2.1 Design and construction of the DMS by phage benchmarking library

We designed a DMS benchmarking (DMS-BM) library (Table S1) to explore the effect of mutation on peptide binding to eleven human bait protein domains (Figure 1a), with distinct binding specificities (Table 1). Twenty-three peptide ligands were retrieved from the Eukaryotic Linear Motif (ELM) database (Kumar et al., 2024) or from the ProP-PD portal (Kliche et al., 2023) (Figure 1b). The design included well-studied interactions such as the p53 degron peptide binding to the SWIB domain of the E3 ubiquitin-protein ligase Mdm2 (MDM2) (Benz et al., 2022), and the cell division cycle-associated protein 2 (CDCA2) LxxIxE motif-containing peptide that binds to the B56 family protein phosphatase 2A (PP2A) regulatory subunit (Hertz et al., 2016). Furthermore, we included peptides from USP10 and CAPRIN1 that both bind to the NTF2-like domain of G3BP1 but have distinct binding motifs (Schulte et al., 2023; Song et al., 2022). Also, two distinct ligands of the talin-1 phosphotyrosine binding (TLN1 PTB) domain were included (Benz et al., 2022). Each SLiM was tiled with two overlapping parental peptides shifted by two amino acids (14 amino acid overlap). Fifteen of the positions in each peptide were subjected to in silico saturation mutagenesis (excluding cysteines for technical reasons), including the overlapping regions of the peptide pairs, resulting in 12,466 peptides (Figure 1b). The mutant peptide pool design was translated to oligonucleotides which were synthesized and genetically fused to the major coat protein P8 for multivalent display on the M13 phage. The sequence coverage of the constructed phage library was found to be 99.6%, with a balanced sequence representation (Figure 1c).

2.2 DMS by phage display correctly defines known SLiM consensuses

The DMS-BM library was used in phage selections (Figure 1d) against the 11 different bait domains (Table 1; Table S2) resulting in binding enriched phage pools. The peptide-coding regions of enriched phage pools from day 1 to day 4 of selections were barcoded and analyzed by NGS. The resulting DNA sequences were translated into peptide sequences, associated with their respective read counts (Table S3) and position specific scoring matrices (PSSMs) were generated for each bait-peptide pair and each selection day. The results were evaluated to choose the appropriate round(s) of selection on which to generate the PSSMs. To this end, we used a benchmarking set of 234 PSSMs based on consensus aligned peptides from motif classes available in the ELM database (Kumar et al., 2024; Tsitsa et al., 2024). The information gathered in ELM is based on the manual curation of peptides binding to the domains from the literature and thus provides a solid publicly available reference set, with the limitation that the number of curated instances of binding peptides for each domain is limited and that the sequences of the potentially biologically relevant binders listed may not correspond to the biophysically best ligands. Indeed, endogenous ligands may have evolved to be suboptimal binders for functional reasons, while phage display often enriches for more hydrophobic peptides (Luck & Trave, 2011). Thus, some discrepancies between the PSSMs generated based on the information available in ELM and the PSSMs generated by the DMS analysis of a specific peptide sequence can be expected. The similarities between the ELM-based PSSMs and the PSSMs defined by the DMS-BM selection results were assessed for each phage selection round (i.e., after 1, 2, 3 and 4 days of selection), as well as for combined selection round results (that is for round 1–2, 2–3, 3–4, 1–3, 2–4, and 1–4). The effect of normalizing the results by library input was also assessed, as some correlation was observed between the peptide abundances in the input library and the outcome of the results of the first day of selection for some of the selections (Table S4). For each DMS-based PSSM, a similarity score p-value and the similarity rank of the PSSM for its corresponding ELM-based PSSM in comparison to the remaining ELM classes screened were calculated (Table S5). To identify the optimal experimental setup, we benchmarked the various selections using a receiver operating characteristic (ROC) analysis of the PSSM similarity score p-value and the rank of the true positives ELM classes (True Positives) in relation to the remaining ELM classes (False Positives). The area under the curve (AUC) was calculated (Figure 1e). We found that the quality of the PSSMs varied by the selection day and that the normalization step slightly improved the results for selection day 1 (Figure S1). However, the most informative results were obtained by combining the data of the second and third rounds of selections using the non-normalized data (AUC 0.91 for the rank and 0.9 for the PSSM similarity), and we thus used these data for the further analysis (Figure S2 and Table S6). Notably, the results of the selection days 2 and 3 were almost as informative by themselves based on the AUC (Day 2: AUC 0.91 for the rank and 0.89 for the PSSM similarity; Day 3: AUC 0.91 for the rank and 0.88 for the PSSM similarity), and the DMS analysis may thus be conducted using only the results of the second or third day of selection. For each PSSM, a completeness score was further calculated and was found to be relatively high (Table 1). A high completeness score (close to 1) indicates that sequencing data generated from the binding-enriched phage pools contained a high proportion of the designed mutations.

