Decision letter: Epistatic mutations in PUMA BH3 drive an alternate binding mode to potently and selectively inhibit anti-apoptotic Bfl-1

Abstract

Article Figures and data Abstract Introduction Results Discussion Materials and methods Data availability References Decision letter Author response Article and author information Metrics Abstract Overexpression of anti-apoptotic Bcl-2 family proteins contributes to cancer progression and confers resistance to chemotherapy. Small molecules that target Bcl-2 are used in the clinic to treat leukemia, but tight and selective inhibitors are not available for Bcl-2 paralog Bfl-1. Guided by computational analysis, we designed variants of the native BH3 motif PUMA that are > 150-fold selective for Bfl-1 binding. The designed peptides potently trigger disruption of the mitochondrial outer membrane in cells dependent on Bfl-1, but not in cells dependent on other anti-apoptotic homologs. High-resolution crystal structures show that designed peptide FS2 binds Bfl-1 in a shifted geometry, relative to PUMA and other binding partners, due to a set of epistatic mutations. FS2 modified with an electrophile reacts with a cysteine near the peptide-binding groove to augment specificity. Designed Bfl-1 binders provide reagents for cellular profiling and leads for developing enhanced and cell-permeable peptide or small-molecule inhibitors. https://doi.org/10.7554/eLife.25541.001 Introduction Anti-apoptotic members of the Bcl-2 family are broadly recognized as promising cancer therapeutic targets. Human anti-apoptotic proteins Bcl-2, Bcl-xL, Bcl-w, Mcl-1 and Bfl-1 have a globular, helical fold and function by binding to short, α-helical Bcl-2 homology 3 (BH3) motifs in pro-apoptotic proteins, as shown in Figure 1A. Competition for binding among BH3-containing proteins regulates mitochondrial outer membrane permeabilization (MOMP), which is an irreversible step toward caspase activation and cell death. The appropriate balance of interactions between pro-survival and pro-death Bcl-2 family members in healthy cells is often disrupted in cancer cells, where overexpression of anti-apoptotic Bcl-2 proteins can promote oncogenesis and confer resistance to chemotherapeutic agents (Opferman, 2016). Figure 1 with 3 supplements see all Download asset Open asset Computational design of a library of PUMA BH3 variants selective for Bfl-1. (A) PUMA BH3 is pan-selective; the design objective was a peptide that binds tightly only to Bfl-1. (B) Sequence of PUMA BH3 showing the heptad numbering convention used in this paper. (C) Overview of the computational library design procedure. (D–E) Scores for members of three libraries designed to target Bfl-1 (blue), Mcl-1 (green) or Bcl-xL (red): (D) PSSMSPOT scores, (E) STATIUM z-scores. (F–I) The affinities of library peptides for different Bcl-2 proteins were predicted to be strongly correlated. (F) PSSMSPOT scores for binding to Bcl-xL versus Bfl-1, (G) PSSMSPOT scores for binding to Mcl-1 versus Bfl-1, (H) STATIUM z-scores for binding to Bcl-xL versus Bfl-1, (I) STATIUM z-scores for binding to Mcl-1 versus Bfl-1. For (D–I), each point represents one peptide sequence and higher scores correspond to higher predicted affinities for the indicated target. Points on the dashed line have the same low specificity as PUMA BH3 (which is shown with a black open circle). https://doi.org/10.7554/eLife.25541.002 There has been considerable progress developing BH3 mimetic peptides and small molecules to inhibit the function of anti-apoptotic Bcl-2 proteins by blocking their interactions. One outstanding example is the small molecule venetoclax, which targets Bcl-2 and was recently approved by the FDA for treatment of chronic lymphocytic leukemia (Souers et al., 2013; Roberts et al., 2016). A major challenge in developing venetoclax was achieving specificity, which is important because Bcl-2 family members support survival of healthy cells. For example, the small molecule ABT-263 inhibits both Bcl-2 and Bcl-xL, but Bcl-xL cross-reactivity leads to dose-limiting thrombocytopenia (Rudin et al., 2012; Roberts et al., 2012; Schoenwaelder et al., 2011). In the laboratory, highly selective inhibitors of