Received: 6 June 2022 Revised: 14 August 2022 Accepted: 19 August 2022 DOI: 10.1002/pro.4429 F U L L - L E NG TH PA P E R Molecular determinants of TRAF6 binding specificity suggest that native interaction partners are not optimized for affinity Jackson C. Halpin1 | Dustin Whitney1 | Federica Rigoldi1 | Venkat Sivaraman1 | Avinoam Singer1 | Amy E. Keating1,2,3 1MIT Department of Biology, Cambridge, Massachusetts, USA Abstract 2MIT Department of Biological TRAF6 is an adaptor protein involved in signaling pathways that are essential Engineering, Cambridge, for development and the immune system. It participates in many protein– Massachusetts, USA protein interactions, some of which are mediated by the C-terminal MATH 3Koch Institute for Integrative Cancer domain, which binds to short peptide segments containing the motif PxExx Research, Cambridge, Massachusetts, USA [FYWHDE], where x is any amino acid. Blocking MATH domain interactions is associated with favorable effects in various disease models. To better define Correspondence Amy E. Keating, MIT Department of TRAF6 MATH domain binding preferences, we screened a combinatorial Biology, 77 Massachusetts Ave., library using bacterial cell-surface peptide display. We identified 236 of the Cambridge, MA 02139, USA. best TRAF6-interacting peptides and a set of 1,200 peptides that match the Email: keating@mit.edu sequence PxE but do not bind TRAF6 MATH. The peptides that were most Funding information enriched in the screen bound TRAF6 tighter than previously measured native National Cancer Institute, Grant/Award peptides. To better understand the structural basis for TRAF6 interaction pref- Number: P30-CA14051; Koch Institute Support (core); National Institutes of erences, we built all-atom structural models of the MATH domain in complex Health; National Institute of General with high-affinity binders and nonbinders identified in the screen. We identi- Medical Sciences, Grant/Award Numbers: fied favorable interactions for motif features in binders as well as negative 5R01GM129007, F32GM114959, F32GM137510, T32GM007287 design elements distributed across the motif that can disfavor or preclude bind- ing. Searching the human proteome revealed that the most biologically rele- Review Editor: Aitziber L. Cortajarena vant TRAF6 motif matches occupy a different sequence space from the best hits discovered in combinatorial library screening, suggesting that native inter- actions are not optimized for affinity. Our experimentally determined binding preferences and structural models support the design of peptide-based interac- tion inhibitors with higher affinities than endogenous TRAF6 ligands. KEYWORD S binding specificity, inhibitor, molecular modeling, peptide library, short linear motif, TRAF6 Jackson C. Halpin, Dustin Whitney and Federica Rigoldi contributed equally to this study. This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited. © 2022 The Authors. Protein Science published by Wiley Periodicals LLC on behalf of The Protein Society. Protein Science. 2022;31:e4429. wileyonlinelibrary.com/journal/pro 1 of 19 https://doi.org/10.1002/pro.4429 1469896x, 2022, 11, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/pro.4429 by Mit Libraries Serials & Journa, Wiley Online Library on [12/12/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License 2 of 19 HALPIN ET AL. 1 | INTRODUCTION 1, 2, 3, and 5 share high sequence similarity, whereas TRAF4 and TRAF6 are more diverged in sequence and Protein–protein interactions assemble signal transduction function.9,20–22 TRAF6 MATH is reported to bind pep- networks that are critical for cellular function and are tides that contain the motif xxxPxExx[FYWHDE] (here often implicated in disease. Knowledge of which proteins referred to as TIM6; Figure 1a), where x is any amino interact, and how, is essential for a mechanistic under- acid. TRAFs 1, 2, 3, and 5 have been reported to bind standing of information propagation in cells and how [PSAT]x[QE]E, PxQxxD, and PxQxT motifs,1 although these networks are perturbed in disease. Identifying mol- other work suggests deviations from these definitions.28 ecules that can mimic and compete with native binding For TRAF6, the proline and glutamate residues, refer- partners provides compounds that can be used as enced here as motif positions (0) and (+2), appear strictly research tools; such inhibitors may also have potential to conserved for TRAF6 binding.19,23–28 A preference for be developed as therapeutics. aromatic or acidic residues at (+5) is maintained in Myriad interactions important for signaling involve peptides that have been experimentally validated to bind the binding of a recognition domain in one protein by a to the TRAF6 MATH domain (Figure 1a).19,23–28 Struc- short linear interaction motif (SLiM) in a partner protein. tures show that the TRAF6 MATH domain binds to Many such domain/motif pairs, including the TRAF6 peptides that extend a beta-sheet in the MATH domain MATH domain/TRAF6-interaction motif pair that is the (Figure 1b–d).19,23,24 Residues in position (+5) bind in a subject of this work, have been compiled in the Eukary- pocket comprised of aromatic and basic residues, engag- otic Linear Motif database.1 Most motif definitions are ing in electrostatic and pi-pi interactions (Figure 1d). based on patterns found in a few examples, leading to Given the low complexity of the TRAF6 motif PxExx incomplete models that do not fully capture the sequence [FYWHDE], we reasoned that there might be other features necessary or sufficient for binding in the cell. A determinants of high-affinity TRAF6 binding. To define deeper understanding of SLiM sequence requirements motif-proximal features important for the interaction of can come from large-scale screens, which can provide a SLiMs with the TRAF6 MATH domain, we used bacte- more comprehensive view of protein recognition domain rial surface-display screening to explore sequence pref- specificity.2–7 erences within a library denoted xxxPxExxx, with x TRAF6 is a member of the tumor necrosis factor being a random amino acid, keeping the proline fixed at receptor-associated factor (TRAF) family of adaptor pro- position (+0) and the glutamate fixed at position (+2). teins with E3 ubiquitin ligase functions.8–10 TRAF6 medi- We screened this library and identified 236 highly ates NF-κB signaling and thereby participates in enriched binders and 1,200 nonbinders. We then used immunity and inflammation-related pathways. TRAF6 structure-based modeling to examine the interaction binds directly or indirectly to tumor necrosis factor recep- between the peptides and the MATH domain. Our anal- tors and members of the interleukin-1 (IL-1) receptor/ ysis revealed residues within the motif that support Toll-like receptor superfamily, among other proteins. high-affinity binding and negative-design elements that Downstream targets for TRAF6-mediated K63-linked ubi- explain why many peptides that contain PxE are not quitylation connect to the regulation of proteins such as suitable TRAF6 ligands. These insights help to elucidate transforming growth factor-β-activated kinase-1 (TAK1), the determinants of TRAF6 binding affinity. We com- IκB kinase (IKK), and mitogen-activated protein (MAP) pared the sequence features of the library-identified kinases, which subsequently lead to the regulation of NF- binders with reported native TRAF6 binders and found κB and AP-1 activity.9,11 Direct inhibition of the C- that most native interaction partners do not match the terminal domain of TRAF6 (the TRAF-C Meprin and top sequences isolated from the library. Notably, there TRAF Homology—or MATH—domain) has been pro- are no sequences in the human proteome that share all posed and explored as a potential therapeutic strategy for of the features that are prominent among the tightest the treatment of a variety of pathologies such as cardio- binders from the screen. These results suggest that vascular diseases, diseases associated with obesity, osteo- native TRAF6 interaction partners may be under func- porosis, and others.12–18 tional selection for moderate affinity, and is consistent TRAF6, like other members of the TRAF family, has with observations that other factors, such as ligand olig- four domains. The N-terminal RING domain works with omerization, are important for triggering TRAF6 bind- the zinc finger domains as an E3 ubiquitin ligase. A ing in certain biological contexts.9,29 The lack of high- coiled-coil domain trimerizes TRAF6. The 17.4 kDa C- affinity binders in the proteome provides an opportunity terminal MATH domain engages peptides containing to out-compete native interactions using designed pep- TRAF interaction motifs (TIMs) and is responsible for tides or mini-proteins that have features uncovered in cellular localization.19 The MATH domains of TRAFs our screen. 1469896x, 2022, 11, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/pro.4429 by Mit Libraries Serials & Journa, Wiley Online Library on [12/12/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License HALPIN ET AL. 3 of 19 F IGURE 1 TRAF6 MATH domain interactions with TIM6 peptide ligands. (a) Alignment of TRAF6-binding sequences from known partners showing the numbering scheme used throughout this paper. (b–d) Structure of the TRAF6 MATH domain (cyan) bound to the CD40* peptide (dark grey), which includes a point mutation relative to human CD40, PDB ID 1LB6. (b) MATH domain in surface representation bound to the CD40* peptide. (c) Bound peptide with positions numbered as in (a). Peptide residues at (+1)–(+5) form a beta-strand that pairs with the MATH domain (paired strand in yellow). (d) Interaction of the (+5) Phe in CD40* with Phe 410 and Arg 392 in TRAF6. 