We compared the PSSMs generated based on the selection results for a given bait-peptide pair against PSSMs for (i) the same bait with the overlapping peptide; (ii) PSSMs for the same bait with a distinct peptide; and (iii) PSSMs for bait–peptide pairs from unrelated baits. As expected, we observed the highest PSSM similarity for overlapping peptides, followed by distinct peptides binding the same bait, and finally, limited similarity with other PSSMs in the dataset (Figure 1f). The DMS-based PSSMs generally encode binding determinants that are similar to the binding determinants described in ELM. The ELM-derived PSSMs for the respective baits are the best matches for the DMS-derived PSSMs for about 55% of the cases, including KEAP1 Kelch TGE motif and the EIF4E ligand (Table S5). Other representative examples include the DMS of the p53 and RNF115 peptides binding to MDM2, which resulted in the expected FxxxWxxL motif (Figure 2a,b). Similarly, the DMS of the PP2A B56 binding peptides from AXIN1 and CDC2A resulted in a [LM]xx[ILV]xE motif, which closely resembles the previously reported B56-binding LxxIxE motif (Figure 2c,d) (Hertz et al., 2016; Wu et al., 2017). In other cases, we find that the ELM-derived PSSMs are not representative of the motifs found in the peptides analyzed. For example, we note the PSSMs of the talin (TLN1) phosphotyrosine-binding (PTB) domain for which the ELM class LIG_PTB_Apo_2 domain is a poor match with the observed motif as TLN1 PTB does not bind classical PTB-binding motifs, as detailed later. For some baits, such as the G3BP1 NTF2-like domain, we noted differences between the PSSMs generated using distinct model peptides. We used two distinct model peptides for G3BP1, one from USP10 and one from CAPRIN1, that are known binders of the same pocket but have distinct binding modes as shown by co-crystallization of the complexes (Schulte et al., 2023; Song et al., 2022). The DMS analysis of the G3BP1 binding USP10 peptide (₃₋LHSPQYIFGDFSPDEFNQ₋₂₀) correctly identified its G3BP1 binding FG motif (Figure 2e). The DMS analysis of the CAPRIN1 peptide (₃₆₂₋LMAQMQGPYNFIQDSMLD₋₃₇₉) resulted in a distinct YxFI motif based on the averaged results of the two overlapping parental peptides. Notably, for the CAPRIN1_364-379 peptide which generated the highest quality data (completeness score 0.87), an extended YxFxxxSxL motif was obtained (Figure S2). This is similar to the extended YNFIxxxxL G3BP binding motif previously observed based on structural analysis (Schulte et al., 2023). As the terminal leucine is missing in the first CAPRIN1_362-377 peptide, the resulting averaged motif is truncated, suggesting that the frame of the peptides used may affect the motif observed. The results demonstrate the potential of the DMS by phage display approach to reveal distinct motifs within different bound conformations of peptide backbones, and to uncover ELM classes that might benefit from further curation.