anti-apoptotic proteins are used for profiling experiments that can establish which anti-apoptotic proteins are essential for cancer cell survival in individual patients and predict chemotherapeutic response in vivo (Ryan et al., 2010; Deng et al., 2007; Montero et al., 2015). There has been progress toward creating a panel of reagents specific for each mammalian anti-apoptotic protein that can advance such diagnostic assays. Useful reagents for this purpose include peptides and small molecules that are selective for Mcl-1 (Foight et al., 2014; Kotschy et al. 2016) or Bcl-xL (Dutta et al., 2015; Lessene et al., 2013). The role of anti-apoptotic protein Bfl-1 in cancer is less characterized than that of Mcl-1 or Bcl-xL, but many lines of evidence suggest that Bfl-1 is also a critical target. In melanoma, Bfl-1 overexpression confers resistance to BRAF inhibitors, and siRNA-mediated knockdown of Bfl-1 induces cell death in melanoma cell lines but not non-malignant cells (Hind et al., 2015; Haq et al., 2013; Senft et al., 2012). Mis-regulation of Bfl-1 is also implicated in hematological malignancies, where elevated levels of Bfl-1 confer resistance to common chemotherapeutic agents. Bfl-1 knockdown suppresses resistance and sensitizes malignant B-cells to chemotherapy (Brien et al., 2007). Bfl-1 expression can also counteract the effects of inhibitors of other anti-apoptotic family members (e.g. Mcl-1, Bcl-2) in leukemia and lymphoma (Fan et al., 2010). Bfl-1 mRNA is over-expressed in myriad malignancies including solid tumor samples from breast, colon, ovary and prostate tissues (Beverly and Varmus, 2009). Thus, Bfl-1 is an intriguing therapeutic target and biomarker for resistance to cytotoxic anticancer drugs. Identifying Bfl-1-selective interaction inhibitors has proven difficult. Small molecules must compete with an extended protein-peptide interface, and developing small-molecule inhibitors of Bcl-xL, Bcl-2 and Mcl-1 required years of work, guided by intensive NMR studies of fragment binding (Souers et al., 2013; Oltersdorf et al., 2005; Kotschy et al., 2016). Screening has identified small-molecule inhibitors of Bfl-1, but these compounds have IC50 values in the high nanomolar to low micromolar range and exhibit only modest specificity for Bfl-1 relative to other Bcl-2 family members (Mathieu et al., 2014; Zhai et al., 2012; Cashman et al., 2010; Zhai et al., 2008). Recently, helical bundle proteins that incorporate a BH3 motif have been designed to inhibit Bfl-1 and other anti-apoptotic proteins. These proteins are tight and selective binders, but their function relies on them being folded, and delivering proteins of molecular weight >13 kDa into cells is problematic given current technologies (Berger et al., 2016). An attractive strategy for inhibiting Bcl-2 family proteins is to develop short peptides that mimic the interaction geometry of native Bcl-2 protein complexes (Figure 1A). Screening BH3-like peptide libraries previously led to identification of a molecule with ~50 nM affinity for Bfl-1 and 30-fold specificity for Bfl-1 over Mcl-1, (Dutta et al., 2013), but this peptide was not shown to induce mitochondrial depolarization in cell-based assays. Identifying Bfl-1 selective peptides is complicated by the extremely large sequence space of short BH3-like helical binders. There are more than 1029 possible peptides of length 23 residues. This sequence space is too large to exhaustively search experimentally. Furthermore, the BH3 motif is a weak motif (only three positions are strongly conserved) that does little to restrict possible binders. Another confounding factor is that Bfl-1 interacts with fewer BH3-like peptides than other anti-apoptotic Bcl-2 family paralogs do (Foight and Keating, 2015; DeBartolo et al., 2014), and no native interaction partners are known to be selective for Bfl-1, suggesting that there may be limited opportunities for achieving specificity. The results described here showcase our computational/experimental roadmap for designing selective peptide inhibitors. We used computational models to design a focused library of ~107 candidate binders and screened it to identify three peptides, FS1, FS2 and