In Silico Engineering of Synthetic Binding Proteins from Random Amino Acid Sequences

Summary

Synthetic proteins with high affinity and selectivity for a protein target can be used as research tools, biomarkers, and pharmacological agents, but few methods exist to design such proteins de novo. To this end, the In-Silico Protein Synthesizer (InSiPS) was developed to design synthetic binding proteins (SBPs) that bind pre-determined targets while minimizing off-target interactions. InSiPS is a genetic algorithm that refines a pool of random sequences over hundreds of generations of mutation and selection to produce SBPs with pre-specified binding characteristics. As a proof of concept, we design SBPs against three yeast proteins and demonstrate binding and functional inhibition of two of three targets in vivo. Peptide SPOT arrays confirm binding sites, and a permutation array demonstrates target specificity. Our foundational approach will support the field of de novo design of small binding polypeptide motifs and has robust applicability while offering potential advantages over the limited number of techniques currently available.

Graphical Abstract

Highlights

• •InSiPS engineers synthetic binding proteins (SBPs) using primary protein sequence
• •SBPs are designed to a bind a target protein and avoid “off-target” interactions
• •Binding and functional inhibition of two of three target proteins in yeast is demonstrated
• •Our new approach offers advantages over alternative tools that rely on 3D models

Biological Sciences; Bioinformatics; Protein Family Determination

Introduction

Proteins are diverse macromolecules that form intricate and complex protein-protein interaction (PPI) networks through selective affinity binding. These properties have driven an expansion in the field of protein and peptide design over the past decade (Kang and Saven, 2007, Pantazes et al., 2011, Saven, 2010, Yu et al., 2014). Specifically, the ability to design synthetic proteins that can bind, label, or inhibit a specified target with high affinity are of primary importance and have the potential to replace antibodies and chemical compounds in a wide range of applications. Current methods to develop engineered binding proteins include peptide aptamer selection (Colombo et al., 2015), directed evolution of display systems (Goldflam and Ullman, 2015), and computational methods, the majority of which modify naturally occurring protein folds rather than designing novel structures ab initio (Benjamin Stranges and Kuhlman, 2013, Karanicolas and Kuhlman, 2009, Mikut et al., 2016, Schreiber and Fleishman, 2013).

Computational protein design (CPD) can allow for the in silico evaluation of amino acid sequences on a scale that goes beyond the constraints of many laboratory approaches (Chica et al., 2005). Natural proteins represent only an infinitesimal portion of potential functional sequences, limiting the scope of most current CPD techniques (Woolfson et al., 2015). Many protein targets lie beyond the reach of natural protein folds or current approaches to developing binding peptides, and searching randomized sequences space has been shown to successfully yield novel functional binding proteins (Cherkasov et al., 2009, Devlin et al., 1990). It is thought that true large-scale de novo protein design can expand beyond the confines of biologically derived molecules into the vast space of “never-born proteins” (Li et al., 2013, Luisi et al., 2006). This unexplored sequence potential coupled with the fact that many protein-based therapeutics have been shown to be effective and well-tolerated in clinical trials (Craik et al., 2013, Otvos and Wade, 2014) have made peptides a quickly expanding category of US Food and Drug Administration (FDA)-approved drugs over the past 20 years (Fosgerau and Hoffmann, 2015, Kaspar and Reichert, 2013).

In addition to peptide therapeutics, CPD has been used in recent years for developing ligand-binding proteins (Tinberg et al., 2013), nanobiotechnology (Wilson, 2015), de novo enzyme design, and the development of antibody mimetics (Lao et al., 2014). Much of the recent focus has been on developing proteins to replace targeted antibody therapies (Huang et al., 2013, Takeuchi et al., 2014). Rationally designed synthetic proteins with high affinity or specificity for a chosen target may become an important alternative to antibody-based biological drugs, which experience numerous limitations including ineffective pharmacokinetics, a relatively large size, immunological complications, ethical questions, and high production costs. Computational tools that can effectively design novel proteins specifically architected to interact with a wide range of targets are now beginning to emerge (Chevalier et al., 2017, Viart et al., 2016).

