Thermodynamic Additivity of Sequence Variations: An Algorithm for Creating High Affinity Peptides Without Large Libraries or Structural Information

Background There is a significant need for affinity reagents with high target affinity/specificity that can be developed rapidly and inexpensively. Existing affinity reagent development approaches, including protein mutagenesis, directed evolution, and fragment-based design utilize large libraries and/or require structural information thereby adding time and expense. Until now, no systematic approach to affinity reagent development existed that could produce nanomolar affinity from small chemically synthesized peptide libraries without the aid of structural information. Methodology/Principal Findings Based on the principle of additivity, we have developed an algorithm for generating high affinity peptide ligands. In this algorithm, point-variations in a lead sequence are screened and combined in a systematic manner to achieve additive binding energies. To demonstrate this approach, low-affinity lead peptides for multiple protein targets were identified from sparse random sequence space and optimized to high affinity in just two chemical steps. In one example, a TNF-α binding peptide with Kd = 90 nM and high target specificity was generated. The changes in binding energy associated with each variation were generally additive upon combining variations, validating the basis of the algorithm. Interestingly, cooperativity between point-variations was not observed, and in a few specific cases, combinations were less than energetically additive. Conclusions/Significance By using this additivity algorithm, peptide ligands with high affinity for protein targets were generated. With this algorithm, one of the highest affinity TNF-α binding peptides reported to date was produced. Most importantly, high affinity was achieved from small, chemically-synthesized libraries without the need for structural information at any time during the process. This is significantly different than protein mutagenesis, directed evolution, or fragment-based design approaches, which rely on large libraries and/or structural guidance. With this algorithm, high affinity/specificity peptide ligands can be developed rapidly, inexpensively, and in an entirely chemical manner.

. Molecular dynamics simulation of TNF1 and TNF1-opt peptide structure. Molecular dynamics (MD) simulations were performed to elucidate potential structure or structural tendencies in TNF1 and the effect of variations on possible conformations. In these simulations, 2600 sampled conformations from a total of 1 µs of MD trajectories were generated for both TNF1 and for TNF1-opt. These conformations were clustered by backbone structural alignment within 1 Å pair-wise RMSD. The fraction of the total number of conformations for the ten largest clusters is shown in the bar graph. Representative backbone conformations for the varied region of the peptide (residues 4-11) from each of the top ten clusters are shown above the graph, with the N-terminal end at the top and the structures ordered from cluster 1-10, left-to-right. Based on an analysis of the distribution of conformations, both peptides may be best characterized as loosely structured, with three main characteristics: 1) Both peptides have a tendency to form a loose and fluid hairpin, with the exact locus of the turn shifting among various positions in the region of residues 9-14. This is consistent with a negative band observed at 234 nm in their circular dichroism (CD) spectra [1,2,3] (Figures S2, S3); 2) The altered region of TNF1-opt, residues 4 through 11, substantially favored an extended conformation (though by no means rigid) in both TNF1 and TNF1-opt; and 3) apart from the foregoing, the structures of both peptides were quite flexible and variable overall.