FS3, that bind tightly and specifically to Bfl-1. Mutational studies and high-resolution structures revealed that the high specificity comes from a BH3 binding mode that is markedly different from what has been seen in prior structures of Bfl-1:BH3 complexes (Herman et al., 2008; Smits et al., 2008). Importantly, FS1, FS2 and FS3 are specific in BH3 profiling, an assay that tests for MOMP in cells. Subsequent rational introduction of an acrylamide moiety to covalently react with Bfl-1 further enhanced Bfl-1 inhibitor specificity. FS1, FS2, FS3 and their chemical derivatives provide new reagents with utility for studying Bfl-1 biology and a launching point for developing Bfl-1 targeting therapeutics. Results Computational analysis prioritizes mutations for targeted library design To reduce the enormous space of possible 23-mer sequences to <107 candidates that could be tested experimentally, we used computational modeling to design focused combinatorial libraries. We first scored mutations throughout the BH3 motif using: (1) a position-specific scoring matrix (PSSM) derived from SPOT peptide array data (PSSMSPOT) and (2) STATIUM, a structure-based statistical potential that previously showed good performance evaluating Bcl-2 protein binding to BH3-like peptides (Figure 1B,C) (DeBartolo et al., 2014, 2012). Mutations were modeled in the BIM BH3 motif, with the intention of testing the mutations in the context of both BIM and PUMA BH3 motifs. These two BH3-only proteins, as well as tBID, interact tightly with Bfl-1. BIM and PUMA bind with low-nanomolar affinity to Bfl-1, but also to anti-apoptotic paralogs Bcl-2, Bcl-xL, Mcl-1 and Bcl-w (Dutta et al., 2013; Foight and Keating, 2015). Thus, our design challenge was to introduce mutations that eliminate off-target binding without destabilizing Bfl-1 binding. Bfl-1 shares 38% binding-groove sequence identity with Mcl-1% and 30% binding-groove identity to Bcl-xL. Bcl-2 and Bcl-w are closely related to Bcl-xL, with 60% sequence identity in the binding groove (Foight and Keating, 2015). To model cross-reactivity, we compared how mutations in BIM were predicted to affect binding to Bfl-1 relative to Bcl-xL and Mcl-1, for which high-quality structures of complexes are available. The predicted binding scores of diverse sequences for the three proteins were highly correlated, and most single mutations were predicted to weaken Bfl-1 binding compared to the wild-type sequence (Figure 1—figure supplement 1). Mutational scoring identified promising positions for introducing sequence variation (helix positions are defined in Figure 1B above the sequence of PUMA BH3). Bfl-1, Bcl-xL and Mcl-1 were predicted to have distinct residue preferences at conserved hydrophobic positions 3d and 4a, consistent with previous observations (Dutta et al., 2010). Many mutations at position 4e were predicted to be strongly Bfl-1 selective, which is supported by the observation that peptide binding by both Bcl-xL, and Mcl-1 is weakened by mutations at this position (Boersma et al., 2008). Mutations at positions 2a and 3g were also predicted to confer Bfl-1 specificity. In native BH3 motifs, these sites are generally occupied by small charged or polar residues that can form hydrogen bonds/salt-bridges with Bcl-xL and Mcl-1 groups that are absent in Bfl-1. Finally, the region around sites 2e and 2g has local structural differences in Bfl-1, Mcl-1 and Bcl-xL. We used in-house software to select degenerate codons at variable sites that optimized the predicted Bfl-1 binding affinity and specificity and that provided chemical diversity in the resulting library (Dutta et al., 2013; Foight et al., 2017). The final library design included >6.8*106 unique sequences (Figure 1—figure supplement 2), most of which were predicted to be Bfl-1 selective by PSSMSPOT and STATIUM (Figure 1F–I). As a control, we designed similarly sized libraries to be selective for Bcl-xL and Mcl-1 (Figure 1—figure supplement 2). PSSMSPOT predicted each library to be enriched in peptides selective for the appropriate target, as shown in Figure 1D. In contrast, STATIUM predicted significantly more cross-reactivity for library members (Figure 1E, Figure 1—figure