We present a powerful massively parallel computational tool that designs high-affinity binding proteins for a given target. This tool is the first of its kind as it employs a unique genetic algorithm, actively minimizes off-target interactions during the design process, and does not employ docking models or require information on the 3D structure of the target. The In-Silico Protein Synthesizer (InSiPS) algorithm begins with a pool of random amino acid sequences and, over many generations of fitness-based selection followed by mutation and crossover events, converges on sequences that are predicted to interact with a specified target and minimize interactions with non-targets (other proteins in the environment) (Figure 1). InSiPS uses the co-occurrence of small interacting motif pairs (Pitre et al., 2012, Schoenrock et al., 2014) to predict PPIs and intelligently design proteins with desired interaction profiles. In this way, previously “undruggable” proteins (Ostrem and Shokat, 2016), and those that lack a well-recognized binding pocket, may be targeted by this method.

Figure 1.

An Overview of the InSiPS Genetic Algorithm

(A) An initial pool of random protein sequences 150 aa in length is created. Next, the primary loop is entered: sequences in the pool are evaluated, and subsequent generations are created. This process repeats for a minimum of 250 generations until a high-fitness peptide is produced.

(B) PPI prediction. Sequences generated by InSiPS are evaluated using the Protein-Protein Interaction Prediction Engine (PIPE) (Pitre et al., 2008). PIPE requires a validated global PPI network as input. Step 1: protein A is compared with all proteins in the known PPI network. A sliding window is used both on A and the proteins in the PPI network until some segment of A, starting at position i, matches a segment of some protein T in the network. All known interactors of T (neighbors) are put into a list to be used in the next step. Step 2: protein B is compared with the proteins in the neighbor's list in the same manner. When a segment of B, starting at position j, is found to match a segment of a protein from this list, the result matrix is incremented at position (i,j). This matrix represents all the segments in proteins A and B that co-occur in experimentally validated PPIs and is used to predict if A and B interact. The interaction algorithm assigns a predicted interaction score between 0 and 1. Any pair scoring over 0.51 is predicted to interact with a specificity of 99.5%. The fitness of a protein sequence is calculated based on predicted interactions with targets-non-targets.

(C) Generating the next generation of candidate sequences. First, a copy, mutate, or crossover operation is randomly chosen with a preset probability proportional to the fitness of a sequence as calculated in (B). This process is repeated until the next generation is complete. The algorithm is terminated after a minimum of 250 generations when the fitness score does not improve over 50 consecutive generations.

The InSiPS algorithm evaluates the predicted affinity and specificity (fitness) of hundreds of thousands of sequences over hundreds of generations, meaning upward of a billion predictions are made in a single run. This scale is difficult to achieve when using competing methods that rely on detailed 3D protein configuration data (Lewis and Kuhlman, 2011) and are thus limited by the computational restraints of working with docking models (Pierce et al., 2014).

Other CPD methods have employed sequence-based approaches to design proteins. For example, Fisher et al. (2011) utilized binary patterning of alternating polar and non-polar residues to yield biologically functional proteins in Escherichia coli (Fisher et al., 2011). Keating and others have developed CLEVER and CLASSY, a method of cluster expansion that maps a complex function of atomic 3D coordinates from structure-based models of protein energetics to more simple linear functions of sequence. This change dramatically speeds up scoring and has been used to successfully design highly specific synthetic protein ligands against multiple basic-region leucine zipper transcription factor families (Grigoryan et al., 2009, Negron and Keating, 2013). However, InSiPS differs from these methods and other sequence-based approaches as it uses sequential optimization of binding via a genetic algorithm without any predetermined pattern or use of structural considerations, instead relying on conserved short linear binding motifs.

Conserved short linear motifs are known to mediate PPIs in a manner that is unique from the more classically accepted interactions between large, rigid domain structures (Chica et al., 2009, Chica et al., 2005, Davey et al., 2012). The more flexible linear motifs are ubiquitous across higher eukaryotes and are proposed to be capable of re-wiring PPI networks through the loss or gain of these functional modules (Neduva and Russell, 2006). Our algorithm uses primary protein sequences and experimentally validated interaction networks to screen for the co-occurrence of such motifs common to known protein pairs. This technique has been used to accurately predict global PPI networks in a variety of organisms including Saccharomyces cerevisiae, Caenorhabditis elegans, Mus musculus, and humans (with a precision of 82.1%) (Pitre et al., 2012, Schoenrock et al., 2014).

