Hundreds of DARPin structures with various topologies were published in the literature and are accessible in the PDB, among which many have 4 or 5 repeated ankyrin motifs. Two DARPins evolved through ribosome display to bind BCL-W, and corresponding to PDB entries 4k5a and 4k5b (Schilling et al., 2014b), were used as known binders in this study. These known binders engage the target in a binding mode which is typical for DARPins, which consists of interactions made by the concave paratope formed by their 5 repeated ankyrin motifs (Binz et al., 2003; Kramer et al., 2010; Pluckthun, 2015; Schilling et al., 2022). By inspecting the sequences and structures of these known binders and other DARPins with available crystal structures in PDB, a common framework sequence was defined for further library expansion. The main features considered during the selection of a DARPin common framework sequence were: 1) 157 amino acids starting with DLGKK and ending with LQKAA sequences; 2) conserved regions at these N- and C-terminal ends; 3) consensus residues deemed essential for the stability of the overall fold along repeated ankyrin motifs; and 4) key residues contributing to binding along repeated ankyrin motifs. These criteria led to a single DARPin common framework sequence, which corresponded to the DARPin of chain F in the PDB entry 4drx (the nomenclature 4drx [F] is used) (Pecqueur et al., 2012).

A set of 18 amino-acid positions within the defined DARPin common framework sequence were manually selected and allowed to vary (referred to as variable positions). These positions, which have high-frequency rates of mutation as observed from sequence alignments of many DARPins from the literature, are: 45, 46, 48, 56, 57, 78, 79, 81, 89, 90, 111, 112, 114, 122, 123, 144, 145 and 147 (standard DARPin numbering is applied). Amino-acid side chains at these positions are lining the concave face of the DARPin scaffold by being located within the β-turn loops and following short α-helices of the ankyrin repeats (Figure 2).

Two positive controls having the 18 variable positions corresponding to known DARPin binders of BCL-W, having PDB entries 4k5a [B] and 4k5b [B], were also built manually into the library. It is important to note that these constructed positive controls share the common framework of the designed library described above and thus differ at several positions from the frameworks of the originating known binders (Figure 2).

An alphabet was created to group amino acids by chemical properties. The following five groups excluding Gly, Cys and Pro were defined: positively-charged (Arg, His, Lys); negatively-charged (Asp, Glu); polar (Asn, Gln, Ser, Thr); non-polar (Ala, Ile, Leu, Met, Val); and aromatic (Phe, Trp, Tyr). Equal probability was given to each group to be selected when mutating sequences. Similarly, amino acids within a group were given equal probability.

The designs were generated using a stochastic procedure in which variable amino-acid positions were mutated either through point mutations or through permutations of amino acids. Multiple starting points in the sequence space were used to generate the designs. The set of mutated designs (M-set) were generated starting from the 4drx [F] sequence chosen as common framework. All variable amino acids were forced to be mutated in this set. To be included in the library, a design sequence had to be sufficiently distant to the designs comprised within the same set. A threshold distance of 10 was fixed which required at least 10 alphabet group changes. The set of permutated designs (P-set) were generated starting from the sequences of the two positive controls. No change in the alphabet group was imposed for this set. A threshold distance of 13 was set, requiring at least 13 amino-acid changes.

The designed sub-library sequences were grafted onto four DARPin template structures followed by side-chain repacking using SCWRL4 (Krivov et al., 2009). The last two alanine residues at the C-terminus of the template sequence were truncated for modeling purposes. Only those side-chains that are different at a given side-chain were mutated and repacked to preserve the structural integrity of the original crystal structures of the DARPin templates. The entire structure was then allowed to be repacked. The DARPin templates from the following PDB entries were used in this study: 4drx [F], 4j7w [A], 5lw2 [A] and 5le6 [A] (Figure 1). The backbone structures of these templates are distinct from those of the two known DARPin binders of BCL-W, 4k5a [B] and 4k5b [B], which were purposely excluded as structural templates to avoid the cognate-docking bias. The selected DARPin template structures underwent the following preparation procedure: 1) addition of missing side chain atoms (no repacking); 2) addition of missing hydrogen atoms and assignment of standard protonation states at pH 7; 3) optimization of the hydrogen-bond network the minH program (Hogues et al., 2014); and 4) AMBER force-field (Cornell et al., 1995; Hornak et al., 2006) energy minimization of added hydrogen atoms and any newly added side-chain atoms with harmonic restraints on all the other heavy atoms of 1,000 kcal/mol/ A ̊ 2 followed by energy minimization of the entire structure with harmonic restraints of 10 kcal/mol/ A ̊ 2 on backbone heavy atoms, 1 kcal/mol/ A ̊ 2 on side-chain heavy atoms, and no restraints on hydrogen atoms.

