catalyzes purine metabolism, converting adenosine into adenine, inosine into hypoxanthine, and guanosine into guanine. It was discovered that people who are deficient in PNP have an immunodeficiency problem. Altered pathways of purine metabolism lead to T-cell deficiency. This is because deoxyguanosine is not properly metabolized and builds up in the blood. It is then transported and phosphorylated by T-cell deoxynucleoside kinases, forming harmful elevated levels of dGTP in the T-cells [142].

This lead to the realization that inhibition of PNP could regulate undesirable T-cell proliferation [143]. Diseases such as rheumatoid arthritis, psoriasis, inflammatory bowl disorders, and multiple sclerosis are caused by inappropriate activation of T-cells by self-antigens and could be targets for PNP inhibitors [142].

3.2.1 Targets

A number of groups have determined the crystal structure of PNP in complex with various ligands. Canduri et al. [144] presented a number of crystal structures of PNP in complex with ligands. They determined the structure of human PNP in complex with guanosine at 2.80 Å resolution, with 3'-deoxyguanosine at 2.86 Å resolution, and 8-azaguanine at 2.85 Å resolution. Canduri et al. [145] have also determined the crystal structure of human PNP in complex with inosine and 2',3'-dideoxyinosine at 2.8 Å resolution. Finally, Canduri et al. [146] reported the crystal structure of human PNP in complex with hypoxanthine and sulfate ion at 2.6 Å resolution. These crystal structures provide information about the location of the binding site of PNP. The ligands described in the aforementioned complexes are the natural ligands that bind during purine metabolism.

In contrast to the structures cited above, dos Santos et al. [147] have reported the crystallographic structure of human PNP in complex with acyclovir. Acyclovir is a potent inhibitor of replicant herpes simplex virus and also inhibits PNP, but with lower activity. This provides a structural template of an inhibitor bound to PNP. Silva et al. [148] described the kinetics and determined the crystal structure of human PNP in complex with a synthetic substrate, 7-methyl-6-thioguanosine.

Recent work has also been done to determine the active sites of PNP and the binding mechanisms of the ligands. Deng et al. [149] used NMR data to determine active site contacts in human PNP. Wielgus-Kutrowska et al. [150] studied the binding of 2-amino-9-[2-(phosphonome-thoxy) ethyl]-6-sulfanylpurine (PME-6-thio-Gua) to PNP from Cellulomanoas sp. They determined that the binding of the ligand to PNP is a one-step process.

Other groups have determined the structures of PNP from a number of other species. Caceres et al. [151] determined a first time model of PNP from Streptococcus agalactiae using molecular modeling and dynamics simulations. Schnick et al. [152] and Shi et al. [153] both report a crystal structure of PNP from Plasmodium falciparum in complex with various ligands and inhibitors. Others include models of PNP from calf spleen [154] and Listeria monocytogenes [155]. These models and structures may provide further insight to the binding site and mechanisms of human PNP.

3.2.2 Inhibitors

There are a number of inhibitors of PNP reported in the open literature, however all consist of small-molecule inhibitors. There are no peptidic inhibitors, making this an interesting challenge for de novo protein design. Because there are no peptidic inhibitors, there is no complex of PNP with a peptide, therefore no design template. One can, however, attempt to construct a template using the information of where and how the small-molecule inhibitors bind to PNP.

One of the earliest inhibitors of PNP was 9-(3,4-dioxopentyl)hypoxanthine [156]. It was found to be 25 times more effective in inhibiting PNP than butanedione. Inhibitors with equilibrium dissociation constants in the picomolar range were designed based on the transition-state structure of PNP [157]. More recent studies have found novel multi-substrate inhibitors [158], transition state analogue inhibitors [142, 159], and analogues of forodesine HCl as inhibitors [160]. Clinch et al. [161] developed two potent inhibitors of human PNP that are currently in clinical trials and which were used as templates for further inhibitor design. In addition, Todorova and Schwarz [162] have investigated the effect of the phosphate substrate on inhibitor binding to human PNP. They determined the thermodynamics of three inhibitors of PNP (acyclovir, ganciclovir, and 9-benzylguanine) as a function of phosphate ion concentration.

3.2.3 Computational Methods

A number of computational methods have been used to aid in the design and development of small-molecule inhibitors of PNP, but none have been used to develop peptidic inhibitors. DeWitte and Shakhnovich [163] developed a de novo design method called SMoG (Small Molecule Growth), which is used for lead design and is based on a simple model for protein-ligand interactions. They applied their method to PNP, comparing the IC₅₀ of known PNP ligands to the SMoG energy score. Shimada et al. [164] used SMoG to generate ligands complexed to actual protein structures and then used these models to construct databases from which knowledge-based protein-ligand potentials were derived. Macedo Timmers et al. [165] developed a set of empirical scoring functions that have the potential to evaluate binding affinities and docking results. These scoring functions were applied PNP in an effort to guide virtual screening. Another method applied to PNP relies on the use of drug-binding databases for virtual screening initiatives [166].