supplement 3). Experimental library screening Oligonucleotides encoding the peptide libraries designed to be specific for Bfl-1, Bcl-xL and Mcl-1 were synthesized in the context of BIM and PUMA BH3 sequences. Pooled BIM-based libraries and pooled PUMA-based libraries were then screened separately for tight and selective binding to Bfl-1. Screening the libraries designed for Mcl-1 and Bcl-xL for binding to Bfl-1, in addition to the library designed to target Bfl-1, provided an opportunity to evaluate the utility of computational library focusing. We used yeast-surface display to identify selective Bfl-1-binding peptides from our mixed libraries (Figure 2A). FACS analysis revealed that the initial libraries had a modest number of cells expressing peptides that bound to Bfl-1 at 100 nM (Figure 2B). This is consistent with predictions that less than 6.5% or 4% of the theoretical library would bind as well or better than PUMA, according to PSSMSPOT or STATIUM, respectively. Figure 2 with 5 supplements see all Download asset Open asset Experimental library screening for Bfl-1 affinity and selectivity. (A) Yeast-surface display configuration. BH3 peptides were expressed as fusions to Aga2; HA tag expression was detected with APC and Bfl-1 binding was detected with PE. (B) FACS analysis showed that only ~5% of cells in the unsorted PUMA libraries bound to Bfl-1 at 100 nM. (C) Library binding to 100 nM Bfl-1 after one round of enrichment. (D–G) Library binding to off-target proteins (100 nM) after one round of enrichment: (D) Bcl-xL, (E) Bcl-2, (F) Bcl-w, (G) Mcl-1. (H) Library binding to 100 nM Myc-tagged Bfl-1 in the presence of excess unlabeled competitor (Mcl-1, Bcl-2, Bcl-w and Bcl-xL; 1 μM each) after six rounds of enrichment. (I) Inhibition constants determined using fluorescence anisotropy for 23-residue peptides corresponding to PUMA BH3, FS1, FS2 and FS3. https://doi.org/10.7554/eLife.25541.006 Figure 2—source data 1 Data collected from competition fluorescence polarization experiments. https://doi.org/10.7554/eLife.25541.007 Download elife-25541-fig2-data1-v1.xlsx Most of the peptides that bound Bfl-1 were cross-reactive with one or more other Bcl-2 family proteins (Figure 2C–G). This cross-reactivity was expected based on the high correlation of predicted binding scores for Bfl-1, Mcl-1 and Bcl-xL and highlights the challenge of identifying specific binders (Figure 1D–E, Figure 1—figure supplement 1). Six rounds of positive, negative and/or competition FACS screening were used to isolate cells that expressed the tightest and most Bfl-1-selective peptides (Figure 2—figure supplement 1). Mcl-1, Bcl-xL, Bcl-2 and Bcl-w were included in the screen as untagged competitors. Early screening provided many Bfl-1 selective hits from the PUMA libraries, but few from the BIM libraries, so the BIM libraries were not pursued (Figure 2—figure supplement 2). After several rounds of competition screening, the PUMA library was enriched in cells displaying peptides that bound to Bfl-1 at 100 nM in the presence of 40-fold excess unlabeled competitor (Figure 2H). Fifty colonies isolated in the final round of screening were sequenced, providing 13 unique sequences: nine sequences were from the Bfl-1 specific library, two were from the Bcl-xL library, and two were from the Mcl-1 library (Figure 2—figure supplement 3). We tested three Bfl-1 selective peptides that were recovered two or more times (FS1, FS2 and FS3). FS1, FS2 and FS3 were all derived from the Bfl-1 targeted library, although FS1 also contained one mutation caused by a spurious single-base pair mutation. FS1, FS2 and FS3 each had reduced affinity for Bfl-1 relative to PUMA, but significantly increased specificity (Figure 2I and Figure 2—figure supplement 4). FS1 bound Bfl-1 with Ki = 15 nM and at least 150-fold specificity for Bfl-1 relative to Bcl-xL, Bcl-2, Bcl-w and Mcl-1. To analyze enrichment trends and to assess the success of our library design, we deep sequenced samples from the naïve pool and from pools collected after 3, 4, 5 and 6 rounds of sorting (sorting conditions are detailed in Figure 2—figure supplement 1). The naive pool was diverse and not dominated by any particular