The BCL-W docking-based screening of the DARPin library was performed using the exhaustive docking engine ProPOSE version 1.03 (Hogues et al., 2018). ProPOSE was run with default parameters using the HITSET flag to force binding towards the set of residues involved in binding BCL-W. Initially, no binding location (or epitope) was defined on the target protein BCL-W and exhaustive docking was performed all around the BCL-W structure. Two BCL-W structures were employed for docking, with PDB entries 4k5a [A] and 4k5b [C] (Figure 1), which correspond to the BCL-W complexed with the two known DARPin binders. For each DARPin library sequence, the four DARPin structural templates carrying the grafted designed sequence were docked against the two BCL-W target structure, resulting in 8 docking experiments. In this study, only the top-1 scored pose generated by ProPOSE was considered for a given complex given its accuracy in pose recovery as top-1 when the protein backbone conformation is known, without the need for rescoring (Hogues et al., 2018). On average, a single docking run took 30 min to execute when parallelized on an Intel Xeon Gold 5,218 using 6 cores.

Epitope restriction on the BCL-W target was introduced after all docking calculations were completed. In this proof-of-concept study, we elected to target the same BCL-W epitope and the DARPin binding mode observed for the two known BCL-W DARPin binders (PDB entries 4k5a and 4k5b) (Schilling et al., 2014b). The similarity of predicted docked poses of designed sequences relative to these known structures was based on CAPRI classification (Lensink et al., 2017). Predictions were compared on the basis of: 1) the backbone RMSD of the ligand upon target superposition; 2) the backbone RMSD of the interface upon superposition of interface atoms; and 3) the fraction of preserved contacts (f_con). The ligand and target were DARPin and BCL-W, respectively. Noteworthy, f_con was used rather than the standard f_nat from CAPRI that is derived from the comparison to a native structure. Moreover, f_con is a position-dependent (amino acid-independent) measure allowing designs with different sequences to be compared. Two predictions were declared as having high, medium or acceptable quality, or as incorrect otherwise, with thresholds defined by the CAPRI classification (Lensink et al., 2017).

The number of top-1 scored poses, N_pose, docked at the targeted epitope from the 8 docking runs was used to retain only those designs that have at least 2 poses docked at the target epitope. To this end, the predicted poses for a given DARPin were grouped using a greedy clustering algorithm with a tolerance of at least medium quality between cluster representatives. In geometric terms, for poses to be considered bound at the targeted epitope occupied by one of the known binders, they were required to have acceptable quality criteria, i.e., 1) f_con of at least 30% with a ligand backbone RMSD >5.0Å and interface backbone RMSD >2.0 Å; or alternatively, 2) f_con between 10% and 30% while having a ligand backbone RMSD <10.0 Å or an interface backbone RMSD <4.0 Å. For poses to be part of the same cluster, they were required to have medium quality criteria, i.e., 1) f_con of at least 50% with ligand backbone RMSD >1.0 Å and interface backbone RMSD >1.0 Å; or alternatively, 2) f_con between 30% and 50% while having ligand backbone RMSD <5.0 Å or an interface backbone RMSD <2.0 Å. No cut-off in score was applied for the clustering.

For each design with N_pose > 1, a consensus score was derived as the arithmetic average over the docking scores of the poses binding to the targeted epitope. Consensus scores over the designs with N_pose > 1 were also normalized into Z-scores to better inform the selection of a top-ranked population based on a minimum number of standard deviations away from the mean calculated from the distribution of all DARPins combining the P-set designs, M-set designs and the positive controls.