3.3 Ubiquitin Specific Protease 7

Ubiquitin specific protease 7 (USP7), also known as Herpesvirus-Associated USP (HAUSP) is a deubiquitinating enzyme that cleaves ubiquitin from its substrates. Ubiquitination is a prominent step in the regulation of many cellular processes. The major function of ubiquitination is to target proteins for degradation [167]. The catalytic core of USP7 consists of three domains: Fingers, Thumb, and Palm domains. In addition, the N-terminal domain of UPS7 consists of a TRAF-like MATH domain. The catalytic cleft is located between the Palm and the Thumb domains.

USP7 binds to p53, which contributes to its deubiquitination and stabilization. In addition, over expression of USP7 leads to p53-dependent cell growth repression and apoptosis. This indicates that USP7 may act as a tumor suppressor [168]. Furthermore, USP7 contains an additional binding site that can be targeted with inhibitors such as ubiquitin aldehyde (Ubal), a potent and irreversible inhibitor of many deubiquitinating enzymes, to regulate deubiquitinating activity [168]. This is important as deviations in the ubiquitin pathway can lead to a number of clinical disorders such as neurodegenerative disorders [169], genetic diseases [170], and immune surveillance/viral pathogenesis [170]. USP7 is also known to interact with a number of viruses, such as the Herpes simplex virus and the Epstein-Barr virus, making it a potential target for anti-viral treatment [168].

3.3.1 Targets

USP7 regulates p53 and is bound by at least two viral proteins (ICP0 from the Herpes simplex virus and EBNA1 from the Epstein-Barr virus). ICP0 binds to a domain between amino acids 599–801, while EBNA1 and p53 bind to a domain between amino acids 62–205 [171, 172]. Holowaty and Frappier [171] showed that binding of EBNA1 to USP7 disrupts the USP7-p53 interaction.

The first crystal structure of USP7 consists of the catalytic core (amino acids 205–560: Fingers, Thumb, and Palm domains) in isolation and in complex with ubiquitin aldehyde (Ubal) [167]. Interestingly, upon Ubal binding, USP7 undergoes a dramatic conformational change in the active site, with some loops moving 5 – 8 Å. The second binding site, located in the TRAF-like domain, binds p53 and MDM2 in a mutually exclusive manner, with MDM2 having the higher binding affinity [173]. Monoubiquitylation of the tumor suppressor p53 mediated by MDM2 promotes its mitochondrial apoptosis, however, the apoptotically active non-ubiquitylated p53 can also be generated via the p53-USP7 complex [174]. Hu et al. [173] determined the crystal structure of the TRAF-like domain of UPS7 in complex with p53 and the TRAF-like domain of USP7 in complex with MDM2. Furthermore, the crystal structure of the p53 binding domain has also been elucidated in isolation and bound to EBNA1 [175].

Conflicting studies show that binding of p53 to USP7 either promotes the deubiquitination and subsequent stabilization of p53 [176, 177] or that disruption of USP7 stabilizes p53 [178, 179]. It is thought that MDM2, rather than p53, is the preferred substrate of USP7, given by the higher binding affinity of MDM2 over p53 [173]. Hence, it may be that USP7 deubiquitinates multiple targets, MDM2 and p53 being two of them, and that the net deubiquitination of the targets determines the stability of p53. USP7 antagonists have the potential to stabilize p53 (which may be sufficient to cause tumor regression) in tumors that over express MDM2 by promoting the degradation of MDM2 [168].

The crystal structures of USP7 in isolation and bound to its substrates provide a number of possible templates for de novo protein design. We have recently verified these multiple binding sites by applying the binding site detection algorithm ConCavity [180]. ConCavity combines evolutionary sequence conservation estimates with structure-based surface pocket prediction in identifying binding sites. Three main binding sites were identified, the Ubal binding site, the p53/MDM2 binding site, and the catalytic cleft. This is consistent with the complexes that have been reported.

3.3.2 Inhibitors

Colland et al. [181] recently identified the first lead-like inhibitor of USP7. Using high-throughput screening, they discovered HBX 41,108, a small-molecule that inhibits USP7 deubiquitination. HBX 41,108 was shown to inhibit USP7 in an uncompetitive way, interacting with the enzyme-substrate complex rather than competing with substrate binding. HBX 41,108 has an IC₅₀ of 0.42 μM and treatment with the small molecule stabilized p53, activated the transcription of a p53 target gene, and inhibited cancer cell growth. It also induced p53-dependent apoptosis in cancer cells containing the wild-type p53.