2.3 DMS refines the talin-1 PTB-binding determinants

For the atypical PTB domain of TLN1, we probed the two distinct peptides, one from the phosphatidylinositol-4-phosphate 5-kinase type-1 gamma (PIP5K1C: ₆₄₀₋FPTDERSWVYSPLHYS₋₆₅₇) and one from the transmembrane phosphoinositide 3-phosphatase and tensin homolog 2 (TPTE2: ₉₂₋LADLIFTDSKLYIPLEYR₋₁₀₉). Neither of these peptide sequences matches the ELM consensus for unphosphorylated PTB domain ligands (x[ILVMFY]xN.x[FY]x). The DMS analysis revealed a common consensus motif found in the two peptides, [WF]xxSxL, which in the TPTE2 peptide takes an extended form of FxxSxLYxP (Figure 2g,h). The TPTE2 peptide has a phenylalanine instead of a tryptophan at the first position of the motif, as compared to the higher affinity PIP5K1C ligand. We therefore determined the affinities for the wild-type TPTE2_92-107 peptide, and its F97W, F97Y, and F97A mutants using a fluorescence polarization (FP)-based assay (Figures S3 and S4; Table S7). While the wild-type TPTE2_92-107 peptide is bound with a K_D value of 4.5 μM, the F97W mutant is bound with 44-fold higher affinity (K_D = 0.1 μM; Figure 2j). The F97Y mutation conferred instead a reduced affinity (K_D = 150 μM; 33-fold loss) while the F97A mutation resulted in loss of binding within the affinity range tested, highlighting the importance of the position for binding. The DMS analysis of the TPTE2 peptide further suggested that an asparagine would be well tolerated at the third position of the motif, and we found that a TPTE2_92-107 S100N mutation conferred a minor increase in affinity (S100N; K_D = 1.7 μM). We further explored the relevance of the putative extended motif in the TPTE2 peptide and found that mutation of Y103W (K_D = 22 μM) and I104A (K_D = 18 μM) reduced the affinity about fourfold, while a conservative I104L mutation had minor effects, supporting that the TPTE2 exploits a longer motif, that is, that the motif-flanking region in the TPTE2 peptide contributes to binding. To gain further insight into how the TPTE2 peptide is bound by the TLN1 PTB domain, we modeled the complex using AlphaFold3 (Abramson et al., 2024) and overlaid it with the solved structure of the TLN1 PTB–PIP5K1C complex (Figure 2k; Figure S5). The structural analysis showed that the [WF]xxSxL part of the two peptides binds in a similar way, with the [WF] at the first position docking into a hydrophobic pocket. While the PIP5K1C peptide loops out from the binding site, the C-terminal residues of TPTE2 peptide make additional contacts with the domain. In particular, Y103 fits into a shallow pocket at the lid region. Taken together, based on the DMS analysis we define a TLN1 PTB domain consensus motif, that can be C-terminally extended.

2.4 Exploring SLiM-based host-virus interactions by DMS by phage display

Having benchmarked the DMS by phage display protocol and showed its potential for uncovering novel details of well-studied interactions, we next applied the approach to less explored host-pathogen interactions. We designed a second DMS library (Table S1), termed DMS-CoV, based on eight viral peptides binding to five human bait protein domains and eleven human peptides binding to five SARS-CoV-2 bait protein domains (Table 2). The studied interactions included, among others: two viral peptides binding to the G3BP1 NTF2-like domain and human ligands of the globular domains of NSP3 and NSP9 (Figure 3a). The interactions were previously found through proteomic peptide phage display (Kruse et al., 2021; Mihalic Benz, et al., 2023) or predicted as binders based on the presence of a consensus binding motif (i.e. the SH3 binding PxxP motif in the SARS-CoV-2 N binding to ABL2 SH3 domain). NGS analysis confirmed that 96.5% of the designed oligonucleotides were represented in the constructed library (Figure 3b). While the sequence coverage was high, there were systematic deviations such that the SH3 binding ₃₅₉₋DAYKTFPPTEPKKDKKKK₋₃₇₆ peptide from the N protein and its variants were depleted in the constructed phage library, possibly due to the lysine-rich peptide interfering with phage virion assembly. Thus, the coverage of the DMS-CoV library at the peptide level was lower than that of the DMS-BM library (Figure 3c). Nevertheless, we used the DMS-CoV library in selections against the defined bait collection. The results obtained using the DMS-CoV library were less informative than the results obtained using the DMS-BM library (Figure 3d), partially due to the lower coverage but likely also due to the fact that several of the interactions probed were of fairly low affinity, and in particular those between viral protein domains and human peptides (Mihalic, Benz, et al., 2023) and due to some specific traits of the motifs. In several cases, only one of the two overlapping parental peptides returned sufficient data, which may indicate that parts of the motifs were truncated in the shifted peptides. For example, for the EZR FERM domain, the analysis correctly identified the YxΦ motif in the N-terminal part of the envelope (E) protein peptide (Figure 3e). The motif is lost in the shifted peptide, which explains the lack of information obtained for the second parental peptide tiling the region. Finally, the viral USP7 MATH domain ligands included in the design failed to be enriched in the selections as they were outcompeted by peptides from the MBOAT1 and AZIN2, which turned out to contain uncharacterized USP7 binding motifs (Figure S6). The results highlight that factors such as the affinity of the interactions probed and the position of the motif in the peptide should be considered when designing libraries for multiplexed DMS by phage display experiments.