As a proof of concept, we aimed to design synthetic binding proteins (SBPs) that could functionally inhibit non-essential endogenous proteins in the yeast Saccharomyces cerevisiae. For ease of experimentation, 18 initial targets were chosen that fit our desired criteria of localizing to the cytoplasm, being of moderate size (<1,500 amino acids [aa]) and relatively steady abundance (500–5,000 molecules per cell) in addition to possessing readily observable phenotypes when the encoding gene is deleted or the protein product is functionally inhibited. Of 18 targets, 3 with varying fitness scores were selected for wet-laboratory experimentation: (1) Psk1, a serine/threonine kinase, which plays a role in regulating sugar metabolism; (2) Pin4, a protein involved in G2-M phase progression following DNA damage; and (3) Rmd1, a protein involved in meiotic nuclear division. The engineered SBPs were specifically designed to avoid interactions with all other yeast cytoplasmic proteins (designated as non-targets). A length of 150 aa was chosen for this project to ensure that the majority of predicted interaction motifs are included within the SBP sequence. Only a subsection of this 150-aa polypeptide, which also contains a 6xHIS tag for affinity purification, is likely required for binding. We evaluated the ability of anti-Psk1, anti-Pin4, and anti-Rmd1 to bind or inhibit their respective targets using a series of phenotypic assays and binding experiments. See Tables S2 and S3 for a complete list of InSiPS results for all target proteins considered in this study.

Our results validate the ability of this approach to computationally engineer unique SBPs. Because InSiPS does not begin with a template, but rather a pool of random sequences, the algorithm has the potential to search sequence space beyond biological barriers and is not constrained to naturally occurring sequences. Moreover, because of efficient parallelization of the algorithm, hundreds of thousands of predictions can be made during each “generation” of the genetic algorithm, allowing strong selective pressure to be applied to maximize binding affinity and specificity. Ultimately, InSiPS-engineered proteins may be useful for research, for biotechnology, or as pharmacological agents.

Results

InSiPS Designs SBPs against Three Target Proteins

A preliminary run of InSiPS was used to evaluate initial SBP designs against 18 yeast targets. Our genetic algorithm assigns each candidate protein sequence a fitness (interaction) score between 0 and 1, balancing affinity for the target with specificity (i.e., reduced binding to non-targets; see Transparent Methods for more details). In all cases, InSiPS was able to design SBPs with substantially stronger predicted affinity for the designated target than the highest likely “non-target” protein. In addition, all SBPs produced very low average non-target scores (Table 1), highlighting the selectivity of SBPs, and showed limited sequence homology to known yeast proteins (Figure S1).

Table 1.

InSiPS Predictions Suggest High Affinity of SBPs for Target Proteins and No Predicted Off-target Interactions

Synthetic Binding Protein	Fitness	Target Score	Max Non-Target Score	Max Non-Target	Average Non-Target Score	Closest Yeast Homolog to SBP
Anti-Psk1	0.465	0.718	0.352	Ubi4p	0.072	Mmp1p
Anti-Pin4	0.380	0.630	0.398	Cdc39p	0.0797	Esl1p
Anti-Rmd1	0.344	0.563	0.389	Sec14p	0.132	YAP1801p

Higher interaction scores indicate an increased likelihood of interaction, and a score of >0.51 would be deemed likely to interact at 99.5% specificity. The fitness function weighs scores between the SBP and both targets and off-targets {Fitness(SBP) = [1 - MAX(score(Non-targets)) x score(Target)}. A higher fitness function indicates an increased likelihood of a protein having specificity for the target. The target score represents the score between the SBP and the target protein. The max non-target is score between the SBP and the next most likely off-target interaction (no interactions predicted). The average non-target score is the averaged score between the SBP and all proteins localized to the cytoplasm. The highest scoring non-target protein and yeast protein most homologous to the SBP are also listed (see Figure S1 for alignment and Table S1 for sequences of synthetic proteins).

Of the three designed SBPs selected for wet-laboratory experimentation, anti-Psk1 demonstrated the highest fitness (0.465) followed by anti-Pin4 (0.380) and anti-Rmd1 (0.344). All three anti-target proteins showed relatively strong interaction scores against their respective targets (Psk1, 0.718; Pin4, 0.630; Rmd1, 0.563). In all cases, the scores of all the non-target proteins are below the threshold of 0.51 at which the algorithm would predict an interaction, meaning no off-target interactions are predicted. InSiPS appears to work better for some targets than others. Anti-Psk1 exceeds the other two anti-target proteins in terms of maximizing target score and minimizing of the max non-target score. InSiPS was able to effectively design binding proteins for all targets considered (see Tables S2 and S3), but only three targets were selected for wet-laboratory experimentation.