It is interesting that the small-molecule inhibitor interacts with the enzyme-substrate complex, rather than with just the enzyme itself. Molecular docking studies were carried out by Colland et al. [181] in order to understand the structural basis of HBX 41,108 binding to USP7. HBX 41,108 was docked to the ubiquitin-bound form of USP7. The inhibitor interacts with Asp295, Val256, Phe283, Trp285, His294, Leu299, and Val302, which form a hydrophobic groove. A crystal or NMR structure of this enzyme-substrate-inhibitor complex confirming these interactions would provide a very good design template for protein design, allowing the design of peptidic inhibitors of USP7. However, peptidic inhibitors could also be designed to inhibit the interaction of the substrate with USP7, using one of the known crystal structures mentioned previously, and targeting the ubiquitin binding site or the TRAF-like MATH domain of USP7.

3.3.3 Computational Methods

While computational methods have not been utilized to design novel inhibitors of USP7, various in silico methods have been used to examine the MDM2-mediated p53 ubiquitination network [182], develop a pharmacophore model for inhibitors of ubiquitin isopeptidases, and identify USP7 as a tumor necrosis factor receptor-associated factor domain-containing protein [183].

3.4 Histone Demethylases

Demethylases are a recently identified class of proteins that reverse Lys and Arg methylation. The first histone demethylase was identified by Shi et al. [184]. Histones are part of chromatin and aid in the packaging of DNA into the chromosomes [185]. DNA winds around histones like spools, forming nucleosomes [186]. Histones affect gene regulation by chemical modifications along the N-terminal histone tail that protrudes from the histone core around which the DNA is wound. These chemical modifications consist of phosphorylation, acetylation, methylation, adenoribosylation, ubiquitylation, and suoylation. While phosphorylation and acetylation have long been known to be reversible processes (mediated by kinases and phosphatases for phosphorylation and acetyltransferases and deacetylases for actylation), methylation was long thought to be an irreversible process. This changed with the discovery of two histone demethylases: lysine-specific demethylase-1 (LSD1) [184] and JmjC domain-containing demethylase-1 (JHDM1) [187]. These two enzymes are capable of demethylating histone tails, making histone methylation a dynamic process, such as phosphorylation and acetylation. Since then, other demethylases have been identified, expanding the class of JmjC domain-containing demethylases to include JHDM3A/JMJD2A, GASC1/JMJDC, and JHDM2A [188]. LSD1 is limited to the demethylation of dimethylated and monomethylated lysine. In contrast, JmjC domain-containing demethylases have been shown to demthylate trimethylated lysine [189–191].

Since histone methylation is linked to a number of important biological processes (stem cell maintenance and differentiation, X inactivation, and DNA damage response [185, 192]), it is expected that histone demethylation is also involved in these processes. Demethylases can act as transcriptional repressors due to the fact that methylation of histone H3 activates gene expression. This implies that inhibitors of demethylases can be used as therapeutic treatments for diseases that involve improper changes in gene expression, such as cancer [193, 194].

3.4.1 Targets

A number of crystal and solution structures of LSD1 have been determined. The earliest is a solution structure of the SWIRM domain of LSD1 [195], which is comprised of 6 compact α helices. Stavropoulos et al. [186] presented the crystal structure of LSD1 at 2.9 Å resolution, finding that LSD1 has a highly symmetric, closely packed domain structure, from which a long helical `tower' domain extends. Yang et al. [196], Yang et al. [197], and Forneris et al. [198] determined the crystal structure of LSD1 in complex with CoREST, a co-repressor that, when bound to LSD1, enables demethylation of histone H3 Lys4 (H3-K4) within nucleosomes. Other crystal structures of LSD1 include the structure of LSD1 at a resolution at 2.8 Å [199], LSD1 bound to an H3 peptide [200], and LSD1 in complex with tranylcypromine, a known monoamine oxidase inhibitor [201].

The most prominent member of the JmjC domain-containing demethylase family is JMJD2A. Chen et al. [202] first reported the crystal structure of JMJD2A with and without α-ketoglutarate in the presence of Fe(II). The structure of the core domain consists of the JmjN domain, the JmjC domain, the C-terminal domain, and a zinc-finger motif. Following this, Chen et al. [203] determined the structures of the catalytic core of JMJD2A in complex with methylated H3-K36 peptide substrates in the presence of Fe(II). In addition, structures of JMJD2A-Ni(II)-Zn(II) bound to tri-, di- and monomethyl forms of H3-K9 and the di- and trimethyl forms of H3-K36 have been reported [204, 205]. JMJD2A has also been found to bind another histone, H4. Lee et al. [206] presented the crystal structures JMJD2A in complex with the trimethyl form of H3-K4 and JMJD2A in complex with the trimethyl form of H4-K20, showing that the two histones adopt radically different binding modes. Other JmjC domain-containing demethylase structures include the DNA binding domain of the H3-K4 demethylase RBP2 [207], the DNA binding domain of JARID1C [208], and solution structure of JARID1A [209].