2.5 DMS results for SARS-CoV-2 peptides binding to the NTF-like domains of G3BP1

Among the viral peptides binding to human proteins, we found two interactions particularly interesting. Firstly, our analysis confirmed the expected YxxL AP2M1 binding motif in the probed peptide from NSP14 (Figure 3f), but also suggested that an additional AP2M1 motif can emerge in the peptide upon an isoleucine to tyrosine substitution (YxxV), resulting in two potentially overlapping AP1M1 binding sites in the same peptide. Secondly, the DMS analysis correctly showed that the two viral G3BP1 binding peptides, N_11–28 (K_D = 2.3 μM) and NSP3_956–973 (61 μM) (Kruse et al., 2021) share an (E/T)FG motif (Figure 3g,i), similar to the FG motif found in USP10 (Figure 2). A T16E mutation in the N peptide conferred a sevenfold loss of affinity, partly explaining the higher affinity of the N peptide for G3BP1 as compared to the NSP3 peptide (N_12–26 T16E = 16 μM; Figure 3h). In addition, the DMS results for the N peptide suggested that the interaction is supported by motif flanking residues (ITFGxP), which is consistent with the binding determinants (ITFG) resolved through co-crystallization of the G3BP1 NTF-N peptide complex (Biswal et al., 2022). The results further indicated that the proline contributes to binding, as a P20F mutation conferred a sixfold loss of affinity for N_12-26. The DMS results further suggested that the affinity of the N peptide for G3BP1 could be improved by mutating a glycine in a wild-card position to alanine, and affinity measurements confirmed that the G19A mutation conferred a twofold increase in affinity (K_D wildtype N_12–26 = 2.3 μM; N_12–26 G19A = 1.1 μM). Other mutations tested in the flanking residues conferred moderate or minor losses of affinity (Figure 3h).

2.6 Refining the motifs in human peptides binding to protein domains from SARS-CoV2 proteins

We next turn to the analysis of human peptides binding to viral protein domains expressed by the SARS-CoV-2 genome. We previously uncovered peptide-based interactions of three NSP3 domains, NSP9, and NSP16 (Mihalic, Benz, et al., 2023), which were further explored here.

2.6.1 NSP3

The large multidomain protein NSP3 has several peptide-binding domains including NSP3 ADRP, NSP3 UBl1, and NSP3 SUD-M. For NSP3 ADRP, we tested two peptides previously identified as binders (AZIN2_1–18 and MBOAT1_16–33). Our previous alanine scanning SPOT array analysis suggested the core motif in the MBOAT peptide to be HPLSE (Mihalic, Benz, et al., 2023), and the current DMS analysis confirmed that this is a critical region for binding (Figure 4a,d). Based on the DMS results, we attempted to improve the affinity of the MBOAT1 peptide for NSP3 ADRP by a set of point mutations in the flanking regions, but the mutations resulted in minor (2–3 fold) losses of affinity in comparison to the wild-type MBOAT1_16–31 peptide (K_D = 49 μM) (Figure 4g, Figure S4, and Table S7).

For the NSP3 UBl1 domain, we also subjected two peptides to DMS analysis (NYNRIN_1031–1048 and NCOA2_1072–1089), of which NCOA2 is the higher affinity ligand (Mihalic, Benz, et al., 2023). Consistent with its higher affinity, the most informative results were obtained for the NCOA2_1072–1089 peptide (Figure 4b,e) that converged on an extended ExxLLxxxYxxL motif. The extended motif partially matches the LxxxY motif previously suggested based on SPOT array alanine scanning (Mihalic, Benz, et al., 2023). In an attempt to increase the affinity of the interaction, we tested two mutations (G1077Q and D1081Q) and evaluated their effects on binding. Each of the mutations conferred minor increases in affinity in comparison to the wild-type peptide (Figure 4h; K_D = 19 and 24 μM for G1077Q and D1081Q, respectively, in comparison to 39 μM for wild-type NCOA2_1073–1088).