2 | RESULTS specific binding register to facilitate analysis and model- ing. These sequences were presented in the context of 2.1 | Library screening by cell-surface flanking sequences from CD40; see Table S1 for details of display reveals strong positional the display constructs. To isolate cells displaying peptides preferences for peptides that bind TRAF6 that bound to the TRAF6 trimer, we carried out one round of initial enrichment using magnetic microbeads Bacterial-surface display can provide information about (see Methods). This procedure generated a smaller the binding of short peptides to protein interaction library, enriched in TRAF6 binders, that we designate domains.3,30 For this work, we developed surface-display MACSLib; this library was used as the input for subse- constructs in which TIM6 peptides were fused to the C- quent enrichment experiments. terminus of re-engineered OmpX,31,32 such that when the To identify high-affinity binders in MACSLib, two construct was expressed, the TIM6 peptides were pre- separate 5-round enrichment sorts were performed using sented on the outer membrane of Escherichia coli cells. FACS to separate binding library members from non- To measure binding to TRAF6, peptide-displaying cells binding members (details in Methods). The stringency of were incubated with biotinylated TRAF6 homotrimers the binding assay was gradually increased by using a consisting of the coiled-coil and MATH domains (con- lower concentration of T6cc for each round, from 300 to struct termed T6cc). The amount of TRAF6 bound to the 3 nM. Following sorting, the population of binding cells cells was then quantified by adding streptavidin- in each round was deep sequenced to monitor the enrich- conjugated phycoerythrin and analyzing the cells via ment of individual sequences. The two replicate sorting fluorescence-activated cell sorting (FACS). The level of experiments gave similar results, with sequences from peptide expression was quantified simultaneously, using rounds four and five reflecting similar preferred residues, a FLAG-binding antibody conjugated to allophycocyanin indicating convergence of the selection process (details in Methods). (Figure 2). In both replicate experiments, the three To evaluate the TRAF6 MATH domain interaction sequences LNLPEESDW, RNVPEESDW, and motif space, we constructed a combinatorial library by TNWPEENDW ranked among the top four binders based introducing random residue variation around the core on sequencing read counts, and 14 of the top-20 most- TIM6 element PxE. We used degenerate NNK codons to represented sequences were the same in the two datasets. encode any of 20 amino acids at “x” positions in the Examination of enriched sequences, particularly those in sequence xxxPxExxx. The proline at position (+0) and the final rounds, indicated a strong preference for Trp at the glutamate at position (+2) were held fixed to increase position (+5). In addition, preferences were evident for the proportion of binders in the library and to force a Asn at (2), Glu at (+1), Ser/Asn at (+3), and a polar or 1469896x, 2022, 11, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/pro.4429 by Mit Libraries Serials & Journa, Wiley Online Library on [12/12/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License 4 of 19 HALPIN ET AL. FIGURE 2 Sequence logos for TRAF6-binding and nonbinding peptides. TRAF6 binders were identified by initial MACS enrichment followed by two replicate 5-round FACS enrichment experiments. MACSLib and enrichment sort replicate logos are built from the unique sequences with read count ≥ 20 in the samples. A total of 2,865 sequences obtained after the MACS enrichment represent only a subset of the library at that stage but reflect the diverse residue content provided as input to the enrichment sorts. The final list of binders (red box) was generated by combining results from both replicates and further filtering for sequences that enriched across multiple rounds. Nonbinders were defined as sequences with read count ≥ 20 in the nonbinder pool (black box). Sequences of peptides selected for structural modeling are summarized in the blue box. Residue height in the logos represents the frequency of that residue in the sequence set. The number of sequences in each set is shown in parentheses. acidic residue at (+4). A final set of 236 high-confidence TABLE 1 Validation of binding for peptides enriched in the TRAF6 binders was generated by taking the set of cell-surface display screen sequences present in rounds 4 and 5 of either replicate Peptide Single clone FACS BLI and filtering for sequences that enriched over at least sequence Kd* (μM)a K (μM)bd 2 rounds of sorting (Figure 2 red box; see Methods for KQEPQEIDF 1.2 ± 0.21 240 ± 23 details). We also generated a population of nonbinders by (CD40*) collecting cells from the original unenriched library that TNWPEENDW 0.084 ± 0.022 37 ± 6.9 gave a strong peptide-expression signal but no TRAF6 LNLPEESDW 0.046 ± 0.0090 28 ± 3.1 binding signal. The logo for nonbinders did not show c strong enrichment of any particular features (Figure 2, RNVPEESDW 0.031 ± 0.0024 24 ± 3.8 black box), and the diversity of the nonbinders confirmed aSingle clone FACS Kd* measurements were performed with trimeric that the input library included all 20 amino acids at each TRAF6 (T6cc). b of the “x” positions. BLI Kd measurements were performed with monomeric TRAF6 (T6m). cAverage of 2 replicates; in all other cases, the reported values are the To verify that the screening hits bound to TRAF6 in a average of 3 replicate binding curves ± the standard error of the mean. concentration-dependent manner, we performed single- clone titration experiments. After titrating TRAF6 trimers into a clonal population of peptide-displaying cells, we measured Kd for 14 peptides selected from the enrich- determined an apparent cell-surface dissociation constant ment data as well as a peptide from CD40 with an (Kd) by fitting the binding signal versus TRAF6 concen- affinity-enhancing point mutation (KQEPQEIDF, here tration to a standard binding model (Methods; termed CD40*).19,28 Interestingly, all of the top peptides Equation (1)). Binding between TRAF6 trimers and from the enrichment bound TRAF6 with an apparent peptide-displaying cells was multivalent, and the avidity affinity tighter than the CD40* peptide, with some bind- enhanced the apparent dissociation constants. We ing over 10-fold tighter (Table 1 and Figure S1; see 1469896x, 2022, 11, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/pro.4429 by Mit Libraries Serials & Journa, Wiley Online Library on [12/12/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License HALPIN ET AL. 5 of 19 F IGURE 3 Biolayer interferometry (BLI) measurements of TRAF6 monomer in solution binding to different peptides on the BLI tip. (a) Binding signal is plotted against TRAF6 concentration and fit to a standard single-site binding equation (Equation 1). Error bars are the standard error of the mean of three replicate measurements. (b) Average dissociation constant from three independently-fit replicate binding curves. The error bars are the standard error of the mean. The dissociation constant fit for the CD40* peptide has a large error because the highest TRAF6 concentration used was 150 μM. Methods). For example, RNVPEESDW, LNLPEESDW, designated as nonbinders (termed MD nonbinders) (see and TNWPEENDW bound TRAF6 with Kd values of Methods and supplementary information). Figure 2 31, 46, and 84 nM, respectively, whereas CD40* bound shows logos summarizing features of the two subsets of with Kd = 1.2 μM. sequences. To further validate the cell-surface interactions identi- We first tested whether FlexPepBind (FPB), a peptide fied in the screen, we measured TRAF6-peptide binding modeling protocol in the Rosetta suite,33 could distin- by biolayer interferometry (BLI), using purified mono- guish the MD binders from the MD nonbinders. As input meric TRAF6 MATH domain (construct termed T6m) in to FPB, we prepared models of MATH domain-peptide solution and purified peptides attached to the sensor tip complexes using the structure of TRAF6 bound to CD40* (see Table S1 for construct details). By BLI, (KQEPQEIDF) (PDB ID 1LB6) as a template.19 Starting RNVPEESDW, LNLPEESDW and TNWPEENDW bound from this initial docking position, the binding pose of the to TRAF6 with Kd values of 24.0, 27.5, and 37.2 μM, peptide was sampled, and the lowest interface score over respectively, while CD40* bound with a Kd of 238 μM all sampled poses was assigned to each peptide complex (Figure 3). The BLI data validate the cell-surface display (see Methods for details). Figure 4a shows the score dis- results and support the conclusion that top hits from the tributions for the 48 binders and 41 nonbinders by FPB screen bind with higher affinity than CD40*, which is score, which achieves a good separation of the two popu- one of the tightest known TRAF6 peptide binders19 lations. CD40*, scored with the same protocol, gave an (Figure 3). Based on these observations, we conclude that FPB interface score of 34.2, which is in the weaker end despite the screening assay being performed in the envi- of the range of binding peptides, consistent with the ronment of the cell surface, and in a multi-valent context, affinity measurements discussed above. enrichment sorting returned high-affinity binders. Because 9-residue peptides will sample an ensemble of conformations when bound to the TRAF6 domain, we also tested a molecular dynamics-based protocol for eval- 2.2 | Structural modeling explains uating peptide-domain interactions. Starting with com- positive and negative binding plexes modeled on the structure of TRAF6 bound to determinants KQEPQEIDF, as described above, we computed a detach- ment temperature (Detach T) for each model, corre- For computational analysis of the structural determinants sponding to the temperature at which the distance of TRAF6-TIM6 binding, we chose a subset of high- between the alpha-carbon of TRAF6 Phe 471 and the affinity binders identified from enrichment sorting center of mass of the peptide increased beyond 7 Å when (termed MD binders) and a subset of sequences the temperature was gradually increased from 300 K. 1469896x, 2022, 11, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/pro.4429 by Mit Libraries Serials & Journa, Wiley Online Library on [12/12/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License 6 of 19 HALPIN ET AL. FIGURE 4 FlexPepBind (FPB) and Detach T scoring of TRAF6 MD binders and MD nonbinders. The FPB interface score (a) and Detach T (b) are plotted for a set of TRAF6 peptide binders (orange) and a set of nonbinders (blue) identified by high-throughput screening. FIGURE 5 Structurally conserved features of high-affinity complexes. (a) Five beta-sheet hydrogen bonds involve main-chain atoms of residues at positions (+1), (+3), (+5) (yellow), and residues 472, 470, and 468 in the MATH domain (green). The image shows a snapshot from a simulation of TRAF6 MATH in complex with the peptide LNLPEESDW. (b) Positions of Pro at (+0) and Glu at (+2) (sidechains in sticks) from different frames of the equilibrated MD simulation of the peptide LNLPEESDW bound to TRAF6 MATH (color scale: red-white- blue for snapshots from the beginning-middle-end of the equilibrated part of the simulation). Pro binds into the pocket shown with cyan mesh, and Glu caps the short helix marked in blue. (c) The two most populated clusters for Trp conformations at position (+5) for 65% of the high-affinity binders. This sidechain arrangement allows simultaneous pi–pi interaction with Phe 410 and cation–pi interaction with Arg 392. The expanded region highlights snapshots from the two most common conformations. Detach T, like FPB interface score, was able to separate molecular dynamics simulations to analyze TRAF6 com- binding peptides from most nonbinders, as shown in plexes with peptide ligands CD40*, each of the 48 MD Figure 4b, with no binders giving Detach T values lower binders, and each of the 41 MD nonbinders (see than 700 K. Methods). With CD40* and all MD binders, simulations To explore the structural origins of sequence trends showed the persistence of 5 hydrogen bonds that posi- apparent in our enrichment sorting results, we used tioned the peptide as an extension of the beta-sheet in 1469896x, 2022, 11, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/pro.4429 by Mit Libraries Serials & Journa, Wiley Online Library on [12/12/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License HALPIN ET AL. 7 of 19 F IGURE 6 Overview of the most significant contacts between the highest affinity binders and the TRAF6 MATH domain, as captured in molecular dynamics simulations. Panels (a)–(f) illustrate specific interactions discussed in the text. Key peptide residues are represented in sticks: grey for Pro at (+0) and Glu at (+2), purple for residues at positions that favor a particular amino acid, and orange for residues at positions that favor a group of amino acids