Structure visualization was performed in PyMOL (The PyMOL Molecular Graphics System, Version 2.0, Schrödinger, LLC). Statistical analyzes were run in R (R Development Core Team, 2011). ClustalW2 was used to run the multiple sequence alignments (Larkin et al., 2007).

The sequence datasets generated for this study have been made available as a MongoDB with example scripts that can be found at the GitHub repository https://github.com/gaudreaultfnrc/Darpins.

Each DARPin design included a N-terminus tag (MRGSHHHHHHGS) and two alanines at their C-terminus as described in (Schilling et al., 2014a). The protein sequences were optimized for Escherichia coli expression using a multifactor algorithm (https://www.genscript.com/tools/gensmart-codon-optimization), then synthesized by GenScript. After inserting each gene in pET24a (+) via NdeI and NotI restriction enzyme sites, the final plasmids were transformed into NRC E. coli BL21-T7 strain (rhaB lacZ::Ptac-T7 RNAP). For each clone, a 2.8-L Fernbach baffled flask containing 500 mL Animal-Product Free (APF) LB Miller (Athena Enzyme Systems Cat. 0133) plus 50 μg/mL kanamycin was inoculated with an overnight preculture to get an initial OD_600nm of 0.1. The flasks were incubated at 37°C, 200–250 rpm until an OD_600nm between 0.8 and 1.0 were reached. To induce protein expression 1 mM isopropyl β-d-1-thiogalactopyranoside (IPTG) was added and the culture incubated for another 4 h at 37°C, 200–250 rpm. The cultures were harvested, and the cell pellets stored at −80°C.

Before purification, a cell pellet was resuspended in Lysis buffer 50 mM NaPO₄, 300 mM NaCl, 10 mM imidazole, pH 7.4 with cOmplete protease inhibitors EDTA-free (Millipore Sigma Cat. 11836170001) and lysed by two passages on a French Pressure Cell Disruptor. Finally, the cell lysate was clarified by centrifugation at 10,000 x g, 4°C, for 15 min and filtration on 0.45 µm filter. A fraction of the clarified lysate (15 mL) was applied on a 3 mL HisPur Cobalt Spin Column (Thermo Fisher Cat. 89969) and the column was washed with 20 mM NaPO₄, pH 7.5, 500 mM NaCl, 0.3 mM TCEP, 15 mM imidazole. Elution was done with 20 mM NaPO₄, pH 7.5, 500 mM NaCl, 0.3 mM TCEP, 100 mM imidazole and pooled after visualization on SDS-PAGE. For some of the proteins, the purification was repeated to increase purity. Buffer exchange for DPBS (Thermo Fisher Cat. 14190144) was done with PD-10 desalting columns (Cytiva Cat. 17085101) and final concentration measured by Qubit Protein Assay (Thermo Fisher Cat. Q33211).

The design of BCL-W was based on (Schilling et al., 2014a) with an N-terminal Avi-tag followed by a bacteriophage lambda protein D fusion tag to improve protein solubility (Forrer and Jaussi, 1998) (see Supplementary Data). A 6xHis tag was added to the C-terminus of BCL-W for purification. Gene optimization, synthesis and cloning in pET24a (+) vector was done as described above for the DARPins. To allow in vitro biotinylation, the NRC E. coli BL21-T7 strain (rhaB lacZ::Ptac-T7 RNAP) was first transformed with pBirAcm (Avidity), a plasmid expressing biotin ligase under tac promoter (IPTG inducible). After growing a chloramphenicol resistant colony in APF LP Miller medium containing 10 μg/mL chloramphenicol, electrocompetent cells were prepared using standard procedures. The plasmid pET24a (+)-BCL-W was then transformed in BL21-T7/pBirAcm strain and selected on APF LB Miller agar containing 50 μg/mL kanamycin and 10 μg/mL chloramphenicol.