Other groups have described the binding of the LSD1/CoREST1/HDAC co-repressor complex to SUMO-2/3 [210], the binding of HIF-1 α to JMJD1A and JMJD2B [211], and the use of LSD1 as a potential therapeutic target [212, 213].

3.4.2 Inhibitors

A recent review by Mai and Altucci [214] focuses on the chemical aspect and potential application in cancer therapy of molecules that inhibit not only histone demethylases but also DNA methyltransferases, histone deacetylases, histone acetyltransferases, and histone methyltransferases. Culhane et al. [215] identified two peptidic inhibitors of LSD1. The first peptide (1) exhibited time-dependent inhibition of LSD1. While the second peptide (2) did not, it was still moderately potent against LSD1 with an IC₅₀ of 15.6 ± 1.7 μM. Table 2 provides the sequences of the two peptides. Schmidt and McCafferty [188] have studied monoamine oxidase inhibitors (MAOIs) as potential inhibitors of LSD1. They showed that trans-2-phenylcyclopropylamine (2-PCPA) is the most potent MAOI and that it is a time-dependent, mechanism-based irreversible inhibitor of LSD1 with an IC₅₀ of 20.7 ± 2.1 μM. In addition, 2-PCPA has been used as template to develop other mechanism-based inactivators of LSD1 [216].

Table 2.

Histone Demethylase Inhibitors. The IC₅₀ of 1 was Not Available. Peptides 1 and 2 are Modified at the N-Terminus. See Culhane et al. [215] for Details

Name	Sequence	Length	IC50 (μM)	Ref.
6a	small molecule ligand	n/a	1.4	[217]
7	small molecule ligand	n/a	6.6	[217]
1	QTARKSTGGKAPRKQLA	17		[194, 215]
2	QTARKSTGGKAPRKQLA	17	15.6	[215]
2-PCPA	small molecule ligand	n/a	20.7	[188, 197]
8	small molecule ligand	n/a	27	[217]
5a	small molecule ligand	n/a	28	[217]
1a	small molecule ligand	n/a	78	[217]
1c	small molecule ligand	n/a	100	[217]
3	small molecule ligand	n/a	540	[217]

Rose et al. [217] identified inhibitor scaffolds for JMJD2 histone demethylases. The best inhibitor (6a) had an IC₅₀ of 1.4 μM. The seven highest-ranked inhibitors based upon enzyme-inhibitor interactions from this study (6a, 5a, 7, 3, 8, 1c, 1a) are shown in Table 2. Recently, Hamada et al. [218] designed JHDM inhibitors based upon the crystal structure of JMJD2A in complex with NGO and histone trimethylated lysine peptide. They tested four compounds and all showed inhibitory activity against JMJD2 proteins. Other known inhibitors of JHDM include NGO, succinic acid, and 2,4-lutidinic acid.

3.4.3 Computational Methods

In the literature examined to date, the design of demethylase inhibitors has been done experimentally, rather than computationally. New small-molecule inhibitors were designed based upon scaffolds of known demethylase inhibitors. There are many known crystal and solution structures of demethylases in complex with histones and various inhibitors, which can be used as templates for computational protein design. Even though there are no structures of peptidic inhibitors, the knowledge that can be gleaned from the small-molecule complexes can provide a starting point for computational design.

4 SUMMARY

A number of computational protein design methods were described and recent advances outlined. These methods included both deterministic (such as DEE, SCMF, and quadratic assignment-like models) and stochastic (such as Monte Carlo, genetic algorithms, and combinatorial library approaches) methods, and showed varying incorporation of backbone flexibility (from rigid backbones and discrete rotamers, to continuous backbones with continuous rotamer angles). A number of successes using the various computational methods were highlighted. Reviews on protein design by Floudas [1] and Floudas et al. [2] cite a number of successes in the area of protein design, from the design of active sites on α-lytic protease to improve specificity toward the substrate [219] to the design of inhibitors for complement component C3 [5, 6]. Despite these successes, there still remain many challenges for protein design. These include improved methodologies, such as improved sequence selection using flexible-backbone templates and simultaneous sequence and structure selection with flexible templates, and new design applications, such as design of inhibitors or peptides, among others.

Four application domains for computational de novo protein design were described, outlining potential drug targets. Peptides can be designed to inhibit HIV-1, PNP, USP7, and histone demethylases. To date, there have been a number of successes with peptidic inhibitors of HIV-1 and histone demethylases, however inhibitors of PNP and USP7 consist only of small-molecule inhibitors. All four systems have great promise for computational protein design, using the information of both peptidic and small-molecule inhibitors bound to the target protein as design templates.