with similar features. All of the highlighted interactions are present for >30% of simulation time for all MD binders. (a) Asn at (2) can simultaneously form H-bonds with Glu 448 and Thr 475. (b) Glu at (+1) forms a bi-dentate salt- bridge interaction with Arg 402 when both sidechains are fully extended. (c) Residues at positions (+3), (+4), and (+5), are located in an electrostatically positive region (as indicated by blue coloring), PDB ID 1LB6. (d) Interactions involving residues at (+1) and (+5), shown as space-filling Glu and Trp in purple, along with MATH domain residues 402, 410, 392, and 394 (space-filling, cyan), narrow the pocket at position (+3) so that small residues, such as the Asn pictured in orange sticks, are preferred at this site. (e) Residues at position (+4) face solvent, and acidic residues at this site, such as the pictured Asp, can form a salt bridge with Lys 469. (F) Trp at (+5) engages in edge-to-face pi-pi and cation-pi interactions with residues Phe 410 and Arg 392, respectively. the MATH domain, as seen for CD40* in PDB structure sidechain. Most frequently, the indole group was 1LB6 (Figures 1c and 5a). The hydrogen bonds involved inserted into the receptor pocket (Figure 5c), allowing backbone atoms of residues in positions (+1), (+3), and for simultaneous pi-pi interaction with Phe 410 and (+5) that were highly stable during all 80 ns of cation-pi interaction with Arg 392. This conformation equilibrated-MD simulation. Invariant TIM6 residues Pro represented the most populated cluster for 65% of the at (+0) and Glu at (+2) also preserved their crystallo- MD peptides and resembles the conformation of Phe at graphic positions throughout all trajectories, with only position (+5) in the complex of CD40* bound to TRAF6 minor displacements (Figure 5b). Pro at (+0) was accom- MATH (Figure 1d).19 In particular, clustering the (+5) modated in the pocket created by residues Phe 471, Met Trp conformations from simulation frames by RMSD 450, and Tyr 473, while the negatively charged Glu at showed that more than 60% of the conformations were (+2) capped a 3–10 helix formed by residues Leu within 1 Å of the sidechain arrangement shown in 456, Leu 457, and Ala 458 (Figure 5b) in the MATH Figure 5c. We also observed structures in which the Trp domain. indole was flipped out of the pocket but maintained a Trp at (+5) was present in most of the binders binding interface, including backbone H-bond interac- obtained from enrichment sorting, even though this res- tions with Pro 468 and occasional pi–pi or cation–pi idue is not common in known native interaction part- interactions with Phe 410 or Arg 392. Such conforma- ners of TRAF6 (Figure 1a). Indeed, only 14 out of tions were shared among 20% of the MD binders. The 236 binder sequences identified in the enrichment remaining 15% of the MD binders showed unclustered screen did not have tryptophan at position (+5). Our Trp (+5) conformations in which the backbone was still simulations showed different conformations for the Trp involved in an H-bond interaction with Pro 468, but the 1469896x, 2022, 11, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/pro.4429 by Mit Libraries Serials & Journa, Wiley Online Library on [12/12/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License 8 of 19 HALPIN ET AL. indole group was flipped out and did not contact the TABLE 2 Fraction of binder and nonbinder sequence sets with MATH domain. certain sequence features The preference for Asn at (2) in binders from the Sequence feature Binders Nonbinders screen can be explained by its sidechain interactions with Proline any position between (+1) 0/236 386/1200 nearby TRAF6 residues. In all of the MD binders, the and (+5) Asn formed a stable interaction with the backbone of Thr Positive charge (R, K) at (+3), 2/236 287/1200 475 on TRAF6 for >40% of simulation time. In more than (+4), or (+5) 70% of the MD binders, the Asn also formed a hydrogen Large/medium residues (Q, H, I, L, 4/236 396/1200 bond with Glu 448 on TRAF6 for >30% of the simulation F, Y, W) at (+3) time (Figure 6a). Longer residues at (2) (e.g., Gln) were unable to interact with both TRAF6 amino acids. This interaction pattern appears to be important for high- affinity binding: Asn at (2) is present in 190 of the (+4), and (+5) (Figure 6c), suggesting that positively 236 binders from the enrichment. charged residues would be destabilizing at these sites. A preference for Glu at position (+1) can be Indeed, Arg or Lys are found at one or more of these explained by a salt–bridge interaction that this residue positions in 287 of 1,200 nonbinders (24%) but in only forms with Arg 402 in the MATH domain (Figure 6b). 2 of 236 binders (0.8%). Mutations by Pullen et al. con- Despite a Cα-Cα distance of 10 Å, the Arg 402 sidechain firm that these substitutions disrupt binding at these can form a salt bridge with the (+1) Glu when both are three positions in the context of a CD40 peptide.28 Steric fully extended toward one another. In the simulations, constraints at position (+3) are further expected to disfa- this interaction was stable for more than 80% of equili- vor medium or large residues at this site. Consistent with brated trajectory time and was completely missing for this, residues Q, H, I, L, F, Y, or W are found at position peptides in which Glu is substituted with Asp due to the (+3) in 396 of the 1,200 nonbinders (33%) but in only 4 of shorter sidechain of the smaller residue. the 236 binders (2%). Overall, 851 of the 1,200 nonbinders At position (+3), interactions involving residues at (71%) have at least 1 of the unfavorable sequence features (+1) and (+5), and the positions of MATH domain resi- described above (see Table 2 for summary). The nonbin- dues 392, 394, 402, 410, and 474, narrow the pocket, pro- ders also lack key residues that form stabilizing interac- viding an explanation for why small residues, such as tions in the highest affinity binders. Only 158 of the 1,200 Asn and Ser, are preferred at this site (Figure 6d). Resi- nonbinders (13%) contain Asn at (2), Glu at (+1), Asp dues at position (+4) are less sterically constrained and at (+4), or Trp at (+5), while all of the 236 binders con- some can form a hydrogen bond or salt bridge with sur- tain at least one of these interactions. Only 7 of 1,200 face Lys 469 (Figure 6e); all MD binder peptides with Glu nonbinders (0.6%) contain two or more of these stabiliz- at (+4) formed a salt bridge with Lys 469 in more than ing residues, while 224 of 236 binders (95%) contain two 60% of simulation frames. or more of these interactions. Analyzing our models of the MD binders helped explain why many nonbinders did not form tight interac- tions with TRAF6, despite including the conserved Pro at 2.3 | Candidate TRAF6 interaction (+0) and Glu at (+2). At positions (+1), (+3), and (+5), motifs in the proteome do not share the the MD binders make hydrogen bonds that complete a sequence features of the top screening hits beta-sheet with TRAF6. Proline residues are disfavored in beta structures because they lack the required NH group We investigated whether any human proteins contain for this interaction and prefer backbone dihedral angles close matches to the high-affinity sequences identified by far from the typical range in β-sheets.34 Thus, Pro at any screening. We defined a position-specific scoring matrix position between (+1) and (+5) is expected to be highly (PSSM) to score candidate TRAF6 interaction motifs unfavorable. Indeed, Pullen et al. showed that mutation based on how well they match our top binders. We used to proline at any of these positions in a peptide from pLogo,35 a log-odds-based method, to construct the CD40 (sequence KQEPQEINFPDDLP) abrogated binding PSSM, using the 236 binding sequences from the enrich- in peptide array experiments.28 A total of 386 of the 1,200 ment as the foreground and the 1,200 nonbinder nonbinders (32%) identified in our screen have such a sequences as the background. The nonbinder sequences substitution, which is likely sufficient to prevent high- were considered a fair approximation of the sequence affinity binding. None of the 236 binders contain a pro- composition of the input library, assuming that TRAF6 line at these positions. Furthermore, the TRAF6 MATH binders are rare in the library. Indeed, we do not observe domain is electrostatically positive near positions (+3), any apparent residue preferences in the nonbinder set 1469896x, 2022, 11, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/pro.4429 by Mit Libraries Serials & Journa, Wiley Online Library on [12/12/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License HALPIN ET AL. 9 of 19 F IGURE 7 TRAF6 motif scoring. (a and b) The PSSM scores of selected TRAF6 binding peptides are correlated with their apparent cell-surface binding affinity for TRAF6. (a) Reported dissociation constants are the average of fits to 2–3 replicate titrations. The standard error of the mean (SEM) is reported for each sequence. (b) Correlation between Kd and PSSM score (data from a). Error bars are the standard error of the mean of 2–3 replicates. (c, top) Position-specific scoring matrix generated from the screening data and used to score candidate binding motifs. (c, bottom) Sequence logos of all unique TRAF6 motif matches (motif: xxxPxExx [FYWHDE]) in the human proteome compared to experimentally validated TRAF6 MATH domain binders. (d) Distribution of normalized PSSM scores of all TRAF6 motif matches in the proteome, using the matrix shown in the top of panel c. (Figure 2). To test if the PSSM score of a sequence repre- Despite the low scores for proteome sequences, we sents how well that peptide binds to TRAF6 on the cell explored whether the screen-based PSSM could be used surface, we scored the sequences used in single clone in conjunction with other metrics to identify proteome titrations. Scores were normalized to the range 0 to sequences with the potential to interact with TRAF6 with 1, with 0 the lowest and 1 the highest possible PSSM high affinity. We constructed a table from the proteome score. We found that PSSM score is correlated with motif matches that includes a variety of scores and filters apparent cell-surface affinity, suggesting that our model for each hit, compiled from multiple sources (Table S4; is a good predictor of TRAF6 binding within this see the supplementary information for details). We sequence space (Figures 7a,b). included indicators of whether the motif is predicted to The PSSM was used to score TRAF6 motif matches in be structurally accessible for binding (IUPred score37 and the human proteome to identify SLiMs with the potential AlphaFold pLDDT score38–40), whether the candidate to bind with high affinity. TIM6 matches (10,451 hits protein is involved in similar biological processes as total) were obtained using the SLiMSearch tool36 (regular TRAF6 (based on shared Gene Ontology (GO) annota- expression: …P.E…[FYWHDE]). The logo of hits is shown tions with TRAF6, from SLiMSearch),36,41,42 whether the in Figure 7c, along with a logo of the experimentally vali- protein has been annotated to interact with TRAF6 dated TRAF6 binding peptides from Figure 1a. Figure 7d (HIPPIE database43), and whether the motif has any shows the distribution of PSSM scores for the sequences unfavorable sequence features identified in our structural retrieved