While MD simulations suggest less rigidity in the TNF1-opt altered region, these simulations along with CD spectroscopy suggest that any tendency towards forming a hairpin present in TNF1 is retained in TNF1-opt. Similar structural tendencies in TNF1 and TNF1-opt imply that the four variations in TNF1-opt are not significantly structurally connected and therefore do not dramatically alter any structure or structural tendencies present in the lead peptide, supporting the idea that relatively unstructured peptides serve as good scaffolds for affinity optimization via thermodynamic additivity. Figure S2. TNF1 CD spectrum and Dichroweb predicted secondary structure. Figure S3. TNF1-opt CD spectrum and Dichroweb predicted secondary structure. Figure S4. TNF1 lead peptide SPR response against a panel of five proteins. (TNF-α, Transferrin, Ubiquitin, AKT1 and Neutravidin). Relative SPR responses from the TNF1 lead peptide against immobilized TNF-α, Transferrin and Ubiquitin are shown. All measurements were taken on the same SPR chip with five different proteins immobilized on separate addressable spots on the SPR chip. Sensorgrams are normalized to the predicted maximum SPR response (R max ) for the specific protein molecular weight and immobilization level. TNF1 SPR responses against immobilized AKT1 and Neutravidin were also measured on the same chip as the three proteins shown. TNF1 showed no measurable binding response against AKT1 or Neutravidin, with AKT1 immobilized on an active spot and Neutravidin immobilized on the reference spot on the SPR chip. TNF1 peptide concentration was held constant at 50 µM for all measurements.   On-rates (k on ), off-rates (k off ) and dissociation constants (K d ) shown were determined from kinetic fits of several sensorgrams across the same concentration series used in equilibrium binding response analysis.   Silver-stained gel image of the final eluted fraction from TNF-α pull-down assays performed with immobilized TNF4 and TNF4-opt in the presence of excess E.coli lysate (~2 mg/mL). Purified 10 µM TNF-α and the TNF-α spiked E.coli lysate used in the pull-down assay are also shown. Several bands appear on the gel independent of the immobilized peptide and are indicated with an asterisk (*), these bands are attributed to background bead binding. After subtracting background bands, one additional off-target band at approximately 65 kDa that appears in the TNF4-opt pulldown eluted fraction is noted. (Bottom) The band near 17 kDa in the SDS-PAGE image was validated as TNF-α with a Western blot. Figure S11. TRF26 point-variant library heat map. Point-variant peptides were immobilized on a gold SPR surface in an array format. The gold surface was blocked with BSA to reduce non-specific binding, and 10 µM transferrin was flowed over the SPR array. Fold-change color map corresponds to the equilibrium binding response of a particular point-variant relative to the TRF26 lead peptide.

D4S
D4S S11K M7K D4S, S11K D4S, M7K a The relative binding free energy differences observed between these variant peptides agrees well with the predicted difference calculated by assuming additivity and using the component contributions reported in Table 1 of the main text.

SPR Assays. SPR Chip Preparation.
Unless otherwise noted, SPR immobilization was done on a Biacore A100 SPR system (GE Healthcare, Piscataway, NJ). A standard coupling protocol was employed to immobilize a neutravidin capture surface via exposed primary amines. The immobilization was performed at 25 °C using Sodium Acetate Similarly, neutravidin was coupled to spots 2 and 3 via a 12 and 16 min. injection cycle respectively, followed by quenching. An additional 7 min injection of ethylene diamine was then pulsed over spots 1-3 to ensure no activated sites remained. This immobilization process was then repeated to couple neutravidin to the carboxy terminus on addressable spots 4 and 5 to give a range of neutravidin immobilization levels across all 5 spots in each of the 4 flow cells.

TNF1 and TNF4 Lead Generation from a 10 4 Random Peptide Screen. Peptide Library and SPR Chip Preparation.
Random 20-mer peptides were prepared in 384-well plates and diluted with Biacore HBS-EP buffer (GE Healthcare, Piscataway, NJ) containing 1 mg/ml carboxymethyl-dextran (Sigma-Aldrich, St. Louis, MO) to reduce non-specific binding to the CM-5 SPR chip surface. The final peptide concentration was 50 µM in each well. Peptides were then screened on a Biacore A100 SPR instrument (GE Healthcare, Piscataway, NJ). A CM-5 SPR chip (GE Healthcare, Piscataway, NJ) was prepared as described above with biotinylated Transferrin, TNF-α, Ubiquitin and AKT1 captured on spots 1, 2, 4 and 5 respectively on each of the four flow cells, Neutravidin was directly immobilized on spot 3 and used as a reference on each of the four flow cells.