Expression of BCL-W, cell lysis and clarification were done as described for the DARPins with some exceptions. Both antibiotics, kanamycin and chloramphenicol, were used, and biotin was added to a final concentration of 5 mM during the culture (25 mL). The cells were lysed in a buffer containing 50 mM NaPO₄, 300 mM NaCl, 10 mM imidazole, pH 8.0 (plus cOmplete EDTA-free protease inhibitors). The clarified lysate (2.5 mL) was applied on a 0.2 mL HisPur Cobalt Spin Column (Thermo Fisher Cat. 90090) and the column was washed with 20 mM NaPO₄, pH 7.5, 500 mM NaCl, 0.3 mM TCEP, 20 mM imidazole. Elution was done with 20 mM NaPO₄, pH 7.5, 500 mM NaCl, 0.3 mM TCEP, 300 mM imidazole and pooled after visualization on SDS-PAGE. Buffer exchange for DPBS (Thermo Fisher Cat. 14190144) was done with G-25 MiniTrap desalting columns (Cytiva Cat. 28918007) and final concentration measured by Qubit Protein Assay (Thermo Fisher Cat. Q33211). Purity levels are given in Supplementary Table S1 and SDS-PAGE gels are provided as Supplementary Data.

Surface plasmon resonance was used to screen the top 18 DARPin designs for binding to the biotinylated BCL-W using a Biacore T200 instrument (Cytiva Inc., Marlborough MA) at 25°C and with PBST running buffer (Teknova, Hollister CA) containing 0.05% Tween 20, 3.4 mM EDTA and an additional 350 mM NaCl. The strategy employed was to capture the biotinylated BCL-W onto the SPR surface with a CAP sensor chip (Cytiva Inc.) and flow a three-point concentration series of the DARPin scaffold using a 10-fold dilution series from 1 μM to cover a wide concentration range. From the resulting sensorgrams, the affinity constant of binding candidates can be determined. A CAP immobilization chip was prepared following the manufacturer’s instructions. Each injection cycle consisted first of a 120-s injection at 5 μL/min of a 5-fold dilution of CAP reagent to indirectly immobilize streptavidin over flow-cells 1 and 2. This was followed by a 240-s capture of 5 μg/mL biotinylated BCL-W at 5 μL/min over flow cell 2 only to form the 60–62 RU BCL-W surface, and finally a three-point concentration injection of the DARPin scaffold or running buffer only using single-cycle kinetics was performed at 50 μL/min for 90 s with a 300-s dissociation phase. At the end of the dissociation phase, any BCL-W/DARPin complex was stripped from the SPR surface using a 60-s injection of 6 M GuCl/0.25 M NaOH taken from the CAP sensor chip reagent kit. The sensorgrams were double referenced and analyzed using the Biacore BiaEval software. Affinities of the DARPin scaffolds for BCL-W were determined using the steady state model, or the 1:1 binding model when kinetic rate constants could be evaluated.

Differential scanning calorimetry (DSC) was used to determine the thermal transition midpoints (T_m) as previously performed (Schrag et al., 2019). DSC was carried out in a VP-Capillary DSC system instrument (Malvern Instruments Ltd., Malvern, United Kingdom). Samples were diluted in DPBS buffer to a final concentration of 0.4 mg/mL. DPBS blank and sample scans were carried out by increasing the temperature from 20°C to 100°C at a rate of 60°C/h, with feedback mode/gain set at “low”, filtering period of 8 s, pre-scan time of 3 min, and under 70 psi of nitrogen pressure. All data were analyzed with Origin 7.0 software (OriginLab Corporation, Northampton, MA). Thermograms were corrected by subtraction of corresponding DPBS blank scans and normalized to the protein molar concentration. The T_m values were determined using automated data processing with the rectangular peak finder algorithm for T_m. Melting temperatures are listed in Supplementary Table S1 and DSC thermograms are provided as Supplementary Data.

The flowchart in Figure 1 presents the overall computational design process devised and implemented for this rigid-docking based proof-of-concept engineering study based on the DARPin scaffold. It includes 6 steps: 1) definition of a single DARPin common framework sequence; 2) expansion of the common framework sequence into a DARPin sequence library with variable positions; 3) selection of a small DARPin sub-library consisting of diverse sequences; 4) grafting of the sequence sub-library onto DARPin structural templates; 5) docking of DARPin sub-library to target protein structures, the core component of the process; and 6) ranking docked DARPin variants for experimental testing. The first three steps operate in the sequence space, whereas the last three in the 3D structure space. All the steps are described in detail in the sub-sections of the Methods section. The following sub-sections focus more in-depth on results obtained in steps 3), 4), 5) and 6) of the process.