For NSP3 SUD-M, we tested the two model peptides, PRDM14_197–214 and TET3_459–476, of which the PRMD14_197–214 peptide (₁₉₇₋QFTEEDLHFVLYGVTPSL₋₂₁₄) is the higher affinity ligand (Mihalic, Benz, et al., 2023). Consistently, the DMS selection was dominated by the PRDM14 peptide and its variants (Figure 4c,f). The DMS analysis suggested that the peptide contains an extended FxxExLxFVLxGxT motif, which is similar to the previous results obtained through SPOT array alanine scanning (underlined; Mihalic, Benz, et al., 2023). We designed two mutations to improve the affinity of the interaction (D201E, H203M) and also tested a Y207E thought to be largely neutral to binding, and, as a control, included mutations that were expected to decrease the affinity (D201A, D201N). Affinity measurements revealed that the conservative D201E mutation had the most beneficial impact on binding (K_D = 5.4 μM, 14-fold increase in affinity compared to wild-type peptide K_D of 74 μM), followed by the H203M (K_D = 13 μM). The Y207E mutation also led to a minor improvement of affinity (K_D = 40 μM) (Figure 4i, Figure S4, and Table S7). As expected, the D201A and D201N mutations conferred reduced affinity.

In summary, the DMS analysis of the peptides binding to the NSP3 domains validated their key residues and pinpointed ways to improve their affinities, in particular, for the NSP3 SUD-M domain.

2.6.2 NSP9

NSP9 is a component of the SARS-CoV-2 5′ mRNA capping machinery. We previously reported that it binds to a large number of peptides from human proteins containing a GΦxΦ[GD] motif, where Φ is a hydrophobic amino acid (Mihalic, Benz, et al., 2023), and these interactions were recently confirmed by surface plasmon resonance experiments (Baker et al., 2024). Here, we probed its binding to three peptides, from AXIN1, NEK9 and NOTCH4, respectively, where analysis of AXIN1 and NOTCH4 peptides produced the most robust data. The DMS analysis of the AXIN1 peptide resulted in a G[VF]x[IL]D motif (Figure 5a), while the NOTCH4 peptide instead revealed a similar yet distinct GxWLG motif (Figure 5b,c). Affinity measurements of the wild-type and mutant NOTCH4 peptides and NSP9 showed that a glycine to aspartic acid (G1611D; K_D = 160 μM), or a glycine to proline (G1611P; K_D = 410 μM) substitution at the last position of the motif conferred losses of affinity (2- to 5-fold) as compared to the affinity for the wild-type NOTCH4 peptide (K_D = 72 μM), thus supporting the motif variations between the two model peptides (Figure S4 and Table S7).

As the NSP9-binding NOTCH4 peptide has previously been shown to have an antiviral effect (Mihalic Benz, et al., 2023), we further attempted to generate a higher affinity ligand based on the DMS data. We tested seven mutations and found each of them to confer minor increases in affinities, with a conservative A1608V mutation at the second position of the motif having the largest effect (A1608V, K_D = 16 μM versus K_D = 72 μM for the wild-type; Figure 5d,e). As a negative control, we tested an A1608S mutation, which as expected led to a decreased affinity (K_D = 220 μM). We combined the A1608V mutation with a P1613V mutation (K_D = 33 μM) into a double mutant, which resulted in a further increase in affinity (A1608V/P1613V K_D = 5.9 μM). Having generated a higher affinity NSP9 ligand, we used AlphaFold3 (Abramson et al., 2024) to model the complex. In contrast to the wild-type peptide, the NOTCH4 A1608V/P1613V variant was confidently docked, with the model suggesting that the GVWLG part of the peptide binds through beta-strand addition (Figure 5f). The proposed NSP9 binding region coincides with residues previously mapped by NMR to be perturbed by peptide binding (Mihalic, Benz, et al., 2023). Thus, the previously NMR-mapped binding site and the AlphaFold3-based model confidently pinpoint the peptide binding region on NSP9.

Finally, we tested the antiviral activity of the affinity-matured NOTCH4 A1608V/P1613V peptide by fusing it to a cell-penetrating Tat-tag and evaluated its antiviral effect in comparison to the Tat-tagged NOTCH4 wild-type peptide. VeroE6 cells were treated with the Tat-tagged peptides and infected with SARS-CoV-2 (multiplicity of infection: 0.5). The viral RNA was quantified 8 h post-infection using qPCR. The analysis showed that the Tat-tagged NOTCH4 A1608V/P1613V peptide is a more potent antiviral inhibitor than the wild-type peptide, consistent with the higher affinity (Figure 5g), while not having any effect on cell viability (Figure 5h).