using SLiMSearch.36 Notably, no sequences in analysis (Table 2). Applying filters based on these criteria the proteome occupy the sequence space favored in the narrowed the list of potential biologically relevant TRAF6 screen (i.e., no sequences have a high score). The motifs in the proteome from 10,000 to 1,000 highest-scoring sequence in the proteome is FNE- sequences matching the xxxPxExx[FYWHDE] motif. PEENFW, with a score of 0.85. Only 4 motifs in the pro- Among these candidate motifs, we chose a few with high teome have a PSSM score above 0.75, and only 1 of those PSSM scores, indicating potentially high affinity, for fur- is predicted to be disordered by IUPred (IUPred >0.437). ther analysis (Table S4). Additionally, the motifs in the proteome that are experi- One of the highest-scoring hits in the proteome is the mentally validated to bind to the TRAF6 MATH domain sequence QNFPVESDW (PSSM score = 0.85) from have low PSSM scores (Table S3). RNF103. RNF103 acts as an E3 ubiquitin-protein ligase 1469896x, 2022, 11, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/pro.4429 by Mit Libraries Serials & Journa, Wiley Online Library on [12/12/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License 10 of 19 HALPIN ET AL. that is localized to the ER membrane; it is involved in the conclusion is that the proteome contains few sequences ER-associated degradation (ERAD) pathway.44 This that match our screen-derived PSSM, and most native sequence has the highly favorable residues Asn and Trp sequences with high scores do not appear to be good at positions (2) and (+5), respectively. The sequence interaction candidates based on other metrics. Biology also contains favorable Ser and Asp residues at (+3) and does not use the sequence space of highest affinity for (+4), respectively. The TRAF6 motif match is not pre- native TRAF6 interactions. dicted to be disordered by IUPRED. However, the average AlphaFold pLDDT score of the motif is 38.7 (correspond- ing to predicted disorder) and the motif residues appear 3 | DISCUSSION accessible in the AlphaFold-predicted structure.38–40 Although RNF103 localizes to the ER membrane, the The discovery of TRAF6 interaction partners over candidate motif (positions 474–482) maps to the cytosol, decades of experimental research has led to a definition given its location between the last transmembrane helix of the TRAF6 MATH domain binding motif as xxxPxExx and the cytosolic RING domain of RNF103. [FYWHDE]. This was arrived at by compiling aligned, The sequence GMGPVEESW starts at position 350 in verified TRAF6 binding sequences and identifying their RIPK1. The sequence has a PSSM score of only 0.46, but common sequence features.19,23–28 In this work, we has a highly favorable Trp at (+5) and lacks any of the explored the TRAF6 motif sequence space systematically, major unfavorable sequence features identified by struc- using cell-surface screening of a combinatorial library tural modeling (Table 2). RIPK1 is a serine–threonine that presented the core PxE motif flanked by random res- kinase involved in regulating TNF-mediated apoptosis, idues. The top hits obtained from this screen bound with necroptosis, and inflammatory pathways.45 It has been affinities comparable to or higher than known TRAF6 annotated as a TRAF6 interaction partner in the HIPPIE interaction partners reported in the literature.19,24 Analy- database, but the details of the interaction are unknown. sis of screening hits highlighted which residues were RIPK1 has been found to bind to other TRAF proteins46 most preferred at each position and identified features and also to TICAM1.47 We speculate that RIPK1 may that differentiate binding sequences from nonbinding interact with TRAF6 via the MATH domain engaging sequences among protein segments that contain the core this short segment. element PxE. Another hit with potential for biological significance is Two different methods of structure-based modeling the sequence GNFPEENND, which spans positions 1,065– could distinguish the best-binding peptides from the 1,073 in the leptin receptor. This sequence contains an Asn background, and we used molecular dynamics simula- at (2), Glu at (+1), and Asn at (+3), which are all favor- tions to study the bound ensembles of diverse binders. able residues according to our model. It has a PSSM score This analysis provided a structural explanation for the of 0.49. The leptin receptor binds leptin, which is secreted residue preferences observed in our screening data. In from adipose cells. In obese mammals, leptin levels are ele- particular, in all high-confidence binders that we ana- vated, leading to chronic low-grade inflammation.48 TRAF6 lyzed, Asn at position (2) can form favorable interac- is a well-known regulator of the inflammation response, tions with Glu 448 on TRAF6, Glu at position (+1) can suggesting a possible link between the two pathways. form a salt bridge with Arg 402 on TRAF6, and Trp at SLiMs are known to evolve rapidly,49 so although position (+5) can form pi-pi interactions with Phe conservation of a motif can support its functional rele- 410 and cation-pi interactions with Arg 392. Our struc- vance, lack of conservation does not necessarily indicate tural analysis also highlighted negative design elements that a motif is not functional. It is notable that the motif that can disfavor PxE-containing segments binding to instances in RNF103 and the leptin receptor are not TRAF6 MATH. The logo of nonbinders in Figure 2 does highly conserved across species (Figure S3). For RIPK1, not indicate any strong features, but our data support a the tryptophan at (+5) in the motif is not conserved model in which a variety of sequence features, including beyond mammals. However, the presence of a TRAF6 proline in positions (+1) to (+5), a large residue at posi- motif in this region of RIPK1 appears widely conserved tion (+3), or a positively charged residue at (+3), (+4), or across species; the TRAF6 motifs in other species typi- (+5) can disfavor binding. Thus, within this 9-residue cally include acidic glutamate at (+5). Interestingly, the stretch that includes PxE, several positive features and position of the motif within RIPK1 also varies among spe- the absence of a variety of negative-design elements are cies (Figure S3), which has been previously observed for important for making a functional TRAF6 binder. SLiM evolution.49 Experimental follow-up will be The sequence preferences observed in this study, required to assess whether any of the proteome hits inter- determined using a large and diverse library, can be com- act with TRAF6 in a biological context. Overall, our pared with point mutations in the native CD40 TIM6 1469896x, 2022, 11, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/pro.4429 by Mit Libraries Serials & Journa, Wiley Online Library on [12/12/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License HALPIN ET AL. 11 of 19 peptide that were analyzed using peptide SPOT arrays.28 propagates the signal further downstream.9 Artificial olig- Overall, there is good agreement between the two studies omerization of TRAF6 alone is sufficient to activate sig- regarding the favorable and unfavorable residues for naling through certain pathways, implying that TRAF6 binding. Contrary to our observations, however, the oligomerization is a key part of the signal transduction.29 SPOT array suggests that large residues (H/I/L/F/Y/W) In this scheme, conserving a weak, fast-exchanging inter- are tolerated at (+3), in the context of the CD40 pep- action between individual motifs and monomeric MATH tide.28 This discrepancy could arise from epistasis domains is likely important for supporting rapid, ligand between positions in the motif. For example, Trp at (+5) binding-dependent assembly and disassembly of a TRAF6 (as opposed to Phe, in CD40) may place stricter spatial signaling complex. constraints on the pocket at position (+3), such that large When profiling the binding preferences of a SLiM- residues are occluded from position (+3) when there is a binding domain using high-throughput screening experi- Trp at (+5), but not when the (+5) position is Phe (as in ments, different libraries may have different applications. CD40). In support of this hypothesis, the CD40 SPOT For identification and prediction of new interaction part- array shows reduced binding when W is substituted for F ners in the proteome, biologically relevant libraries, such at the (+5) position. Foight et al. also found coupling as a library of tiled sequences from the proteome, are between mutations, and a sensitivity of mutations to likely more effective than a randomized library.4,30 In motif context, in the interactions of peptides with TRAF2, contrast, for the purpose of synthetic biology or inhibitor TRAF3, and TRAF5 MATH domains.3 design, a randomized library has the potential to better Our analysis of the proteome revealed that no identify high-affinity sequences that are more likely to sequences map to the high-affinity, TRAF6 binding out compete native binders. sequence space that we identified using cell-surface dis- High-affinity TRAF6 binders isolated in this work can play, and that our resulting PSSM is not a good predictor serve as lead peptides for inhibitor development. TRAF6 of endogenous TRAF6 binding sequences. Indeed, most signaling is implicated in inflammation and cardiovascu- of the well-studied peptides from verified TRAF6 MATH lar disease.12,13,18 Targeting TRAF6 MATH is reported to domain binders lack the features of the highest-affinity improve insulin sensitivity in obese mice, improve heart binders identified in the screen or only contain 1 or 2 of function in mouse models of non-ischemic cardiac fail- the most favorable interactions we identified (Figure 1a, ure, reduce atherosclerosis, and inhibit osteoclastogenesis Table S3). Only 3/12 have an Asn at (2), and only and bone resorption.14–17 A RANK peptide attached to a TICAM1 contains Trp at (+5). Our findings imply that protein transduction sequence to promote cell entry is the core binding motif is either not under selection for currently sold commercially as a TRAF6 inhibitor high affinity, or high affinity is detrimental to TRAF6 (e.g., Novus Biologicals NBP2-26506).16 The reported function. affinity of a RANK peptide with sequence RKIPTEDEY Other library-based studies have found enrichment of for TRAF6 is 78 μM, determined by isothermal titration hydrophobic residues in binders of protein domains that calorimetry.19 The same study reported a Kd of 84 μM for does not reflect the composition of the native binding the CD40* peptide, and we showed that peptides from partners.2,50 The distinct properties of native vs. library- our screen bind 10-fold tighter than CD40* (Figure 3). selected binders could be due to natural selection for Thus, peptides from our screen, possibly further opti- binding specificity, solubility, or peptide intrinsic disor- mized by adding an optimal flanking sequence, can serve der, rather than affinity. Weak SLiM binding that allows as higher-than-native-affinity inhibitors. Having a broad for transient and short-lived interactions can also provide range of peptide sequences that can disrupt TRAF6 bind- advantages.51 For example, complexes that use multiple ing, as we have generated here, can support further weak interactions rather than one higher-affinity