Applications Of Computationally Designed Proteins


Computational design of ligand-binding sites is a particularly useful development for building new biosensors. Biosensors transduce microscopic binding events into macroscopically observable signals. Traditionally they have been developed by identifying a naturally occurring protein that binds the target analyte, discovering a suitable macroscopic signal, and building a detector that reports the concentration of the detected analyte to the user. While this strategy has produced many useful biosensors, each one is unique and requires substantial development time and optimization. Clearly, it would be useful to develop a modular system in which the binding site and the signaling site can be changed separately without destroying the communication between them, so that the development of new biosensors would require redesign of only one component—the analyte binding site. Many interesting platforms have been suggested (Hellinga and Marvin 1998), but periplasmic binding proteins, introduced previously in this chapter, have been the most thoroughly explored.

With the aim of creating a modular biosensor platform, E. coli maltose-binding protein was converted into a fluorescent maltose sensor by attaching a small molecule fluorophore to a region that is allosterically coupled to the ligand-binding site (Marvin et al. 1997). The maltose-binding site was redesigned to bind Zn2+ while retaining the fluorescent response to ligand binding, indicating that, in theory, new biosensors could be created in a modular fashion by altering the binding specificity of the binding pocket (Marvin and Hellinga 2001). [Surprisingly, it was later determined by x-ray crystallography and small-angle x-ray scattering that the designed Zn2+ binding site does not seem to form in its intended geometry (Telmer and Shilton 2005)]. Building on that work, ribose-binding protein was successfully converted into a zinc receptor with a secondary coordination sphere included in the initial designs (Dwyer et al. 2003). Once the method of stereochemistry-based design was available, ribose-binding protein (and other members of the periplasmic binding protein family) were designed into sensors for many other molecules (described previously in this chapter) (Allert et al. 2004; Looger et al. 2003). The redesigned ribose-binding proteins (zinc-BP and TNT-BP) were expressed in a bacterial strain containing a chimeric two-component signal transduction pathway connecting ligand binding in RBP to transcriptional activation of a reporter gene (| -galactosidase). This resulted in unique ligand-mediated synthetic signal transduction pathways. In theory, if the reporter gene were green fluorescent protein (GFP) or luciferase, then the engineered bacteria would glow in response to the analyte, making a truly biological biosensor.

Therapeutic proteins and Antibodies Cytokine Design

Now that the field of computational design of protein-protein interactions is becoming more established, reports of its application to therapeutically relevant proteins are surfacing. One example of computational design in a therapeutic setting is the inactivation of TNF-a signaling by rationally designed dominant-negative TNF-a variants (Steed et al. 2003) using Xencor's Protein Design Automation (PDA) software (Filikov et al. 2002; Hayes et al. 2002), which is based on the work of Dahiyat and Mayo (Dahiyat and Mayo 1997). TNF-a is a homotrimeric cytokine that activates the inflammation pathway upon binding TNF receptor-1 (TNF-R1), and is implicated in a wide range of diseases including rheumatoid arthritis and inflammatory bowel syndrome (Chen and Goeddel 2002). Analysis of a homology model of the cytokine-receptor complex indicates that there is significant steric separation between residues of TNF-a that form the homotrimeric interface and those that form the binding interface with TNF-1R. The PDA software predicted one or two nonim-munogenic mutations that disrupt interactions with TNF-1R while preserving the structural integrity of the TNF variants and their ability to homotrimerize (and ideally heterotrimerize with endogenous TNF). In preclinical trials its ability to block soluble TNF activity in cell-based assays and two mouse arthritis models was established (Zalevsky et al. 2007). While this is a good example of computational design in a therapeutic setting, and is advantageous over other TNF-a inhibitors, destroying one interaction while preserving another distal one is not terribly challenging, and perhaps could have been achieved by inspection alone.