To evaluate the novelty of the anti-target proteins, the sequences were compared against the yeast proteome using BlastP. The results show limited sequence similarity to yeast proteins (Figure S1). Anti-Psk1 most closely resembles the known Psk1 interactor Mmp1 and aligns with 34% coverage and a maximum 52% identity over 38 aa. Anti-Pin4 has sequence homology to a single yeast protein, Esl1, a known interactor of Pin4 with 29% sequence identity over a region representing only 3% of the total protein sequence. Anti-Rmd1 was found to have significant sequence homology with two proteins, Sec14 with a maximum 52% sequence identity over 29 aa, and Pmt5 with a maximum 28% sequence identity. However, neither of these proteins is known to interact with Rmd1.

Anti-Psk1 and Anti-Pin4 Show Functional Inhibition of Targets In Vivo

We hypothesized that the expression of our anti-target proteins may inhibit the function of the target proteins if biologically significant binding occurs in vivo. To this end, we expressed the SBPs in S. cerevisiae and performed three assays that examined conditional viability or growth rate and effects on protein distribution (Figure 2).

Figure 2.

Strains Expressing Anti-Psk1 and Anti-Pin4 Can Phenocopy Deletion Mutants of the Target Proteins and Alter Target Protein Expression or Assembly

(A and B) Viability of cells under strain-specific stress condition shows that anti-Psk1 and anti-Pin4 expression can produce phenotypes that resemble loss of function mutants of the target proteins. (A and B) Average normalized colony-forming unit (CFU) counts from four trials are displayed as mean ± SD. Stress conditions in trial were (A) exposure to UV light for 30 s for the anti-Psk1 trial and (B) exposure to cycloheximide (65 ng/mL) for the Pin4 trial.

(C and D) Expression of anti-target SBPs produces growth defects under strain-specific stress conditions that resemble deletion of the target. Three replicates of each culture or condition were grown for 12 h in liquid YPG (Yeast extract/Peptone/Galactose) + drug, media and OD600 was measured hourly. This experiment was repeated three times. Error bars represent SD among replicates, and a polynomial line of best fit is presented. (C) Δpsk1 sensitivity to H₂O₂ resembles the phenotype of strains expressing anti-Psk1. Cells were grown in media containing 0.75 mM H₂O₂. (D) Δpin4 sensitivity to 13 μM hygromycin resembles the phenotype of strains expressing anti-Pin4.

(E and F) Observed alteration of fluorescence profile of GFP-tagged targets when anti-target proteins are expressed. Aliquots of WT + anti-target protein cells from the same culture were used to inoculate complete media with either 4% galactose (where anti-target SBP is expressed) or 4% glucose (where anti-target SBP is repressed), and overall fluorescent signal from three independent cultures for each condition were measured over time and normalized to the growth rate. See also Figure S2 showing that anti-Pin4 can cause sensitivity to arsenite.

To test if the anti-Psk1 SBP can functionally inhibit Psk1, we induced oxidative stress and compared viability and growth rate to wild-type (WT) and Δpsk1 strains as the loss of PSK1 is known to increase sensitivity to UV light and oxidizing agents (Hanway et al., 2002). Anti-Psk1 expression phenocopies the Δpsk1 mutant, consistent with inhibition of protein function (Figure 2). As seen in Figures 2A and 2C, Δpsk1 is sensitive to UV irradiation and exposure to H₂O₂, phenotypes that were also seen when anti-Psk1 is expressed. UV exposure decreased viability in Δpsk1 by 85% and by 83% when anti-Psk1 is expressed. To determine if the sensitivity observed was the result of the anti-Psk1 protein and not other factors, we expressed anti-Psk1 in the Δpsk1 strain and observed no significant alteration to viability. Growth curve analysis showed that exposure to H₂O₂ decreased the growth rate of cells expressing anti-Psk1 relative to WT cells, which strongly resembles the Δpsk1 phenotype (Figure 2C). In addition, expression of anti-Psk1 resulted in a significant change (net –decrease) in the fluorescent signal of GFP-tagged-Psk1 protein, suggesting possible aggregation or degradation of the target (Figure 2E).

We performed the same assays as above to test for any functional inhibition of Pin4 by the anti-Pin4 protein. The deletion of PIN4 is known to cause yeast cells to become sensitive to inhibitors of protein synthesis such as cycloheximide and hygromycin B (Brown et al., 2006). Culturing in the presence of 65 ng/mL cycloheximide decreased viability in Δpin4 by 67%, WT+ anti-Pin4 by 35%, and Δpin4+ anti-Pin4 by 34% (Figure 2B). Growth curves show similar sensitivity of WT+ anti-Pin4 and Δpin4 to hygromycin B (Figure 2D). These results show that anti-Pin4 sensitizes cells to translational inhibitors, suggesting functional inhibition of Pin4. In addition, cells expressing anti-Pin4 become sensitive to arsenite in a manner similar to those expressing Δpin4 (Figure S2). The expression of anti-Pin4 decreased the net fluorescent signal of GFP-tagged Pin4. These results together suggest partial functional inhibition of target protein function.