binding efforts to develop inhibitors with desirable properties, site provide opportunities for regulation, and enhanced such as low immunogenicity and cell permeability. specificity, as is the case for tandem recognition of SLiMs by SH2, SH3, WW, and other domains.52–55 The TRAF proteins provide a different example of the benefits of 4 | MATERIALS AND METHODS weak binding for signaling. TRAF6 uses avidity to signifi- cantly enhance binding affinity to oligomeric receptor Vectors, bacterial cells, and cloning: The expression con- proteins. In concentration regimes where the binding structs and cell surface display constructs are detailed in affinity of a single-motif peptide is not significant, recep- Table S1. The TRAF6 trimeric construct, here termed tor oligomerization can trigger TRAF6 trimers binding to T6cc, consisted of residues 310–504 of human TRAF6 three or more motifs in the tails of clustered cytoplasmic (including the MATH domain and coiled-coil trimeriza- receptors, which then promotes ubiquitylation that tion domain), an N-terminal BAP tag for biotinylation, 1469896x, 2022, 11, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/pro.4429 by Mit Libraries Serials & Journa, Wiley Online Library on [12/12/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License 12 of 19 HALPIN ET AL. and a hexahistidine tag for purification. The construct before storing at 80C in aliquots for later use. Concen- was expressed using a pDW363 vector to ensure biotiny- trations of T6cc are reported as monomer concentrations. lation. A monomeric TRAF6 construct lacking the tri- For solution binding studies, a monomeric variant of merization domain and BAP tag, here termed T6m, TRAF6 (T6m) was expressed in Rosetta2(DE3) cells over- consisted of residues 350–501 of human TRAF6 and a night at 18C and purified similarly to T6cc. T6m was hexahistidine tag. The construct was expressed in a purified into a final buffer of 50 mM Tris pH 8.0, 180 mM pDW363 vector, although the lack of a BAP tag ensured NaCl, 5% glycerol, and 1 mM DTT. SUMO-peptide fusion no biotinylation of this protein. SUMO-peptide fusion proteins were co-expressed with the biotin ligase BirA constructs contained a BAP tag, hexahistidine tag, and (from the pDW363 vector) in Rosetta2(DE3) E. coli for SUMO tag. The construct was expressed in a pDW363 5 hours at 37C. Media was supplemented with 0.05 mM vector to ensure biotinylation. E. coli strains BL21(DE3), D-(+)-biotin. The protein was purified by Ni2+-NTA DH5α, and MC1061 were used for protein expression, affinity chromatography followed by gel-filtration into a cloning, and surface display, respectively. For bacterial final buffer of 20 mM Tris pH 8.0, 150 mM NaCl, 1 mM surface display of TRAF6-binding peptides, the eCPX vec- DTT, and 10% glycerol. tor designed by the Daugherty group56 was modified at Magnetic bead presorting (MACS): To generate the C-terminus to append a FLAG sequence, a linker TRAF6-bound beads, 2 ml of vortex-mixed Invitrogen containing a SfiI site, and the CD40* peptide sequence, DynaBeads™ Biotin Binder beads were incubated with which included 9 residues resolved in the X-ray structure T6cc (33 pmol biotinylated T6cc/10 μl beads) for 2 hours of CD40* bound to TRAF6 (PDB ID: 1LB619) plus 8 flank- at 4C and then washed in PBS buffer, as described by ing residues on each side of this core region. The CD40*- Angelini et al.57 The TRAF6-decorated beads were then derived sequence used was PTNKAPHPKQEPQEIDFPD added to cultures of induced cells expressing the peptide DLPGSNT. library (induced with 0.2% w/v arabinose for 2 hours at Mutant library construction: The library was con- 37C). After incubation for 3 hours at 4C, beads were structed using primers (from IDT) with NNK codons magnetically isolated for 60 seconds before aspiration included in positions marked “x” the motif xxxPxExxx, and replacement of PBS buffer. Beads were then gently such that the theoretical size of the library was shaken in the fresh buffer for 5 minutes at 4C. The bead 207 = 1.28 * 109 unique members. The variable sequence wash cycle was repeated 7 times before beads were placed was flanked by SfiI restriction sites for cloning. In paral- in LB media for growth overnight. 100 μl of the final lel, a linear vector for cell-surface display containing the growth stock was serially diluted on LB + agar constant sequence of the display construct, with SfiI sites +25 μg/ml chloramphenicol plates. Colony-forming units matching the library insert, was amplified by PCR. The were tabulated to back-calculate the number of cells in insert and linear vector fragments were purified by PCR the MACS-sorted library, which yielded 1.42 * 105 cells. Cleanup Kits (Genesee Scientific) before SfiI digestion. Bacterial FACS preparation: For enrichment sorts and Following purification of the digested fragments, a 5:1 single-clone FACS cell surface titrations, 5 mL cell cul- ratio of insert: vector was added to a 200 μl T4 DNA tures were grown overnight at 37C in LB + 25 μg/mL Ligase (New England Biolabs) reaction and then incu- chloramphenicol and 0.2% w/v glucose. The next day the bated for 16 hours at 4C. The ligated mixture was elec- culture cell density was measured by OD600, and approxi- troporated into fresh electrocompetent MC1061 cells in mately 3.25 * 105 cells of each stock were isolated for new four separate transformations. Transformed cells were growth in 5 mL LB. Upon reaching an OD600 of 0.5–0.6, transferred into 10 ml warm Super Optimal Broth media cells were induced with 0.2% w/v arabinose for 2 hours at with 20 mM glucose (SOC media) and incubated at 37C 37C. Density was again measured, and cells were pel- for 1 hour. The 10 ml culture was then added to 1 L of leted by centrifugation and resuspended in PBS + 0.5% LB + 25 μg/ml chloramphenicol and grown to an OD600 BSA. Cells were then aliquoted into a 96-well Multi- of 0.6–0.8 before centrifugation and resuspension in Screen HTS® GV sterile filtration plate (2 x 107 cells per LB + 20% glycerol for freezing for storage. sample) and washed with fresh PBS + 0.5% BSA. Cells Protein purification and preparation: T6cc was co- were then incubated in 30 μl of αFLAG-APC [PerkinEl- expressed with the biotin ligase BirA (from the pDW363 mer] (prepared at a 100:1 dilution in PBS + 0.5% BSA) at vector) in BL21(DE3) E. coli for 5 hours at 37C. Media 4C for 15 min. Next, cells were resuspended in 50 μl of was supplemented with 0.05 mM D-(+)-biotin. The pro- TRAF6 solution (25 μl PBS + 0.5% BSA mixed with 25 μl tein was then purified using Ni2+-NTA affinity chroma- of the chosen TRAF6 concentration) and incubated at tography followed by gel-filtration chromatography into 4C for 60 min. Following a wash with 200 μl PBS + 0.1% a final buffer of 20 mM Tris pH 8.0, 150 mM NaCl, 5% BSA, cells were resuspended in 30 μl streptavidin-PE glycerol, 1 mM DTT. Purified protein was concentrated (SA-PE) [ThermoFisher] (prepared at a 100:1 dilution in 1469896x, 2022, 11, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/pro.4429 by Mit Libraries Serials & Journa, Wiley Online Library on [12/12/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License HALPIN ET AL. 13 of 19 PBS + 0.1% BSA) and incubated at 4C for 15 minutes. in Kaleidagraph58 using non-linear least-squares fitting to Cells were then washed in 200 μl PBS + 0.1% BSA, resus- determine the dissociation constant. pended in another 200 μl PBS + 0.1% BSA, and placed on Enrichment sorting of MACS-presorted library: To iso- ice prior to FACS analysis or sorting. FACS analysis was late the best TRAF6 binders, we performed a five-round performed using an HTS Canto II instrument and sorting enrichment sort using the MACS-sorted library took place on a FACS Aria III cell sorter (MACSLib) as the input. On each day, the library was (BD Biosciences). Sorted cells were collected in 1.5 ml sorted for TRAF6 binding as described above (Bacterial microcentrifuge tubes containing 500 μl Luria-Bertani FACS preparation) using a single permissive gate set to media with 25 μg/ml chloramphenicol. collect successfully expressed TRAF6 binders. The gate Single-clone titration experiments. For single-clone was set manually each day using positive and negative titration experiments, samples for FACS analysis were binding controls. Selection for TRAF6 binding was gradu- prepared as described above using eight concentrations ally increased by using a lower concentration of T6cc of TRAF6 for each clone: 0 nM, 3 nM, 10 nM, 30 nM, each day (concentrations used: 300 nM, 100 nM, 30 nM, 100 nM, 300 nM, 1 μM, 3 μM. Binding curves were gener- 10 nM, 3 nM). Collected cells were grown overnight ated by plotting the mean PE value vs. TRAF6 concentra- before splitting half of the pool to continue the sort and tion and fit to the following equation to determine a Kd the other half to harvest for plasmid DNA and subse- value: quent Illumina sequencing. We performed two duplicate   5-day enrichment experiments, generating 10 total pools y¼Finitþð  Þ x ð Þ for deep sequencing.Fsat Finit þ  1x Kd Nonbinding clone FACS sorting. Using the unenriched (pre-MACS) library as input, a gate was drawn to define where y is the mean PE fluorescence value and x is the the region where peptide-expressing cells are found in concentration of TRAF6. Finit, Fsat, and Kd were treated the absence of TRAF6. This gate was used to collect 2 * as floating parameters; Finit is the y value in the absence 10 4 cells in the presence of a high TRAF6 concentration of TRAF6 and Fsat is the y value at which the binding ([T6cc] = 6 μM) to isolate clones with no detectable bind- curve saturates. Although the cell-surface binding data fit ing to TRAF6. well to a hyperbolic binding equation, this assay is not Illumina amplicon preparation: Figure S2 gives an likely to be at equilibrium, and we discourage interpreta- overview of this procedure. Sorted pools were grown tion of the apparent binding constant K as a true equi- overnight at 37d C in LB, and bulk plasmid DNA was har- librium dissociation constant. vested by QIAprep miniprep kit (Qiagen). We then PCR Biolayer Interferometry (BLI): BLI experiments were amplified the variable region of the plasmids from each carried out on an Octet Red96 instrument (ForteBio). cell-sorted pool, appending a MmeI restriction site to the Streptavidin-coated tips (ForteBio) were pre-incubated 50 end. At the 30 end, we appended: (a) an unused, ran- for 10 min in BLI buffer (20 mM Tris pH 8.0, 207 mM domized 9 nt barcode UID sequence, (b) a 6 nt indexing NaCl, 1 mM DTT, 1% Glycerol, 0.1% BSA, and 0.1% sequence for multiplexing (Illumina TruSeq), and (c) a Tween-20). Biotinylated SUMO-peptides were immobi- custom reverse-read annealing sequence. Barcodes are lized on streptavidin tips. Loaded tips were then given in Table S1, amplicon construction is depicted in immersed in a solution of the TRAF6 MATH domain, Figure S2, and a lookup table is provided in Table S2. which had been diluted to the relevant concentration in Amplified fragments were digested with MmeI. A BLI buffer. Association data were collected at room tem- double-stranded DNA fragment with a 2 nt overhang perature at an orbital shake speed of 1,000 rpm (sampling matching the MmeI cut site was then ligated to each rate) until the signal plateaued. Subsequently, TRAF6 MmeI-cleaved fragment. This fragment contained the bound tips were transferred to a well containing the standard 50 Illumina adapter sequence and one of 24 pre- above buffer, and dissociation data were collected until selected 5 nt barcodes for sample multiplexing. 