4.1 Library design

The DMS-BM and DMS-CoV phage libraries were designed based on previously reported ligands (Benz et al., 2022; Kruse et al., 2021; Kumar et al., 2024; Mihalic, Benz, et al., 2023). Each wild-type peptide was tiled by two overlapping peptides shifted by 2 amino acids. All wt peptides were mutated at 15 positions, including all overlapping positions to all-natural amino acids except cysteine. The peptides were reverse translated to oligonucleotides optimized for E. coli expression, and flanking regions for library creation were added (5′ CAGCCTCTTCATCTGGC and 3′ GGTGGAGGATCCGGAG).

4.2 Phage library constructions

The oligonucleotides (GenScript) were PCR amplified with Phusion PCR Master Mix (Fisher Scientific) using 90 s 98°C initial denaturation; 18 cycles of 15 s 98°C denaturation, 15 s 55–58°C annealing, and 15 s 72°C elongation; and 5 min 72°C final elongation. The PCR products were purified using the MinElute PCR Purification Kit (Qiagen). The PCR-amplified oligonucleotides (0.6 μg) were 5′ phosphorylated with 20 units of T4 polynucleotide kinase (Fisher Scientific) at 37°C for 1 h in 1× TM buffer (10 mM MgCl₂, 50 mM Tris–HCl, pH 7.5) supplemented with 5 mM dithiothreitol (DTT) and 1 mM adenosine triphosphate (ATP). Following 5 min of cooling on ice, the oligonucleotides were annealed to 10 μg of dU-ssDNA phagemid (90°C for 3 min, 50°C for 3 min, and 20°C for 5 min) in TM buffer. DNA polymerization and ligation were initiated by adding 10 μL 10 mM ATP, 10 μL 10 mM dNTP, 15 μL 100 mM DTT, 30 Weiss units of T4 DNA ligase (Fisher Scientific), and 30 units of T7 DNA polymerase (Fisher Scientific), followed by incubation at 20°C for 16 h. The reaction was stopped by three freeze–thawing cycles. Remaining wild-type dU-ssDNA was digested by incubating with 5 μL FastDigest SmaI (Fisher Scientific, 37°C, 30 min). dsDNA was purified using the QIAquick PCR & Gel Cleanup Kit (Qiagen). The dsDNA phagemid library was electroporated into E. coli SS320 cells (Lucigen) pre-infected with M13KO7 helper phages (ThermoFisher). Electroporated cells were rescued in 25 mL pre-warmed super optimal broth (SOC) medium (0.5 w/v% yeast extract, 2 w/v% tryptone, 10 mM NaCl, 2.5 mM KCl, 10 mM MgCl₂, 10 mM MgSO₄, and 20 mM glucose, pH = 7.0) and incubated at 37°C for 30 min. The phage-producing bacteria were grown overnight (±18 h) in 0.5 L 2YT medium (1 w/v% yeast extract, 1.6 w/v% tryptone, and 0.5 w/v% NaCl) at 37°C, and then harvested. Phage libraries were stored at −80°C in 10 v/v% glycerol.

4.3 Bait expression and purification

The pETM33 (EMBL), PH1003 (Sidhu Lab), pET42a(+) (EMD Biosciences), or pGEX-4T1 (GenScript) vectors containing cDNA encoding bait proteins (Table S2) were used to express GST-tagged baits. Overnight cultures (2YT, 37°C, 200 rpm, 18 h) of E. coli Bl21 DE3 gold cells (Agilent) transformed with the appropriate vector were used to inoculate 500 mL 2YT (supplemented with Kan (50 μg/mL) or Carb (100 μg/mL)), followed by incubation (37°C, 200 rpm). Protein production was induced at OD₆₀₀ 0.6–0.8, with 1 mM isopropyl β-d-1-thiogalactopyranoside (IPTG) for 18–20 h at 18°C, 200 rpm. The bacteria were pelleted (5000g, 5–7 min) and stored at −20°C. Bacteria were dissolved in lysis buffer (phosphate buffered saline (PBS, 37 mM NaCl, 2.7 mM KCl, 8 mM Na₂HPO₄, 2 mM KH₂PO₄), pH 7.4, 1% Triton X-100, 10 μg/mL DNase I, 5 mM MgCl₂, lysozyme (Thermo Scientific), cOmplete Mini, EDTA-free, Protease Inhibitor tablet (Roche, 1 tablet/10 mL)), incubated at 4°C for 1 h and sonicated (2 s pulse, 2 s pause for 20 s). Cell debris was removed (16,000g, 4°C, 1 h). The supernatant was incubated with glutathione (GSH) Sepharose 4 Fast Flow beads (Cytiva; 1 mL resin/bacterial pellet) (4°C, agitation, 1 h), the resin was pelleted by centrifugation (500g), the supernatant was removed and the beads were washed with PBS. Proteins were eluted stepwise with 1 mL elution buffer (50 mM Tris–HCl, 10 mM reduced glutathione, pH 8.0). Protein purities and sizes were confirmed through SDS-PAGE (BioRad Mini-PROTEAN TGX Stain-Free Precast gels, 200 V, 30 min). Purified bait proteins were flash frozen using liquid nitrogen in 16 v/v% glycerol and stored at −80°C until further use.