The third target selected for experimentation, Rmd1, was analyzed to detect if phenotypic changes can be produced through expression of the anti-Rmd1 protein. However, no significant alteration to conditional viability or sensitivity to β-mercaptoethanol or L-1,4-dithiothreitol exposure similar to Δrmd1 was observed and no significant change to the fluorescent profile of GFP-tagged Rmd1 was detected when anti-Rmd1 was expressed. For these reasons, we chose not to experiment further using the anti-Rmd1 peptide as the aim of this project was to demonstrate functional inhibition through binding. This observation suggests that not all designed anti-target peptides are functionally effective.

Yeast Two-Hybrid Analysis Indicates Binary Interactions between Target/Anti-target Proteins

Together, the results displayed in Figure 2 suggest that anti-Psk1 and anti-Pin4 possibly bind to and alter the endogenous functionality of their respective targets. To confirm that binary PPIs between Psk1/anti-Psk1 and Pin4/anti-Pin4 occur in vivo, we performed a series of yeast two-hybrid (Y2H) assays. Three reporter genes were present in our Y2H strain, which are all induced by Gal4 reconstitution, and three independent reporter assays were employed to test for interactions (see Transparent Methods for details).

In this way, the reconstitution of GAL4 by a physical interaction between the bait (target) and prey (anti-target) will induce growth on minimal media lacking uracil, activate lacZ activity, and provide resistance to 3-aminotriazole. All three reporter assays showed binding signals indicating physical interactions between the two target-anti-target combinations in vivo (Figure 3). Figure 3C indicates that binding affinity between anti-Pin4/Pin4 may be lower than the anti-Psk1/Psk1 affinity as this β-galactosidase assay is the most quantifiable of the three Y2H assays employed.

Figure 3.

Yeast Two-Hybrid Assay Indicates Physical Interactions between Target and Anti-target Proteins

(A and B) (A) Positive Y2H results using uracil reporter assay. A growth curve in minimal media lacking uracil shows that Pin4/anti-Pin4 and Psk1/anti-Psk1 bait-prey combinations grow better than the negative control strain, indicating a PPI between bait and prey proteins through expression of the URA3 reporter. Triplicate trials produced similar positive results, but the results from a single trial are shown. (B) Positive Y2H result for Pin4/anti-Pin4 and Psk1/anti-Psk1 bait-prey combinations based on resistance to 3-aminotriazole (3-AT). Normalized colony-forming unit (CFU) counts for triplicate trials on minimal media lacking histidine +25 mM 3-AT resistance in test bait-prey combinations and in the positive control are presented as mean ± SD.

(C) Positive Y2H result for Pin4/anti-Pin4 and Psk1/anti-Psk1 bait-prey combinations using a β-galactosidase reporter. Miller units are used to quantify β-galactosidase activity by measuring the hydrolysis of ortho-Nitrophenyl-β-galactoside (ONPG) spectrophotometrically. Relative β-gal activity (fold change) from triplicate trials is shown relative to negative control ± SD. The Psk1/anti-Psk1 interaction produced a stronger signal than Pin4/anti-Pin4.

Peptide SPOT Array Analysis Shows that Binding on Targets Occurs at Predicated Loci

Positive results in our Y2H assays further support binding between target and anti-target proteins in a biological system. To probe these interactions in vitro and to further evaluate target interaction sites, a walking peptide SPOT array was used (Figure 4) (Jia et al., 2005). Because InSiPS predicts regions on both the target and anti-target proteins responsible for binding, it provides a starting point to probe the binding regions. Using InSiPS-predicted interaction sites, we designed walking peptide arrays that probed the predicted interaction site of the target and flanking regions using 18-aa motifs shifting at single amino acid intervals. As seen in Figures 4A and 4B, very specific residues on both interacting partners are proposed to facilitate binding (dark green regions). Interactions between both anti-Psk1/Psk1 and anti-Pin4/Pin4 were shown to occur within or directly adjacent to Protein-Protein Interaction Prediction Engine-predicted interaction regions (Figure 4). This further supports the biological validity of our CPD method for engineering binding proteins and demonstrates the specificity of the engineered proteins.

Figure 4.