50 and 30 the signal plateaued. Due to the fast kinetics of the inter- Illumina anchoring sequences were appended to the action, we elected to calculate Kd values using the steady- amplicons in a subsequent PCR amplification. More than state signal of the association step. The raw association 50 amplicons were Sanger sequenced (QuintaraBio) to data of a SUMO-only control was subtracted from that of assess amplicon quality, which revealed the expected the SUMO-peptides. The normalized signal of the associ- sequences and variable positions. The sequencing length ation step was averaged over 10 seconds after reaching a of each amplicon was 65 nt, so forward and reverse plateau and plotted against the concentration of TRAF6 paired-end 40 nt reads overlapped by 15 nt. Immediately MATH domain. The binding curve was fit to Equation (1) prior to Illumina sequencing, the MIT BioMicro Center 1469896x, 2022, 11, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/pro.4429 by Mit Libraries Serials & Journa, Wiley Online Library on [12/12/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License 14 of 19 HALPIN ET AL. verified fragment size for all pools by agarose gel and more reads on Day 4 and/or Day 5 in either replicate multiplexed all pools at equimolar amounts. enrichment. The resulting list of binders was then further Illumina data collection and processing: Illumina filtered to include only those sequences that enriched sequencing was performed on a NextSeq500. The reads two or more times (defined as an increase in read fre- were demultiplexed using custom python scripts: https:// quency from one day to the next day) during either github.com/jacksonh1/NGS_demultiplexing. Reads that enrichment replicate, yielding a final list of 236 unique did not exactly match one of the barcode/index pairs TRAF6-binding peptides. (first 5 nts of the forward read and first 6 nts of the Nonbinder data analysis: The NGS data from the non- reverse read, respectively) were discarded. Additionally, binder FACS sample (Nonbinding clone FACS sorting) we required each of the first 5 nts of the forward read to were analyzed to define sequences of peptides that do not have a Phred score of 20 or greater. Next, the “reformat. bind to TRAF6. Sequences were filtered to include only sh” tool from the BBTools suite (Version 38.94) was used those DNA sequences coding for peptides matching the to de-interleave the paired-end reads and filter for reads xxxPxExxx motif and having a read count of 20 or more. with an average Phred score greater than or equal to Amino acid sequences containing the characters “*” or 20 (using the parameter: “minavgquality = 20”).59 In our “X” were removed. The final list of nonbinders contained dataset, the forward reads covered the entire variable 1,200 unique peptides. region of the displayed peptide. Therefore, reverse reads Generation of PSSM for proteome scanning: To gener- were discarded after de-interleaving, and only the higher- ate a PSSM from the enrichment and nonbinder data, we quality forward reads were used for further analysis. For used pLogo, which uses log-odds-based scoring to gener- each sample, we used custom Python scripts to count the ate a PSSM from a given set of foreground sequences and abundance of each sequence in each sample at the DNA background sequences.35 We used the 236 unique level, using an alignment-based counting strategy. Here, TRAF6-binding peptides determined from the enrich- the forward reads were aligned to a counting template ment experiment as the foreground and the 1,200 unique sequence covering the variable region of the display con- nonbinding peptides from the nonbinder sample as the struct: *********CCT***GAA*********CCGG, where * rep- background. resents variable nucleotide positions. Sequences that Scoring motif matches in the proteome: To generate mismatched 3 or more times to constant positions of the a table of TRAF6 motif instances in the proteome, we template (non * positions) were removed. Sequence used SLiMSearch 436 to find all matches to the consen- counts were then further collapsed to just the TRAF6 sus TRAF6 binding motif (regex: “…P.E…[FYWHDE]”) motif region: *********CCT***GAA*********. The result in the human proteome. The PSSM generated with was a list of sequences and their associated read counts pLogo (described above), was then used to score the for each sample. NGS data, processed data files, and hits from SLiMSearch using custom python scripts. We Python scripts are available at https://github.com/ used the SLiMSearch “shared functional annotations” jacksonh1/TRAF6_screen and https://doi.org/10.6084/ feature to allow filtering the hits to proteins that share m9.figshare.20485914.v3 (see supplementary information Gene Ontology (GO) terms with TRAF6. The set of for file descriptions). All sequence logos in this study GO terms used to filter the hits can be restricted by were generated using the logomaker python library.60 the likelihood that a given term is shared by any two Enrichment data analysis: For each replicate enrich- proteins in the proteome (“sig” or “p-value”). We used ment experiment, the nucleotide sequences were trans- this feature and custom python code to create filters of lated into amino acid sequences, and only those DNA different cutoff values (sig < = 0.01, 0.001, 0.0001, and sequences coding for peptides matching the xxxPxExxx 0.00001) to allow filtering hits to those that share motif were kept for further analysis. Amino acid TRAF6 GO terms with sig less than or equal to the sequences containing the characters “*” or “X” were then given cutoff value. The HIPPIE database was used to removed (corresponding to sequences containing stop label motif instances in proteins that are annotated to codons or “N” bases). Read frequencies were calculated interact with TRAF6.43 AlphaFold 2 structure predic- by dividing the read count of each sequence in each sam- tions for the human proteome were downloaded from ple by the total number of reads in that sample. When a the AlphaFold Protein Structure Database.38–40 Per- sequence had fewer than 20 reads, the frequency was set residue AlphaFold pLDDT scores for each motif in the to 0 to minimize effects from low read counts. To deter- table (+3 flanking residues) were extracted from the mine a set of TRAF6 binding sequences, we first removed predicted structures using custom python scripts. For any sequences that did not have 50 or more reads on at proteins in the table with no corresponding AlphaFold least one of the five enrichment days or the input library Protein Structure Database entry, the pLDDT columns (MACSLib). We then filtered for sequences with 20 or were left blank. 1469896x, 2022, 11, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/pro.4429 by Mit Libraries Serials & Journa, Wiley Online Library on [12/12/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License HALPIN ET AL. 15 of 19 Selection of binding and nonbinding peptides for pose for each peptide, allowing both backbone and side- modeling studies: 48 binders were chosen from the bind- chain atoms to move, using the refinement flag, and ing sequences identified in the enrichment experiment applying harmonic constraints around the crystallo- for structure-based modeling (MD binders). Three of the graphic distances between the peptide and TRAF6 to 48 binders (RNVPEESDF, RNVPEESTW, and reduce the conformational sampling space. Specifically, WNMPAEYDF) came from an earlier analysis of the we restrained backbone hydrogen bond distances enrichment data and are not present in the final set of between the peptide residue in position (+0) and TRAF6 236 binders. However, all three sequences enrich at least residue G472 and between the peptide residue in position once during the enrichment experiment and are likely (+2) and TRAF6 residue G470 (using the observed dis- real binders despite not making the final cutoff. In addi- tances in structure 1LB6). Models were ranked based on tion, 41 nonbinder sequences were selected from the non- total interface score, calculated as the sum over energy binder pool for structural modeling (MD nonbinders). terms contributed by interface residues of both partners. Computational Rosetta modeling: The pipeline for Interface residues were defined as those with Cβ (Cα for modeling mutated peptide interactions proceeded as fol- Gly) within 8 Å of any atom in the TRAF6 protein. We lows. First, all structures were alchemically mutated onto used the lowest-interface score complex for our analysis. the crystal structure of TRAF6 bound to the CD40* pep- The following is the command-line flag array for model- tide (seq: KQEPQEIDF, PDB: 1LB6) using FoldX. For ing peptides using Rosetta: ($name indicates a wildcard each mutated pose, we used Rosetta relax to remove ste- inserted to match the peptide to be run). ric and angle violations. Next, the Rosetta FlexPepDock $rosdir/FlexPepDocking.static.linuxgccrelease -s module was used to create 500 poses of each using the $name\_Dock_0001.pdb -native $nativepdb -lowre- lowres_preopt flag to more aggressively sample the space s_preoptimize -pep_refine -nstruct 500 -use_input_sc in case of necessary residue rearrangement. The -ex1 -ex2 -out:file:silent $name\_Dock. silent -out:file: talaris2013 score function was used for all model scoring silent_struct_type binary. in Rosetta. The top pose by Rosetta score was isolated For the Detach-T protocol, the crystal structure was from each mutated sequence and used to rationalize resi- initially minimized and equilibrated for 20 ns with due preferences for the binders. The Rosetta version used CHARMM36a using ACEMD code. The resulting struc- was rosetta_bin_linux_2017.08.59291_bundle. ture was then mutated using the VMD-Mutator tool to Scoring peptide binding affinity: We implemented two introduce changes into structure 1LB6.63 We ran MD different computational pipelines for scoring peptide simulations with the temperature increasing from 300 up binding to TRAF6: FlexPepBind (FPB)33 and an in-house to 1,000 K, using a temperature step of 10 K every 100 ps protocol based on computing a detaching temperature for a total time of 5 ns, restraining protein CA that were (DetachT) by using short molecular dynamics simula- more than 15 Å from the binding pocket to avoid protein tions at increasing temperatures. diffusion in the unit cell. For each peptide, we recorded Structures were prepared using TRAF6-CD40* com- the temperature at which the distance between the geo- plex structure 1LB6 as a model. All sequences were metrical center of TRAF6 residue 471 and the center of 9 amino acids long and shared the PxExxAr short linear mass of the peptide segment composed of residues P TRAF6-interacting motif (TIM6). We assumed that all (0)  x (+1)  E (+2) increased to greater than 7 Å. peptides bound in the canonical TRAF6 binding groove Molecular Dynamics simulation of a subset of with a position similar to that of the CD40* peptide TRAF6-peptide complexes: For 89 complexes (48 MD KQEPQEIDF.19 binders and 41 MD nonbinders), we performed short Binding energies were computed using the FPB pro- molecular dynamics (MD) simulations to identify key gram implemented