Binding Specificity on a Microarray. TNF1 and corresponding variants were spotted on aminopropyltriethoxysilane glass slides (made in-house) functionalized with Sulfo-SMCC (Thermo Fisher Scientific, Rockford, IL), leaving a reactive maleimide group free for covalent attachment via the C-terminal cysteine in the peptides. Peptides were diluted to 1 mg/ml in PBS and robotically spotted in triplicate in a humidity chamber and allowed to react overnight. Spotted slides were first washed with 90% TFA for 15 minutes, then rinsed with nanopure water and dried in a centrifuge. Spotted slides were then washed with 100 mM NaBH 4 in 2X SSC buffer containing 0.05% SDS for 15 minutes, rinsed with 1X SSC then PBST and finally with nanopure water and dried in a centrifuge. Microarrays were preblocked with 3% BSA (wt:wt) and 1 mM mercaptohexanol (to quench remaining maleimides) in PBST buffer for 2 hours in a humidity chamber then briefly rinsed with PBST. 200 nM TNF-α and 200 nM E.coli lysate (calculated as a total protein concentration in the lysate) labeled with Alexa-555 and Alexa-647 (Invitrogen, Carlsbad, CA) respectively in PBST buffer containing 1% BSA was incubated with the microarray for 1 hour in a humidity chamber. Dye-swap experiments were done in an identical manner with the exception of TNF-α and E.coli lysate being labeled with Alexa-647 and Alexa-555 respectively. After binding, the microarray was washed 3 times with PBST, 15 minutes each wash, and then 3 times with nanopure water, 15 minutes each wash. Finally the microarray was dried in a centrifuge and scanned with an Agilent (Santa Clara, CA) microarray scanner at 10 µm resolution and PMT settings adjusted to give similar dynamic range of intensities across both channels. Image analysis was done with Genepix (Molecular Devices, Sunnyvale, CA) and resulting data processed in Matlab (Natick, MA).
Binding Specificity Pull-Down Assays. 400 µL of SulfoLink® Coupling Resin slurry, equivalent to 200 µL of settled beads, (Thermo Scientific, Rockford, IL) was placed in a centrifuge column and washed 5x with coupling buffer (0.1 M phosphate buffer pH 8 containing 1 mM EDTA). Then, TNF1, TNF1-opt, TNF4 or TNF4-opt peptides dissolved in coupling buffer were immobilized on the resin by incubating 15 nmol of the corresponding peptide (200 µL of 75 µM solution) at room temperature for 1 hour. The resin was washed 3x with coupling buffer and the unreacted sites quenched with 400 µL of 100 mM mercaptohexanol in coupling buffer at room temperature for 45 min. After washing the column 5x with 1M NaCl in 1x HBSN buffer, and 5x with 1x HBSN-T (1x HBSN containing 0.05% Tween-20), the resin is ready for the pull-down assay.
The peptide-bound resin was incubated with 100 µL of 10 µM TNF-α (PeproTech, Rocky Hill, NJ) in E. coli lysate (~2 mg/mL) for 1 hour at room temperature. The resin was washed consecutively, 5x with 1x HBS-NT, 3x with 1M NaCl in 1x HBS-N and 1x with 1x HBS-N. Finally, TNF-α was eluted by using 0.2 M NaOH at pH 9.0 and 100 µL fractions were collected. 75 µL of the first three fractions were combined and concentrated by centrifugation using a 3 kDa filter cut-off. Samples were run on a 4-12% SDS-PAGE gel (NuPAGE® Novex® Bis-Tris 4-12%, Invitrogen, Carlsbad, CA) for 10 min at 100 V and for 35 min more at 200 V and imaged using a Pierce Silver Stain Kit (Thermo Scientific, Rockford, IL). TNF-α was detected by Western Blot with an anti-TNF-α antibody (Invitrogen, Carlsbad, CD) and imaged on a Typhoon Trio (GE Healthcare, Piscataway, NJ).
TRF26 point-variant screen. The TRF26 lead peptide and point-variants were ordered in PEPScreen® 96-vial format from Sigma-Genosys and prepared as described above. From the 96-well stock plates, peptides were printed in microarray format on a bare gold Biacore FlexChip SPR (GE Healthcare, Piscataway, NJ) surface using a PerkinElmer (Waltham, MA) microarray spotter. Covalent attachment of the peptides was achieved through the Cterminal cysteine sulfhydryl group with the gold surface. After printing, the gold surface was treated with 1 mM β-mercaptoethanol to coat the remaining free gold surface with a monolayer of hydroxyl groups. β-mercaptoethanol treatment was done with the SPR chip in the Biacore FlexChip SPR system (GE Healthcare, Piscataway, NJ), allowing for real-time monitoring of monolayer attachment progress and completion. After monolayer attachment, the surface was washed extensively within the FlexChip system and 10 µM unlabelled transferrin in PBST buffer was flowed over the array surface. Binding response to each peptide spot was monitored and recorded over a 2 minute injection of transferrin. This experiment was repeated on a freshly prepared SPR array with the following additional step. After attachment of the β-mercaptoethanol monolayer, the gold surface was treated for 2 minutes with 1% Bovine Serum Albumin (BSA) in PBS to block non-specific binding. After BSA blocking, the surface was washed extensively and 10 µM unlabelled transferrin in PBST (Sigma-Aldrich, St. Louis, MO) was flowed over the SPR array for 2 minutes. The results from the both the unblocked and BSA blocked binding assays were comparable in that the point-variants with the highest binding responses were identical, however numerous peptides that showed moderate to low binding response in the unblocked assay showed dramatically reduced binding in the BSA blocked assay indicating that these moderate to low responses were likely non-specific.