More impressive is the redesign of the TNF-related apoptosis inducing ligand (TRAIL). TRAIL is a potential anticancer drug that selectively induces apoptosis in cancer cells by interacting with the death receptors DR4 and DR5 (Jin et al. 2004; Kelley and Ashkenazi 2004) and is in Phase II trials for non-Hodgkin's lymphoma and non-small cell lung cancer (Genentech). It also binds decoy receptors that do not induce apoptosis (DcR1 and DcR2). Quax and colleagues computationally evaluated the effect of a single amino-acid mutation at 34 different positions of TRAIL on its interaction energy with each of the four receptors (Van Der Sloot et al. 2006). Seven mutations at five positions were predicted to confer selectivity for DR5. Of the various combinations of these mutations, one variant (with two mutations) showed no measurable affinity for DR4 or Dc1, and 20-fold reduced affinity for Dc2. In apoptosis assays, cells expressing DR5 were susceptible to killing by the mutant TRAIL. DR5-selective TRAIL variants have also been identified by phage display (Kelley et al. 2005); by using negative selection to remove variants that bind the other receptors

(DR4, Dc1, and Dc2), researchers at Genentech were able to isolate DR5-selective variants (which have more mutations than the computationally designed variants) that also induced apoptosis. Interestingly, the two solutions sets, computational and diversity-based, are nonoverlapping.

Antibody Affinity Maturation

A common application of diversity-based engineering is affinity maturation of antibodies by phage display (Sidhu and Koide 2007), ribosome display (Lipovsek and Pluckthun 2004), and yeast display (Gai and Wittrup 2007) (reviewed in Marvin and Lowman 2005). Although these methods have proven robust, they are limited in the maximum genetic diversity they can search; a library of 6 x 1023 Fab variants, even if present as single protein molecules, would weigh 50 kg! In silico, there are no such limitations. This has enabled computational design, with improved modeling of water and electrostatic interactions (Levy et al. 2003), to improve the affinity of antibodies beyond the achievements of diversity-oriented engineering (Lippow et al. 2007). The affinity of cetuximab (Erbitux®, ImClone Systems) for EGFR was improved 10-fold from 0.5 nM to 0.05 nM by mutations that modify the electrostatic interactions on the periphery of the binding interface; affinity of D44.1 (a common test antibody) for lysozyme was improved 100-fold from 4.4 nM to 0.04 nM. A similar approach was used to improve the affinity of an anti-VLA1 antibody from 10 nM to 1 nM (Clark et al. 2006). As computational design becomes more widely available, and its potential to create more efficacious antibodies is explored, it will eventually be adopted by companies that focus on therapeutic proteins.

surpassing the Technical Limitation of the Lab

The true value of computational design is its ability to surpass the technical limitations of the various diversity-oriented design approaches. Perhaps the best example is the design of antibody Fc variants with enhanced effector function (Lazar et al. 2006), for which no diversity-oriented approach is viable. Antibody-dependent cellmediated cytotoxicity (ADCC), a key effector function that determines the clinical efficacy of some monoclonal antibodies, is mediated primarily through a set of closely related Fcy receptors with both activating and inhibitory activities (excellently reviewed in Desjarlais et al. 2007). It has been shown by standard site-directed mutagenesis (Shields et al. 2001) and glycoform engineering (Shields et al. 2002) that the affinity of interaction between Fc and certain FcyRs correlates with cytotoxicity in cell-based assays. Of particular significance, higher affinity for FcyRIIIa (CD16) correlates with higher ADCC. Phage display, a powerful tool for antibody affinity maturation, could in principle be used to identify mutations that increase FcyRIIIa affinity, were it not for the critical role of the ^-linked carbohydrate at N297. Deglycosylation of intact IgG decreases its affinity for FcyRIIIa by 20-fold, and deglycosylation of an Fc fragment decreases its affinity to below detectable levels (Radaev et al. 2001). Because E. coli bacteria lack the molecular machinery necessary for proper glycosylation of a displayed Fc fragment, and because full length IgGs are too large to be displayed, the only method for increasing the affinity of an IgG for FcyRIIIa used to date has been trial-and-error site-directed mutagenesis (Shields et al. 2001).

Lazar and coworkers at Xencor overcame this technical hurdle by using the crystal structure of the Fc/FcyRIIIa complex (Radaev et al. 2001) and their PDA software to predict mutations that would provide more favorable interactions at the Fc/ FcyRIIIa interface (Lazar et al. 2006). Hedging their bets, they made and tested hundreds of combinations of those variants in high-throughput binding assays. They identified variants with over 100-fold enhancement of in vitro effector function, and this resulted in increased cytotoxicity in an in vivo preclinical model. This approach to optimization of FcR binding has enabled Xencor to form partnerships with MedImmune, Genentech, Boehringer Ingelheim, and Human Genome Sciences to incorporate (presumably) these mutations into their antibody-based products.

0 0

Post a comment