in Rosetta version 3.6 with scoring interactions or disruptive elements that influence peptide function ref2015.61 We generated models of peptide- binding. Simulations were performed in NAMD using the protein complexes starting with structure 1LB6 CHARMM36m force field.64,65 Each of 89 TRAF6-peptide (chains A, TRAF6 MATH domain, and B, CD40*), first complexes was solvated with a 15 Å pad of TIP3P water relaxing the structure with the Rosetta Fast-Relaxation (resulting in a final simulation box of ≈80,000 atoms). protocol to remove internal clashes and any angle viola- Simulations were performed at a constant pressure of tions in the receptor and CD40*. We then introduced 1 atm and temperature of 300 K, a non-bonded cut-off of point mutations into the CD40* peptide, keeping the 12 Å, rigid bonds between heavy atoms and hydrogen backbone atoms fixed and optimizing the sidechain con- atoms,66 and particle-mesh Ewald (PME) long-range elec- formations of mutated residues using the fixed-backbone trostatics.67 All complexes were first subjected to 1,000 design package with Resfile flag.62 Next, the Rosetta FPB energy minimization steps. Relaxed models were then module was used to sample 100 variations of the docking equilibrated for 50 ns using a time step of 2 fs with all Ca 1469896x, 2022, 11, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/pro.4429 by Mit Libraries Serials & Journa, Wiley Online Library on [12/12/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License 16 of 19 HALPIN ET AL. atoms restrained by a 10 kcal mol1 Å2 spring constant. (supporting); writing – original draft (supporting); writing – Finally, 100 ns production runs were done using review and editing (supporting). Federica Rigoldi: Con- ACEMD,68 with non-bonded cut-off and PME parameters ceptualization (supporting); data curation (equal); formal set as in the equilibration phase, and the time step analysis (equal); funding acquisition (supporting); investi- increased to 4 fs. To prevent protein diffusion in the water gation (equal); methodology (equal); project administration box, a restraining spring constant (5 kcal mol1 Å2) was (equal); validation (equal); visualization (equal); writing – applied to all Cα atoms of the protein more than 15 Å original draft (equal); writing – review and editing (sup- from the peptide-binding pocket. porting). Venkatesh Sivaraman: Data curation (support- Structures from the production runs were analyzed to ing); methodology (supporting); software (supporting). determine root mean square deviations (RMSD), root Avinoam Singer: Investigation (supporting); project mean square fluctuations (RMSF), and the presence/ administration (supporting); validation (supporting); writ- absence of specific interactions (hydrogen bonds, salt- ing – review and editing (supporting). Amy Keating: Con- bridges) using a Donor (D)-to Acceptor (A) distance cutoff ceptualization (lead); funding acquisition (lead); project of 3.2 Å; hydrogen bonds were additionally required to administration (equal); resources (lead); supervision (lead); have an A–D–H angle of <30. We also checked for struc- writing – original draft (equal); writing – review and editing turally favorable aromatic sidechain arrangements. In par- (equal). ticular, cation-pi interactions were defined using the distance between the indole/phenyl group centroid and FUNDING INFORMATION the guanidium centroid or amino group for Arg/Lys, Research reported in this publication was supported by respectively, and the angle between the respective planes. the National Institute of General Medical Sciences of the The angle was defined between the normal vectors to the National Institutes of Health under Awards F32 planes of the sidechain rings, the guanidium group, or the GM137510 to Jackson C. Halpin, F32 GM114959 to Dus- positively charged groups. To qualify as cation-pi interac- tin Whitney and R01 GM129007 to Amy E. Keating, Avi- tion, the distance had to be below 5.5 A. If the sidechains noam Singer received support from National Institute of had an angle between 45 and 135, the cation-pi interac- General Medical Sciences training award T32 GM007287. tion was defined as T-shaped, otherwise as stacked.69,70 Jackson C. Halpin was also supported by an award from We applied a similar definition for pi–pi interaction, set- the Ludwig Center for Molecular Oncology at MIT. Fed- ting the distance threshold between the centroids of the erica Rigoldi was supported by a Progetto Roberto Rocca two aromatic rings to 7 Å, and the angle range between Post-doctoral Fellowship. This work was supported in 75 and 90 for T-shaped or < 15 for stacked (parallel dis- part by the Koch Institute Support (core) Grant placed or vertical).71–73 MATH domain charge distribu- P30-CA14051 from the National Cancer Institute. The tions for Figure 6 were computed using the Coulombic content herein is solely the responsibility of the authors electrostatic potential (ESP) tool in ChimeraX.74 and does not represent the official views of any of the ACKNOWLEDGEMENTS funding organizations. We thank the Koch Institute's Robert A. Swanson (1969) Biotechnology Center for technical support, specif- DATA AVAILABILITY STATEMENT ically for peptide synthesis and flow cytometry expertise The data that support the findings of this study are and services. We also thank the MIT Biophysical Instru- openly available on Github at https://github.com/ mentation Facilty for access to instrumentation, and jacksonh1/TRAF6_screen and figshare (https://doi.org/ acknowledge use of the MIT Engaging and C3DDB com- 10.6084/m9.figshare.20485914.v3). puting clusters. ORCID AUTHOR CONTRIBUTIONS Jackson C. Halpin https://orcid.org/0000-0003-4085- Jackson Clark Halpin: Conceptualization (supporting); 1064 data curation (equal); formal analysis (equal); funding Amy E. Keating https://orcid.org/0000-0003-4074-8980 acquisition (supporting); investigation (equal); methodol- ogy (equal); project administration (equal); software (lead); REFERENCES validation (equal); visualization (equal); writing – original – 1. Gouw M, Michael S, Samano-Sanchez H, et al. The eukaryoticdraft (equal); writing review and editing (equal). Dustin linear motif resource – 2018 update. Nucleic Acids Res. 2018; Whitney: Conceptualization (supporting); data curation 46:D428–D434. https://doi.org/10.1093/nar/gkx1077. (supporting); formal analysis (supporting); funding acquisi- 2. Tonikian R, Zhang Y, Sazinsky SL, et al. A specificity map for tion (supporting); investigation (equal); methodology the PDZ domain family. PLoS Biol. 2008;6:e239. https://doi. (equal); project administration (equal); visualization org/10.1371/journal.pbio.0060239. 1469896x, 2022, 11, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/pro.4429 by Mit Libraries Serials & Journa, Wiley Online Library on [12/12/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License HALPIN ET AL. 17 of 19 3. Foight GW, Keating AE. Comparison of the peptide binding 18. Zapata JM, Lefebvre S, Reed JC. Targeting TRAfs for therapeu- preferences of three closely related TRAF paralogs: TRAF2, tic intervention. Adv Exp Med Biol. 2007;597:188–201. https:// TRAF3, and TRAF5. Protein Sci. 2016;25:1273–1289. https:// doi.org/10.1007/978-0-387-70630-6_15. doi.org/10.1002/pro.2881. 19. Ye H, Arron JR, Lamothe B, et al. Distinct molecular mecha- 4. Ivarsson Y, Arnold R, McLaughlin M, et al. Large-scale interac- nism for initiating TRAF6 signalling. Nature. 2002;418:443– tion profiling of PDZ domains through proteomic peptide- 447. https://doi.org/10.1038/nature00888. phage display using human and viral phage peptidomes. Proc 20. Zapata JM, Martínez-García V, Lefebvre S. Phylogeny of the Natl Acad Sci U S A. 2014;111:2542–2547. https://doi.org/10. TRAF/MATH domain. Adv Exp Med Biol. 2007;597:1–24. 1073/pnas.1312296111. https://doi.org/10.1007/978-0-387-70630-6_1. 5. Zhang Y, Yeh S, Appleton BA, et al. Convergent and divergent 21. Marinis JM, Homer CR, McDonald C, Abbott DW. A novel ligand specificity among PDZ domains of the LAP and zonula motif in the Crohn's disease susceptibility protein, NOD2, occludens (ZO) families. J Biol Chem. 2006;281:22299–22311. allows TRAF4 to down-regulate innate immune responses. https://doi.org/10.1074/jbc.M602902200. J Biol Chem. 2011;286:1938–1950. https://doi.org/10.1074/jbc. 6. Ueki Y, Kruse T, Weisser MB, et al. A consensus binding motif M110.189308. for the PP4 protein phosphatase. Mol Cell. 2019;76:953–964.e6. 22. Chung JY, Park YC, Ye H, Wu H. All TRAFs are not created https://doi.org/10.1016/j.molcel.2019.08.029. equal: Common and distinct molecular mechanisms of TRAF- 7. Ichikawa DM, Corbi-Verge C, Shen MJ, et al. A multireporter mediated signal transduction. J Cell Sci. 2002;115:679–688. bacterial 2-hybrid assay for the high-throughput and dynamic https://doi.org/10.1242/jcs.115.4.679. assay of PDZ domain-peptide interactions. ACS Synth Biol. 23. Shi Z, Zhang Z, Zhang Z, et al. Structural insights into mito- 2019;8:918–928. https://doi.org/10.1021/acssynbio.8b00499. chondrial antiviral signaling protein (MAVS)-tumor necrosis 8. Walsh MC, Lee J, Choi Y. Tumor necrosis factor receptor- asso- factor receptor-associated factor 6 (TRAF6) signaling. J Biol ciated factor 6 (TRAF6) regulation of development, function, Chem. 2015;290:26811–26820. https://doi.org/10.1074/jbc. and homeostasis of the immune system. Immunol Rev. 2015; M115.666578. 266:72–92. https://doi.org/10.1111/imr.12302. 24. Huang W-C, Liao J-H, Hsiao T-C, et al. Binding and enhanced 9. Wu H, Arron JR. TRAF6, a molecular bridge spanning adaptive binding between key immunity proteins TRAF6 and TIFA. immunity, innate immunity and osteoimmunology. Bioessays. Chembiochem. 2019;20:140–146. https://doi.org/10.1002/cbic. 2003;25:1096–1105. https://doi.org/10.1002/bies.10352. 201800436. 10. Xie P. TRAF molecules in cell signaling and in human diseases. 25. Sato S, Sugiyama M, Yamamoto M, et al. Toll/IL-1 receptor J Mol Signal. 2013;8:7. https://doi.org/10.1186/1750-2187-8-7. domain-containing adaptor inducing IFN-β (TRIF) associates 11. Dainichi T, Matsumoto R, Mostafa A, Kabashima K. Immune with TNF receptor-associated factor 6 and TANK-binding control by TRAF6-mediated pathways of epithelial cells in the kinase 1, and activates two distinct transcription factors, NF-κB EIME (epithelial immune microenvironment). Front Immunol. and IFN-regulatory Factor-3, in the toll-like receptor signaling. 2019;10:1107. https://doi.org/10.3389/fimmu.2019.01107. J Immunol. 2003;171:4304–4310. https://doi.org/10.4049/ 12. Zarzycka B, Seijkens T, Nabuurs SB, et al. Discovery of small jimmunol.171.8.4304. molecule CD40–TRAF6 inhibitors. J Chem Inf Model. 2015;55: 26. Jiang Z, Mak TW, Sen G, Li X. Toll-like receptor 3-mediated 294–307. https://doi.org/10.1021/ci500631e. activation of NF-kappaB and IRF3 diverges at toll-IL-1 receptor 13. Abdullah M, Berthiaume JM, Willis MS. Tumor necrosis factor domain-containing adapter inducing IFN-beta. Proc Natl Acad receptor-associated factor 6 as a nuclear factor kappa B- Sci U S A. 2004;101:3533–3538. https://doi.org/10.1073/pnas. modulating therapeutic target in cardiovascular diseases: At 0308496101. the heart of it all. Transl Res. 2018;195:48–61. https://doi.org/ 27. Tsukamoto N, Kobayashi N, Azuma S, Yamamoto T, Inoue J. 10.1016/j.trsl.2017.10.012. Two differently regulated nuclear factor kappaB activation 14. Antonios C, Tom S, Barbara Z, et al. Blocking pathways triggered by the cytoplasmic tail of CD40. Proc Natl CD40-TRAF6 signaling is a therapeutic target in obesity- Acad Sci U S A. 1999;96:1234–1239. https://doi.org/10.1073/ associated insulin resistance. Proceedings of the National pnas.96.4.1234. Academy of Sciences. 