subset of sequences. In contrast, FS1 (38% of sequences, the most prevalent library member), FS2 (25% of sequences), and many other peptides from the Bfl-1 targeted library were prominent in the final screening pool. Analysis of sequential pools showed that peptides from the Bfl-1 targeted library were substantially enriched relative to peptides from the Bcl-xL and Mcl-1 targeted libraries (Figure 3A). Of the unique sequences in the final pool, 73.9% were from the Bfl-1 targeted library (Figure 3B). Figure 3 Download asset Open asset Evaluation of the library design. (A) Sequences from the Bfl-1 library were preferentially enriched during sorting. Sequences with no more than one amino-acid mutation from the Bfl-1 (blue), Mcl-1 (green), or Bcl-xL (red) targeted libraries are plotted. Other sequences are shown in magenta. (B) The large majority of unique sequences in the final pool originated from the Bfl-1 library (colors as in part A). (C–F) Comparison of PSSMSPOT and STATIUM scores for the library before and after sorting. Peptides from the final sorted pool (red dots) are superimposed on the distribution of scores for the theoretical library (blue contour plots). Points to the left of the dotted lines correspond to peptides predicted to bind more selectively to Bfl-1 than does PUMA, with respect to the indicated competitor protein (Bcl-xL in C and E, Mcl-1 in D and F). Scores for FS1, FS2 and FS3 are indicated. https://doi.org/10.7554/eLife.25541.013 Figure 3—source code 1 Deep sequencing data analysis. https://doi.org/10.7554/eLife.25541.014 Download elife-25541-fig3-code1-v1.zip We scored peptides from the Bfl-1 targeted library that passed all rounds of screening with the STATIUM and PSSMSPOT models used in library design (Figure 3C–F). Most sequences were predicted to have improved selectivity for Bfl-1 relative to PUMA (98–99% with improved specificity over Bcl-xL or Mcl-1 by PSSMSPOT, and 95% or 62% with improved specificity over Bcl-xL or Mcl-1, respectively, by STATIUM). The selected sequences were not among those predicted by either model to be the tightest or most Bfl-1 selective in the theoretical library. The binding mode of Bfl-1-selective peptides FS1, FS2 and FS3 included mutations to larger residues than those in PUMA at their N-termini (red in Figure 4B,C), and smaller residues at their C-termini (blue in Figure 4B,C). Deep sequencing of additional selective sequences supported this trend: Of 612 unique peptide sequences from the final round of sorting that originated from the Bfl-1 targeted library sequences, 364 showed this type of residue size patterning at the same sites (sequence logo in Figure 4A). Figure 4 with 1 supplement see all Download asset Open asset Epistatic mutations in PUMA confer Bfl-1 binding specificity. (A) Sequence logo of unique peptide sequences in the final sorted pool from the Bfl-1 targeted library. (B) Location of mutated sites in FS1, FS2 and FS3. Mutations at positions 2a and 2e are in red and positions 2g, 3d, 4a and 4e are in blue. (C) Structure of Bfl-1 (gray surface) bound to PUMA (green, this work) with residues at positions 2a and 2e in red and those at 2g, 3d, 4a and 4e in blue. (D) Non-additive mutational energies for PUMA/FS2 chimeric proteins indicate coupling between N- and C-terminal mutations. Data are Ki ± SD of three or more independent fluorescence anisotropy competition experiments. https://doi.org/10.7554/eLife.25541.015 Figure 4—source data 1 Data collected from competition fluorescence polarization experiments. https://doi.org/10.7554/eLife.25541.016 Download elife-25541-fig4-data1-v1.xlsx To assess whether the combination of large and small residues played a role in establishing binding specificity, we tested PUMA/FS2 chimeric peptides for binding to all five anti-apoptotic proteins. Mutating PUMA to introduce smaller residues at positions 2g, 3d, 4a and 4e differentially impaired binding to all receptors and resulted in weak yet specific binding to Bfl-1 (Figure 4—figure supplement 1). Mutating residues at the N-terminus of PUMA to larger residues at positions 2a and 2e gave a modest 2.3-fold increase in affinity for Bfl-1. But the same mutations in the context of