Walking Peptide SPOT Arrays Indicate Specific Binding Regions

SPOT arrays containing 18-aa-long printed peptides corresponding to subsequences from within the predicted interaction regions of target protein at single amino acids intervals.

(A and B) Predicted interaction matrices highlight the predicted interaction regions between target (x axis) and anti-target (y axis). (A) The anti-Psk1/Psk1 interaction site was predicted to occur between residues 1209–1246 of the PSK1 protein. (B) The Pin4/anti-Pin4 interaction site was predicted to occur between residues 472–506 on Pin4.

(C and D) SPOT arrays of predicted target binding sites and flanking regions probed with 6xHis-tagged anti-target proteins followed by detection using an anti-His antibody. (C) Specific binding of the anti-Psk1 protein to the target was detected between amino acids 1204–1228 (D) Specific binding of the anti-Pin4 protein to the target was detected between amino acids 436–458.

(E and F) Relative binding of SPOT array peptides indicates highly specific binding regions with highest relative binding. See also Figure S3 for analysis of the anti-Psk1 predicted interaction site.

A very specific site was predicted to mediate the Psk1/anti-Psk1 interaction. This interaction region spans residues 1209–1247 in Psk1 and residues 47–69 (22 amino acids long) on the anti-Psk1 designed protein. Peptides spanning the Psk1 target region and flanking residues were probed with purified 6xHis-tagged anti-Psk1. As expected our observations using a walking array indicate an overlap between the predicted interaction site and the site identified by the walking array.

Interestingly, the predicted interacting region on anti-Psk1 residues is in an area that does not have significant sequence homology to Mmp1 (Figures S1 and S3), the protein in the yeast proteome with the greatest sequence homology to anti-Psk1. Here we see how InSiPS demonstrates its ability to focus on specific small interacting motifs that facilitate direct protein binding.

Walking SPOT array analysis of the predicted Pin4 interaction region indicated a single motif that facilitates the observed binding between Pin4 and anti-Pin4, which again corresponded closely to the region predicted by InSiPS. A single region spanning residues 440–466 demonstrated binding affinity for the anti-target protein in the walking array (Figure 4D). This further supports the premise that InSiPS can successfully engineer proteins that bind through short interaction motifs.

Specific binding of both anti-Psk1 and anti-Pin4 synthetic proteins to short subsequences within the target protein was observed in the walking peptide SPOT array (Figures 4C and 4D). We further studied the interaction between Psk1/anti-Psk1 as this peptide demonstrated the highest fitness score (Table 1), strongest Y2H signal (Figure 3C), and greatest functional inhibition of the target (Figures 1A and 1B).

The walking array results showed that the highest binding intensity occurred between anti-Psk1 and a truncated version of Psk1 spanning residues 1207–1224 (Figure 4C). We probed this interacting region using a permutation array to determine which residues are the most essential for binding by making single amino acid substitutions at all positions and monitoring the effect on binding affinity.

Permutation Array Shows High Specificity of Subsequence in the Psk1-Binding Motif

Our permutation analysis showed that amino acids 1218–1224 on Psk1 appear to be the most essential to the interaction (Figure 5C). This region that has three identical residues with the highest scoring predicted off-target interactor Ubi4, 1219K, 1221L, and 1223D (Figures 5A and 5D). However, the residues that have the greatest influence on the affinity between Psk1/anti-Psk1 binding are two histidines at 1220H and 1222H such that any substitution at these loci abolishes the interaction (Figure 5C). Notably, on Ubi4, the off-target protein predicted to be the most likely to interact with anti-Psk1, the corresponding residues are glutamine (Q) and glutamate (E). Importantly, both single-amino-acid substitutions, H→Q and H→E, demonstrated significantly decreased affinity for anti-Psk1 protein suggesting relatively lower affinity for the Ubi4 (Figure 5D). The specificity of our peptides is shown by the limited number of acceptable substitutions at these two key loci. Despite the sequence homology to this region on Ubi4, key residues 1222H and 1224H are expected to limit binding by anti-Psk1.

Figure 5.

Characterization of Psk1 Interaction Motif and Binding Affinity

(A) Permutation array based on relative binding of the anti-Psk1 binding peptide to the Psk1 interaction motif (1207–1224). Peptides that correspond to the WT Psk1 sequence are outlined in white. Spot coloring is based on the relative binding intensity for each permutated position.

(B) Binding curve of anti-Psk1 with the Psk1-binding motif. Shown in the diagram are equilibrium isotherms for the Psk1(1207–1224) peptide from fluorescent polarization.

(C) Anti-Psk1 recognition motif based on a position-specific scoring matrix created from the permutation array.