TRF26 enhanced point-variant validation and characterization. TRF26 enhanced pointvariants identified in the point-variant screen were synthesized and purified using standard solid-phase FMOC peptide synthesis and HPLC purification. Purified peptides were used for all validation and affinity characterization assays. Validation and affinity characterization of TRF26 point-variants was done using a Biacore T100 SPR instrument (GE Healthcare, Piscataway, NJ) and Biacore CM-5 SPR chips. CM-5 chip flow cells 1, 2, 3 and 4 were activated with EDC/NHS as described in the main article methods. After activation, 25 µg/ml transferrin in Acetate pH 4.5 buffer was flowed across flow cells 2 and 4 for 1000 seconds at 10 µl/min. Transferrin immobilization was adjusted to produce a predicted R max ~ 100 RU. After transferrin immobilization, remaining active NHS groups were quenched with ethylene diamine on flow cells 1, 2, 3, and 4, with flow cells 1 and 3 serving as reference cells. Affinity characterization was done with a concentration series of unlabelled peptide flowed over the immobilized transferrin in HBS-P (GE Healthcare, Piscataway, NJ) buffer containing 1 mg/ml carboxymethyl-dextran (Sigma-Aldrich, St. Louis, MO) to inhibit non-specific binding to the CM-5 surface. Reference subtracted equilibrium binding response was recorded after a minimum 60-second association. Surface regeneration was done with Biacore Glycine 2.5 regeneration solution (GE Healthcare, Piscataway, NJ).

Molecular Dynamics Simulations.
For each sequence, 100 molecular dynamics trajectories, each of 10 ns in length, were generated using AMBER v.9 [5]. Each trajectory was begun from a conformation generated by assigning random values to all rotatable bonds, then randomly rotating bonds to eliminate any steric collisions, then minimizing. Trajectories were run using a 2 fs time step, with bonds to hydrogens constrained with SHAKE [6], AmberParm96 force field parameters, and the GB/SA implicit solvent model, with parameter settings SALTCON = 0.15, SURFTEN = 0.003, and EXTDIEL = 75 to simulate the salt, surfactant, and organic content of the SPR running buffer used for affinity measurements. Temperature for all runs was maintained at 300K via the Andersen thermostat [7] applied at 4 ps intervals. Conformations were sampled at 200 ps intervals after discarding the first 5 ns of each trajectory, yielding a total of 2600 samples for each sequence. A 2600 x 2600 pairwise distance matrix was computed reflecting average RMS distances following structural alignment of the backbone atoms of residues 4 through 11, as computed for each pair of conformations using Pymol's [8] fit function. Clustering was performed by repeatedly identifying the largest subset of samples having RMS distances within a 1 Å threshold, and removing the cluster so identified from the distance matrix. The graphical representations in Figure S1 were produced using Pymol.
TNF1 and TNF1-opt Circular Dichroism Spectroscopy. All circular dichroism (CD) spectroscopy was done using a JASCO J-815 CD (Easton, MD). TNF1 and TNF1-opt peptides were diluted to 10 µM in HBS buffer with 0.015% P-20 detergent. A blank 190-300 nm CD spectrum was taken before all measurements with HBS 0.015% P-20 buffer; this blank spectrum was subtracted from all subsequent peptide spectra. CD spectra for both TNF1 and TNF1-opt between 190-300 nm were taken at 25 °C.