smaller residues at positions 2g, 3d, 4a and 4e improved affinity for Bfl-1 by 28.6-fold (Figure 4D). The different effects of these mutations, when made in different contexts, indicates an energetic coupling consistent with a structural repositioning of the designed peptides in the groove of Bfl-1. To better understand the structural basis for the epistasis, we solved X-ray crystal structures of Bfl-1 bound to PUMA, at 1.33 Å resolution, and of Bfl-1 bound to FS2 at 1.2 Å resolution (Supplementary file 1). In comparison with all available X-ray structures of BH3 peptides bound to human or murine Bfl-1, PUMA and FS2 each adopt new, distinct positions in the binding groove (Figure 5A and Figure 5—figure supplement 1). FS2 is shifted 1.2 Å and rotated 17° in the binding groove compared to its parent peptide PUMA. The peptide C-terminus, which harbors the large-to-small mutations, is repositioned more dramatically than the N-terminus (Figure 5B). Despite the shifts in peptide binding geometry, the structures of Bfl-1 in these newly solved complexes are highly similar. The all-atom RMSD for residues in the binding pocket (within 5 Å of the BH3 peptide) of Bfl-1:FS2 vs. Bfl-1:PUMA is <0.7 Å and is 1.05 Å for Bfl-1:FS2 vs. Bfl-1:BIM (Herman et al., 2008). Figure 5 with 4 supplements see all Download asset Open asset High-resolution structures of PUMA and FS2 bound to human Bfl-1. (A) Binding groove of Bfl-1 (gray, surface) with PUMA (yellow) and FS2 (purple). (B) Cα- Cα shifts between FS2 and PUMA. Sites with larger/smaller residues in FS2 are indicated in red/blue. (C) The canonical Bfl-1:BH3 salt bridge between D3f and R88 is observed in the Bfl-1:PUMA complex but not the Bfl-1:FS2 complex. (D) Tryptophan at 1g is rotated into the Bfl-1 binding groove in the Bfl-1:FS2 complex and away from the binding groove in the Bfl-1:PUMA complex. (E) In contrast with the solvent exposed arginine at position 3c of the Bfl-1:PUMA complex, R3c is oriented into the BH3 binding groove in the Bfl-1:FS2 complex, forming a hydrogen bond with N51 of Bfl-1. (F) Bfl-1 targeted library sequences score better on the Bfl-1:FS2 structure than on the Bfl-1:BID structure used for the initial library design; higher scores predict tighter binding. STATIUM z-scores for the Bfl-1 targeted library are in blue. FS1, FS2 and FS3 are indicated in red and PUMA in green. (G) Sorting enriched sequences that score better on the Bfl-1:FS2 template than on the Bfl-1:BID template. STATIUM z-scores for the input Bfl-1 library are in blue and scores for sequences identified after the final round of screening are in green. https://doi.org/10.7554/eLife.25541.018 Further structural analysis showed that the Bfl-1:FS2 complex supports several key side-chain interactions that are absent in Bfl-1:PUMA and that may be important for selective binding. Surprisingly, aspartate at position 3f (D3f) in FS2, which is strongly conserved in known BH3 motifs, makes different interactions than what is observed in numerous previously solved Bcl-2 complex structures. D3f typically forms a salt bridge with arginine 88 (R88) in helix four in Bfl-1 or the corresponding arginine in Bcl-xL, Mcl-1, Bcl-w or Bcl-2 (Figure 5C). In the Bfl-1:FS2 structure, the carboxylate of D3f is shifted 5.6 Å away from the guanidinium group of R88, and is highly solvent exposed (Figure 5C). Because D3f does not form the canonical D3f:R88 interaction and is solvent exposed, we reasoned that FS2 should tolerate mutations at this site. This was confirmed by the tight binding of six peptides with alanine, serine, asparagine, glutamate, histidine or tyrosine at this position (Figure 5—figure supplement 2). Disruption of the D3f:R88 salt bridge would be expected to reduce affinity for Bfl-1 and for all of the other anti-apoptotic receptors. However, in the Bfl-1:FS2 complex this change may be partially compensated by hydrogen bonding on the opposite side of the FS2 helix between arginine at position 3c (R3c) of FS2 and asparagine 51 (N51) of Bfl-1 (Figure 5E). In Bfl-1:FS2, position 3c is positioned closer to helix 2 of Bfl-1 than in Bfl-1:PUMA, allowing R3c to fill the space left by an adjacent methionine-to-alanine mutation at 3d