(D) Sequence homology between the anti-Psk1-binding site of Psk1 and closest homolog Ubi4. Histidine residues at positions 14 and 16 of the anti-Psk1-binding motif are not conserved in the Ubi4 protein sequence.

To gauge the affinity of the anti-Psk1 protein against the predicted Psk1-binding site, fluorescent peptides corresponding to the Psk1 1207–1224 amino acid sequence were synthesized and purified for binding studies in solution. As shown in Figure 5B, binding of the Psk1 (1207–1224) peptide to anti-Psk1 assumed saturable patterns with a Kd value of 2.2 nM. Together these results demonstrate the ability of InSiPS to engineer a small synthetic binding peptide with limited sequence homology to natural proteins that can bind with high affinity to the designated targets.

Discussion

We have designed and validated a unique computational tool, InSiPS, for engineering proteins that bind specified protein targets with high affinity. InSiPS offers multiple strengths over current CPD tools because of its unique methodology. By functioning as a genetic algorithm and analyzing thousands of candidate sequences over hundreds of generations, InSiPS can engineer high-confidence binding proteins, which limits the need for costly and time-intensive speculative laboratory trials using classical enrichment techniques. Because the algorithm analyzes PPIs using patterns of primary amino acid sequences, it is not limited to working on proteins with known high-resolution 3D structures or domains (Zhang et al., 2013) and is free of the requirement for known 3D pockets. In this way it expands our ability to tackle diverse protein targets. InSiPS also has the unique advantage of actively avoiding interactions with thousands of off-target proteins. This robust algorithm has broad applicability and could theoretically be used to target various proteins. These features, coupled with the wide variability of potential future applications (for example, in the areas of biomarkers and pharmacological agent development), make InSiPS a powerful CPD tool.

The efficacy of the InSiPS algorithm is demonstrated here through the successful engineering, production, and biological analysis of two SBPs, anti-Psk1 and anti-Pin4, which successfully bind to respective targets Psk1 and Pin4. Each binding protein was the manifestation of hundreds of generations of the genetic algorithm and upward of a billion PPI predictions. Both SBPs had limited sequence homology to endogenous yeast proteins (Figures S1 and S2) and demonstrated high fitness scores (0.465, 0.380), predicting strong affinity for their target proteins and low affinity for all other proteins localized to the yeast cytoplasm. Consecutive biological assays verified the ability of these proteins to bind their targets both in vivo and in vitro and inhibit their natural biological functions.

Initial experiments indicated that the expression of anti-target proteins anti-Psk1 and anti-Pin-4 could functionally inhibit target proteins and produce phenotypes that resembled Δpsk1 and Δpin4, respectively (Figures 2A–2D). Y2H analysis indicated binary interactions between target and anti-target proteins in vivo (Figure 3), and a peptide SPOT array was probed in vitro to identify specific binding motifs on Psk1 and Pin4 responsible for mediating interactions (Figure 4). These results suggest that binary interactions are occurring between anti-Psk1/Psk1 and anti-Pin4/Pin4 within predicted target regions, but further study is needed to understand binding dynamics.

The anti-Psk1/Psk1 interaction was further probed as it exhibited the highest fitness score (Table 1), binding affinity (Figure 5), Y2H signal (Figure 3C), and functional inhibition of the target (Figures 1A and 1B). Permutation analysis showed that a single amino acid change to the target binding site can prevent binding of anti-target proteins (Figure 5D). This demonstrates the ability of the genetic algorithm to “evolve” proteins that possess significant binding specificity by searching beyond natural sequence configurations. Fluorescence polarization of the Psk1-binding motif and anti-Psk1 protein demonstrated a saturable pattern with a Kd value of 2.2 nM (Figure 5B). The strong affinity of the anti-Psk1 protein for its target coupled with the specificity observed in the permutation analysis lends further support to the ability of the InSiPS algorithm to efficiently design protein sequences with desired interaction properties.

The current study furthers the emerging field of de novo binding protein or peptide design, which strives to explore beyond natural protein sequence space and create functional high-specificity proteins that often are not found in nature. The majority of previous approaches to protein engineering have involved intelligent manipulation of naturally occurring proteins, but CPD is now quickly entering a new era of de novo design. Most major advances in de novo CPD over the past few years have focused on mastering protein folding and structure prediction and using these principles to design simple novel protein structures (Huang et al., 2016). For the most part, these newly designed proteins are structural with limited functionality and have simply served to lay the groundwork for future endeavors. Fundamental protein structures such as barrels (Huang et al., 2015, Thomson et al., 2014), helical bundles (Huang et al., 2014), protein nanomaterials (King et al., 2012), and oligomers (Boyken et al., 2016) have been developed and may eventually be used for a variety of future applications as we delve further into sequence space and take advantage of the scalability and specificity of polypeptides. However, at this point, very few novel functional polypeptides have been developed computationally de novo.