Expanding a common framework sequence by varying 18 positions lining the concave face of the DARPin fold (Figure 2) resulted in 10²³ theoretical library size. Millions of iterations were run to select a diverse sub-library fulfilling several design criteria (see Methods sub-Section 2.1.3). The resulting diverse sub-library comprised a total of 2,213 designs of which 1,429 were produced by mutations and 784 by permutations (Table 1). The closest designs in sequence are 9 amino-acid substitutions away from any of the two positive controls (Supplementary Figure S1), or 6 groups away when grouping amino acids by homology (see Methods section). The mutation-based designs have an even proportion of amino-acid groups at the variable positions (Supplementary Figure S2). In contrast, permutation-based designs have unevenly distributed amino-acid groups and lack Ala, His and Ser as inherited from the starting positive-control sequences (Supplementary Figure S2). In terms of net charge, mutation-based designs span a wide range from −16 to +3 with a mean net charge of −6.8, whereas permutation-based designs inherit the net charges of their respective parental positive control (Supplementary Figure S3).

In order to generate a sub-library that samples homogeneously the immense theoretical sequence space, designs were imposed to be orthogonal to each other. Clustering based on amino-acid properties indicated that most sequence space regions were covered by both mutation-based and permutation-based types of sequences, with a few areas only covered by the mutation-based set (Figure 3). While proximity in sequence might be perceivable between some of the designs and the two positive controls (Figure 3A), overall, the designed sequences were diverse and nearly equidistant from each other (Figure 3B).

Four crystal structures were used as templates in the modeling of the DARPin ligands (see Methods section). The variance in RMSD among these templates has a mean of 0.95 Å. The template 4drx [F] is more distant due in part to an opening of the last repeated motif of the scaffold. The magnitudes of backbone changes between each of these templates and any of the 2 known DARPin binders of BCL-W are larger than between the 2 known binders (0.42 Å). Thus, backbone RMSDs of 0.91, 0.75, 0.79 and 0.79Å were calculated to the 4k5a [B] known binder, and of 0.96, 0.75, 0.78 and 0.78Å to the 4k5b [A] known binder, for the template structures 4drx [F], 4j7w [A], 5le6 [A] and 5lw2 [A], respectively. More backbone variations could be observed in the unstructured region of the fourth ankyrin repeat, where the known BCL-W binders had a distinct conformational topology at the tip of this loop region. These variations in the templates relative to known binders were critical for testing the method in real-life application mode in which the bound backbone structure will be unknown a priori.

The entire set of sequence designs in the selected sub-library was grafted onto four template structures, then cross-docked against two target (BCL-W) structures, leading to 8 docking runs per DARPin sequence. The two backbone structures used for the target (4k5a [A] and 4k5b [C]) were relatively close from each other, with an RMSD of 0.77 Å. They also engaged their respective known DARPin binders (4k5a [B] and 4k5b [A]) via a well-preserved binding interface with backbone atoms deviating by an RMSD of 0.60 Å. Hence, in this study, the docked poses for novel DARPins were required to bind around the same epitope that is targeted by these two known DARPin binders of BCL-W. In more technical terms, the predicted poses of designed DARPins were required to have an overlap of at least acceptable quality (according to CAPRI classification (Lensink et al., 2017) to either of these known binders. This was met by 1,033 designs (47% of the sub-library), and are referred to as “locus designs”. (Increasing the stringency and imposing at least a medium quality of pose overlap with the known binders reduced the number of locus designs to 559.) We found no bias towards either of the two target BCL-W structures used for docking, as 811 designs docked to structure 4k5a [A] and 632 designs to structure 4k5b [C]. In terms of the template DARPin structures used for docking, 5lw2 [A] was the least successful template structure with 344 docked designs, followed by 380 designs docked on 5le6 [A], 533 on 4j7w [A] and 579 on 4drx [F]. The net charge distribution of the 1,033 locus designs is slightly different relative the entire docked sub-library of 2,213 designs, as it has sharper peaks at the −6 and −4 net charges (Supplementary Figure S3).