2014;111:2686–2691. https://doi.org/ 28. Pullen SS, Dang TT, Crute JJ, Kehry MR. CD40 signaling 10.1073/pnas.1400419111. through tumor necrosis factor receptor-associated factors 15. Bosch L, de Haan J, Seijkens T, et al. Small molecule-mediated (TRAFs). Binding site specificity and activation of downstream inhibition of CD40-TRAF6 reduces adverse cardiac remodelling pathways by distinct TRAFs. J Biol Chem. 1999;274:14246– in pressure overload induced heart failure. Int J Cardiol. 2019; 14254. https://doi.org/10.1074/jbc.274.20.14246. 279:141–144. https://doi.org/10.1016/j.ijcard.2018.12.076. 29. Baud V, Liu Z-G, Bennett B, Suzuki N, Xia Y, Karin M. Signal- 16. Poblenz AT, Jacoby JJ, Singh S, Darnay BG. Inhibition of ing by proinflammatory cytokines: Oligomerization of TRAF2 RANKL-mediated osteoclast differentiation by selective TRAF6 and TRAF6 is sufficient for JNK and IKK activation and target decoy peptides. Biochem Biophys Res Commun. 2007;359:510– gene induction via an amino-terminal effector domain. Genes 515. https://doi.org/10.1016/j.bbrc.2007.05.151. Dev. 1999;13:1297–1308. 17. Seijkens TTP, van Tiel CM, Kusters PJH, et al. Targeting 30. Hwang T, Parker SS, Hill SM, et al. Native proline-rich motifs CD40-induced TRAF6 signaling in macrophages reduces ath- exploit sequence context to target Actin-remodeling Ena/VASP erosclerosis. J Am Coll Cardiol. 2018;71:527–542. https://doi. protein ENAH. Elife. 2022;11:e70680. https://doi.org/10.7554/ org/10.1016/j.jacc.2017.11.055. eLife.70680. 1469896x, 2022, 11, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/pro.4429 by Mit Libraries Serials & Journa, Wiley Online Library on [12/12/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License 18 of 19 HALPIN ET AL. 31. Rice JJ, Daugherty PS. Directed evolution of a biterminal bacte- Receptor-1 Signaling complex. Immunity. 1996;4:387–396. rial display scaffold enhances the display of diverse peptides. https://doi.org/10.1016/S1074-7613(00)80252-6. Protein Eng Des Sel. 2008;21:435–442. https://doi.org/10.1093/ 47. Meylan E, Burns K, Hofmann K, et al. RIP1 is an essential protein/gzn020. mediator of toll-like receptor 3–induced NF-κB activation. Nat 32. Rice JJ, Schohn A, Bessette PH, Boulware KT, Daugherty PS. Immunol. 2004;5:503–507. https://doi.org/10.1038/ni1061. Bacterial display using circularly permuted outer membrane 48. Pérez-Pérez A, Sanchez-Jiménez F, Vilariño-García T, Sanchez- protein OmpX yields high affinity peptide ligands. Protein Sci. Margalet V. Role of leptin in inflammation and vice versa. Int J 2006;15:825–836. https://doi.org/10.1110/ps.051897806. Mol Sci. 2020;21:5887. https://doi.org/10.3390/ijms21165887. 33. Raveh B, London N, Schueler-Furman O. Sub-angstrom model- 49. Davey NE, Cyert MS, Moses AM. Short linear motifs - ex nihilo ing of complexes between flexible peptides and globular pro- evolution of protein regulation. Cell Commun Signal. 2015;13: teins, proteins: Structure. Funct Bioinformat. 2010;78:2029– 43. https://doi.org/10.1186/s12964-015-0120-z. 2040. https://doi.org/10.1002/prot.22716. 50. Luck K, Travé G. Phage display can select over-hydrophobic 34. MacArthur MW, Thornton JM. Influence of proline residues sequences that may impair prediction of natural domain- on protein conformation. J Mol Biol. 1991;218:397–412. https:// peptide interactions. Bioinformatics. 2011;27:899–902. https:// doi.org/10.1016/0022-2836(91)90721-H. doi.org/10.1093/bioinformatics/btr060. 35. O'Shea JP, Chou MF, Quader SA, Ryan JK, Church GM, 51. Davey NE, Van Roey K, Weatheritt RJ, et al. Attributes of short Schwartz D. pLogo: a probabilistic approach to visualizing linear motifs. Mol Biosyst. 2012;8:268–281. https://doi.org/10. sequence motifs. Nat Methods. 2013;10:1211–1212. https://doi. 1039/C1MB05231D. org/10.1038/nmeth.2646. 52. Van Roey K, Uyar B, Weatheritt RJ, et al. Short linear motifs: 36. Krystkowiak I, Davey NE. SLiMSearch: A framework for Ubiquitous and functionally diverse protein interaction mod- proteome-wide discovery and annotation of functional modules ules directing cell regulation. Chem Rev. 2014;114:6733–6778. in intrinsically disordered regions. Nucleic Acids Res. 2017;45: https://doi.org/10.1021/cr400585q. W464–W469. https://doi.org/10.1093/nar/gkx238. 53. Ottinger EA, Botfield MC, Shoelson SE. Tandem SH2 domains 37. Mészaros B, Erdo}s G, Dosztanyi Z. IUPred2A: Context- confer high specificity in tyrosine kinase signaling. J Biol dependent prediction of protein disorder as a function of redox Chem. 1998;273:729–735. https://doi.org/10.1074/jbc.273.2.729. state and protein binding. Nucleic Acids Res. 2018;46:W329– 54. Lin Z, Yang Z, Xie R, Ji Z, Guan K, Zhang M. Decoding WW W337. https://doi.org/10.1093/nar/gky384. domain tandem-mediated target recognitions in tissue growth 38. Tunyasuvunakool K, Adler J, Wu Z, et al. Highly accurate pro- and cell polarity. Elife. 2019;8:e49439. https://doi.org/10.7554/ tein structure prediction for the human proteome. Nature. eLife.49439. 2021;596:590–596. https://doi.org/10.1038/s41586-021-03828-1. 55. Dodson EJ, Fishbain-Yoskovitz V, Rotem-Bamberger S, 39. Jumper J, Evans R, Pritzel A, et al. Highly accurate protein Schueler-Furman O. Versatile communication strategies structure prediction with AlphaFold. Nature. 2021;596:583– among tandem WW domain repeats. Exp Biol Med 589. https://doi.org/10.1038/s41586-021-03819-2. (Maywood). 2015;240:351–360. https://doi.org/10.1177/ 40. Varadi M, Anyango S, Deshpande M, et al. AlphaFold protein 1535370214566558. structure database: Massively expanding the structural cover- 56. Kenrick SA, Daugherty PS. Bacterial display enables efficient age of protein-sequence space with high-accuracy models. and quantitative peptide affinity maturation. Protein Eng Des Nucleic Acids Res. 2022;50:D439–D444. https://doi.org/10. Sel. 2010;23:9–17. https://doi.org/10.1093/protein/gzp065. 1093/nar/gkab1061. 57. Angelini A, Chen TF, de Picciotto S, et al. Protein engineering 41. Ashburner M, Ball CA, Blake JA, et al. Gene ontology: Tool for and selection using yeast surface display. In: Liu B, editor. the unification of biology. Nat Genet. 2000;25:25–29. https:// Yeast surface display: Methods, protocols, and applications. doi.org/10.1038/75556. New York, NY: Springer New York, 2015; p. 3–36. https://doi. 42. The Gene Ontology Consortium. The gene ontology resource: org/10.1007/978-1-4939-2748-7_1. Enriching a GOld mine. Nucleic Acids Res. 2021;49:D325– 58. KaleidaGraph, (2022). www.synergy.com. D334. https://doi.org/10.1093/nar/gkaa1113. 59. B. Brian, BBMap short-read aligner, and other bioinformatics 43. Alanis-Lobato G, Andrade-Navarro MA, Schaefer MH. HIPPIE tools, (2022). sourceforge.net/projects/bbmap/. v2.0: Enhancing meaningfulness and reliability of protein– 60. Tareen A, Kinney JB. Logomaker: Beautiful sequence logos in protein interaction networks. Nucleic Acids Res. 2017;45: python. Bioinformatics. 2020;36:2272–2274. https://doi.org/10. D408–D414. https://doi.org/10.1093/nar/gkw985. 1093/bioinformatics/btz921. 44. Maruyama Y, Yamada M, Takahashi K, Yamada M. Ubiqui- 61. Alford RF, Leaver-Fay A, Jeliazkov JR, et al. The Rosetta all- tin ligase Kf-1 is involved in the endoplasmic reticulum- atom energy function for macromolecular Modeling and associated degradation pathway. Biochem Biophys Res Com- design. J Chem Theory Comput. 2017;13:3031–3048. https:// mun. 2008;374:737–741. https://doi.org/10.1016/j.bbrc.2008. doi.org/10.1021/acs.jctc.7b00125. 07.126. 62. Brian K, Gautam D, Ireton Gregory C, Gabriele V, Stoddard 45. Yuan J, Amin P, Ofengeim D. Necroptosis and Barry L, David B. Design of a novel globular protein fold with RIPK1-mediated neuroinflammation in CNS diseases. Nat Rev atomic-level accuracy. Science. 2003;302:1364–1368. https:// Neurosci. 2019;20:19–33. https://doi.org/10.1038/s41583-018- doi.org/10.1126/science.1089427. 0093-1. 63. Humphrey W, Dalke A, Schulten K. VMD: Visual molecular 46. Hsu H, Huang J, Shu H-B, Baichwal V, Goeddel DV. TNF- dynamics. J Mol Graph. 1996;14:33–38. https://doi.org/10. dependent recruitment of the protein kinase RIP to the TNF 1016/0263-7855(96)00018-5. 1469896x, 2022, 11, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/pro.4429 by Mit Libraries Serials & Journa, Wiley Online Library on [12/12/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License HALPIN ET AL. 19 of 19 64. Phillips JC, Braun R, Wang W, et al. Scalable molecular 71. Marsili S, Chelli R, Schettino V, Procacci P. Thermodynamics dynamics with NAMD. J Comput Chem. 2005;26:1781–1802. of stacking interactions in proteins. Phys Chem Chem Phys. https://doi.org/10.1002/jcc.20289. 2008;10:2673–2685. https://doi.org/10.1039/B718519G. 65. Huang J, Rauscher S, Nawrocki G, et al. CHARMM36m: An 72. McGaughey GB, Gagné M, Rappé AK. π-Stacking interactions: improved force field for folded and intrinsically disordered pro- ALIVE AND WELL IN PROTEINS*. J Biol Chem. 1998;273: teins. Nat Methods. 2017;14:71–73. https://doi.org/10.1038/ 15458–15463. https://doi.org/10.1074/jbc.273.25.15458. nmeth.4067. 73. Burley SK, Petsko GA. Aromatic-aromatic interaction: A mech- 66. Ryckaert J-P, Ciccotti G, Berendsen HJC. Numerical integra- anism of protein structure stabilization. Science. 1985;229:23– tion of the cartesian equations of motion of a system with con- 28. https://doi.org/10.1126/science.3892686. straints: Molecular dynamics of n-alkanes. J Comput Phys. 74. Pettersen EF, Goddard TD, Huang CC, et al. Structure visuali- 1977;23:327–341. https://doi.org/10.1016/0021-9991(77) zation for researchers, educators, and developers. Protein Sci. 90098-5. 2021;30:70–82. https://doi.org/10.1002/pro.3943. 67. Darden T, York D, Pedersen L. Particle mesh Ewald: An Nlog (N) method for Ewald sums in large systems. J Chem Phys. 1993;98:10089–10092. https://doi.org/10.1063/1.464397. SUPPORTING INFORMATION 68. Harvey MJ, Giupponi G, Fabritiis GD. ACEMD: Accelerating Additional supporting information can be found online biomolecular dynamics in the microsecond time scale. J Chem in the Supporting Information section at the end of this Theory Comput. 2009;5:1632–1639. https://doi.org/10.1021/ article. ct9000685. 69. Gallivan JP, Dougherty DA. Cation-π interactions in structural biology. Proc Natl Acad Sci. 1999;96:9459–9464. https://doi. How to cite this article: Halpin JC, Whitney D, org/10.1073/pnas.96.17.9459. Rigoldi F, Sivaraman V, Singer A, Keating AE. 70. Chipot C, Jaffe R, Maigret B, Pearlman DA, Kollman PA. Molecular determinants of TRAF6 binding Benzene dimer: A good model for ππ interactions in pro- specificity suggest that native interaction partners teins? A comparison between the benzene and the toluene are not optimized for affinity. Protein Science. dimers in the gas phase and in an aqueous solution. J Am Chem Soc. 1996;118:11217–11224. https://doi.org/10.1021/ 2022;31(11):e4429. https://doi.org/10.1002/pro.4429 ja961379l. 1469896x, 2022, 11, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/pro.4429 by Mit Libraries Serials & Journa, Wiley Online Library on [12/12/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License