Proteins or peptides with short highly specific binding motifs, like the ones developed in this study, could be developed to function as aptamers, research tools, biomarkers, pharmacological agents, or more. Peptide aptamers continue to be developed for a variety of industrial (Colombo et al., 2015) and medicinal (Hanley-Bowdoin and Lopez-Ochoa, 2015) applications, but the field of using proteins as pharmacological agents has expanded significantly over the past 5–7 years. Peptide therapeutics can be tailored to maximize compatibility, stability, and potency with relative ease, and there are currently over 60 available FDA-approved peptide drugs with another 500+ progressing through the development stage (Kaspar and Reichert, 2013, Rafferty et al., 2016).

Other potential applications of this methodology include the development of biosensors for previously unreachable biomarkers or the intelligent design of peptide aptamers. Certain small protein targets such as the medically relevant angiotensinogenase renin have proven difficult to probe with traditional approaches but can be detected using peptide-based aptamers developed using cDNA display techniques (Biyani et al., 2016). Peptides designed to function as therapeutics can outperform relatively larger (∼150 kDa) antibodies in terms of bioavailability, tumor penetration, and production efficiency and have been used to develop binding assays, cancer therapy, drug delivery, and in vivo imaging (Yu et al., 2017). Peptides, which were previously thought of as ineffective pharmacologically due to poor delivery mechanisms and rapid degradation or clearance (Otvos and Wade, 2014), are changing pre-conceptions as synthetic peptide production decreases in cost, new delivery systems are developed, and biological production using vectors such as viruses or genetic manipulation via CRISPR-Cas9 become more realistic possibilities.

To summarize, we have presented a unique CPD tool that employs a massively parallel genetic algorithm and PPI prediction tool to engineer binding peptides against protein targets. We demonstrated that two synthetic proteins engineered by InSiPS can bind endogenous yeast proteins at predicted motifs and inhibit functionality. Further work will examine if the technique can work in other species and explore the range of potential targets. We invite those interested in using InSiPS to contact the corresponding author. By combining constructive techniques such as InSiPS with modeling technologies, future techniques will aim to develop highly stable, specific binding proteins for a range of applications.

Limitations of the Study

Our method has furthered the field of de novo computational design of short binding polypeptides and offers promising preliminary findings, but more work is required to expand the reach and efficiency of the method. The current framework restricts InSiPS to working on annotated proteins within known PPI networks. The technology may be applicable in other systems, but this has not yet been shown. The algorithm also does not directly consider protein stability but combines aspects of known functional motifs to confer bioactivity. Because InSiPS functions in the realm of flexible linear motifs, limited structural considerations are required. Future work will examine incorporating additional structural prediction tools to predict stability and folding patterns.

Another limitation may be the applicability of our approach to tackle different targets. In the current study, we started with three potential targets. In our first attempt, we generated high-scoring anti-peptides with limited off-target interactions for two of three targets. As each round of computation begins with a random pool of sequences, further iterations of InSiPS may identify suitable anti-peptides for Rmd1. However, with low sample numbers it remains difficult to speculate about the broad applicability of our approach to different proteins. InSiPS may also be constrained if the desired target is a member of a protein family with highly similar sequences or shares significant sequence similarity with other proteins in the cell. Last, the specificity of these proteins remains unclear despite being engineered to avoid off-target interactions. Only indirect and predicted evidence that our anti-target proteins avoid off-target interactions is provided in this article. We did not observe significant changes in observable phenotypes when the target protein is deleted and the anti-target protein is expressed (Figures 2A and 2B) and show specificity through the binding observed in the SPOT arrays (Figures 4 and 5).

Methods

All methods can be found in the accompanying Transparent Methods supplemental file.

Published: January 25, 2019

Supplemental Information

Data S1. Raw Data for GFP-tagged Target Fluorescence Assay, Related to Figure 2

Raw data for the normalized fluorescence experiment presented in Figures 2E and 2F. The cell density (OD₆₀₀) and average fluorescent unit (AFU) reading values for three individual cultures is shown for each condition. In addition, the mean and standard deviation are provided along with the normalized fluorescence values that are presented in Figure 2.