First, locus design DARPins were filtered based on the number of top-1 scored poses, N_pose, that were docked at the targeted epitope from the 8 docking runs for each DARPin. A total of 293 locus designs (13% of the sub-library) had at least 2 poses docked at the target epitope. These were retained for further ranking and were called “consensus designs”. The net charge distribution among the consensus designs had even sharper peaks at the net charges −6 and −4, with the majority of consensus designs at charge −6 (Supplementary Figure S3).

For each of selected 293 consensus designs, a consensus score was derived as the arithmetic average over the docking scores of the poses binding to the targeted epitope. These consensus scores were normally distributed and ranged from −84.2 to −45.3, from strongest to weakest binder (Figure 4). The permutation-based designs were preferentially chosen according to the consensus scores with a median of −65 as opposed to a median of −61 for the mutation-based ones. In total, 152 (52%) and 71 (24%) designs that docked at the targeted epitope did so with values in N_pose of 2 and 3, respectively (inset in Figure 4). The lowest consensus score corresponded to a Z-score of −3.5. Consensus designs with Z-scores below −1.5 were selected for experimental validation, which formed a set consisting of 18 novel DARPins (Table 2). An overlay of all consensus poses for the selected designs is shown in Figure 5A.

The two positive-control interfaces, having the 18 variable positions imported from the two known DARPin binders of BCL-W and grafted onto the common framework sequence, were also docked in the same manner. These positive controls had consensus scores of −75.8 and −73.5, corresponding to Z-scores of −2.1 and −1.7, respectively. With the assumption that none of the designs are true binders, the separation of the positive controls from the designs had an AUC of 0.971 (Figure 4). The AUC dropped to 0.922 when best scores were used instead of consensus scores for designs with N_pose > 1. While working with rigid scaffolds, this observation suggested the need for structural ensembles to achieve better enrichments and thus motivated the use of consensus scores over best scores in the sections that follow. These positive controls were ranked within the range of the top-18 novel designs, and they were also subjected to experimental testing. It is important to note that to properly compare the scores for the two parental known binders of BCL-W with those of the mutants, we needed to base it on the modeled structures of the known binders rather than their crystal structures. Using the crystal structures would be a case of cognate backbone docking and perfect match in shape complementarity leading to out-of-range scores (DARPin/BCL-W docking scores of −144.5 and −123.3 were obtained for 4k5a [B]/4k5a [A] and 4k5b [B]/4k5b [C], respectively). The overlay of all locus docked poses for the two grafted positive control interfaces is shown in Figure 5B to have the same orientation with those of the selected designs (Figure 5A). These poses are further similarly oriented with those of the known binders, as exemplified in Figure 5C. A closer examination reveals that despite an excellent pose recovery for this cross-docking experiment, there are certain noticeable differences in the fine atomic details at the interface, which are likely due mainly to non-cognate backbone coordinates and to a lesser extent to changes of the framework sequence outside the 18 variable positions. Overall, cross-docking of positive controls predicted that they would retain similar binding relative to the corresponding known binders.

As presented in Table 2, most of these top consensus designs (13 of 18) were from the permutation set despite its smaller representation in the initial library. Also, 15 out of the 18 consensus designs had a net charge equal to that of a positive control (−6 or −4), despite the random sequence generation procedure employed. On average, the top-18 designs were 15 mutations away from the positive-control interfaces, with the closest design being 10 mutations away. These top consensus designs had between 2 and 7 top-1 poses bound at the target epitope. Interestingly, a strong bias towards an increased consensus was observed with 14 out of the 18 novel designs (78%) with at least 3 representative poses bound at the targeted epitope. This level has to be contrasted to only 24% of all consensus designs reaching an N_pose > 2. Hence, not only were the top designs predicted to bind stronger to the target, they also did so with a higher number of predicted consensus poses, with an average N_pose of 4.2. Comparably, the two positive controls had N_pose values of 4 and 7 (Figure 4).