when it adopts this hydrogen-bonded position. N51 at this position of helix two is unique to Bfl-1 among the human anti-apoptotic proteins (Figure 5—figure supplement 3). Other structural differences between PUMA and FS2 binding are apparent near the N-terminal end of the peptide. Modeling FS2 mutations in the Bfl-1:PUMA structure suggested that the small-to-large mutation of alanine at position 2a in PUMA to the valine in FS2 would result in steric clashes with helix 4 of Bfl-1 for all backbone-dependent rotamers (Figure 5—figure supplement 4). This change is accommodated by the shift in the Bfl-1:FS2 structure. Also, a rotation of FS2 in the Bfl-1 binding groove partially buries the phenylalanine at position 1g that is solvent-exposed in the PUMA complex, which may be energetically favorable (Figure 5D). Because the altered binding mode of FS2 is expected to impact predictions made using structure-based models, we re-scored the designed Bfl-1 library on the shifted Bfl-1:FS2 structure using STATIUM. FS1 and FS2 scored much better (higher) on the shifted model than on the original model, whereas PUMA scored better on the original model (Figure 5F). Analysis of the entire pool of sequences that passed screening showed that these peptides were enriched in sequences that scored better on the shifted model, compared to the input library, consistent with our observation of size patterning in the majority of these sequences (Figure 5G). Structural analysis of off-target binding to Mcl-1 To better understand the structural basis of FS2-binding specificity, we solved the X-ray crystal structure of FS2 bound to Mcl-1 at 2.35 Å resolution. FS2 binding to Mcl-1 is >100 fold weaker than binding to Bfl-1. Similar to the way FS2 binds to Bfl-1, FS2 engages Mcl-1 in a shifted orientation relative to BIM (Figure 6A,B). As is the case for FS2 binding to Bfl-1, this shift re-positions the highly conserved aspartate at peptide position 3f to a location 4.8 Å away from Mcl-1, disrupting the canonical salt bridge with arginine 92 (Figure 6C). This disruption would be expected to reduce affinity for Mcl-1, but it doesn’t account for the specificity of FS2 for Bfl-1, because the salt bridge is lost in both complexes. There are other differences between the Bfl-1:FS2 and Mcl-1:FS2 structures that may account for some of the affinity difference. For example, R3c in FS2 forms a hydrogen bond with N51 of Bfl-1, but does not form an equivalent interaction with Mcl-1 and is instead solvent exposed (Figure 6D). In Mcl-1, there is an alanine (A55) at this site, and an adjacent histidine (H53) would be expected to clash with R3c if it adopted this conformation. The N-terminus of FS2 is also buried further into the binding groove of Bfl-1 than Mcl-1 (Figure 6E). The Bfl-1 binding groove is wider in this region than the Mcl-1 binding groove, as illustrated by aligning many Bfl-1 and Mcl-1 structures (Figure 6—figure supplement 1). This region of the groove is formed by helices 3 and 4. There is an amino acid insertion in the loop between helices 3 and 4 that is unique to Bfl-1 that likely contributes to the distinct structural environment of Bfl-1 in this region (Figure 5—figure supplement 3). Figure 6 with 1 supplement see all Download asset Open asset Crystal structure of FS2 bound to human Mcl-1. (A) Binding groove of Mcl-1 (blue, surface) with BIM (yellow, 2PQK [Fire et al., 2010]) and FS2 (purple). (B) Cα- Cα shifts between FS2 and BIM when bound to Mcl-1. (C) The canonical Bfl-1:BH3 salt bridge between D3f and R92, formed in Mcl-1:BIM, is not observed in the Mcl-1:FS2 complex. (D) In contrast with the arginine at position 3c of the Bfl-1:FS2 complex, which makes packing and hydrogen-bond interactions the interface, R3c is oriented away from the BH3 binding groove in the Mcl-1:FS2 complex. (E) The Mcl-1 binding groove between helix 3 and helix four is narrower than the Bfl-1 binding groove, and the N-terminus of FS2 is shifted in the Mcl-1:FS2 structure in comparison with the Bfl-1:FS2 complex. https://doi.org/10.7554/eLife.25541.023 Biological activity of designed Bfl-1 inhibitors We tested our designed peptides for Bfl-1 selective targeting by carrying out BH3 pr