4.4 Phage selections

Ten micrograms of GST-tagged bait proteins or GST (negative control) in 100 μL PBS were immobilized in Nunc MaxiSorp flat-bottom 96-well plates (ThermoFisher Scientific, Cat: 44-2404-21) for 18 h at 4°C. Wells were blocked with 200 μL 0.5% bovine serum albumin (BSA) in PBS for 1 h at 4°C under gentle agitation. GST-coated wells were washed four times with 200 μL PT (PBS + 0.05 v/v% Tween 20) and (naïve) phage library (10¹¹ phages, 100 μL in PBS) was added to each GST-coated well. Following incubation (4°C, gentle agitation, 1 h), the phage library was transferred to blocked and washed bait-protein-coated wells. After 2 h of incubation at 4°C, unbound phages were removed by five 200 μL PT washes. Bound phages were eluted with 100 μL log-phase E. coli OmniMAX cells (cultured in 2YT medium supplemented with 10 mg/mL tetracycline) for 30 min at 37°C under gentle agitation. 10⁹ M13KO7 helper phages (ThermoFisher) were added to each well and allowed to infect bacteria for 45 min at 37°C. The hyperinfected bacteria were transferred to 1 mL 2YT supplemented with 30 μg/mL Kan, 100 μg/mL Carb, and 0.3 mM IPTG and incubated overnight (37°C, 200 rpm, ±18 h). Bacteria were pelleted (2000g, 4°C, 10 min) and the phage supernatants were transferred to a fresh 96-deep-well plate, pH adjusted by adding 1/10 volume 10× PBS, and heat-inactivated through incubation at 65°C for 10 min. The phage pools were used for the next day of selection.

4.5 Phage Pool ELISA

Proteins (10 μg) in PBS (100 μL/well) were coated in Nunc MaxiSorp flat-bottom 96-well plates for 18 h at 4°C under gentle agitation. Wells were blocked with 200 μL 0.5% BSA in PBS (4°C, 1 h). Phages (100 μL) were allowed to bind to the bait protein- or GST-coated wells for 1 h at 4°C. Unbound phages were washed away with 4× 200 μL PT, and 100 μL HRP-conjugated anti-M13 bacteriophage antibody was added (Sino Biological Inc., Cat: 11973-MM05T-H, 1:5000 diluted in 0.5% BSA in PT). Following a 1 h incubation at 4°C, wells were washed four times with 200 μL PT and once with 200 μL PBS. 100 μL TMB substrate (Seracare, Cat: 5120-0047) was used to detect the bound antibody, and the enzymatic reaction was stopped through the addition of 100 μL 0.6 M sulfuric acid (H₂SO₄). The absorbance at 450 nm was measured with a SpectraMax iD5 microplate reader (Molecular Devices).

4.6 Sample preparation for NGS and data analysis

Peptide-coding regions were amplified and barcoded using Phusion PCR Master Mix (Fisher Scientific) for 22 cycles. PCR products (25 μL) were normalized using Mag-Bind Total Pure NGS magnetic beads (Omega Bio-Tek, Cat: M1378). Normalized PCR products (10 μL) were pooled and purified using gel purification. DNA was eluted with 30 μL TE buffer (10 mM Tris–HCl, 1 mM EDTA, pH 7.5). The amplicon pool was sent for NGS (Illumina MiSeq v3, 1×150bp read setup, 20% PhiX, performed by the NGS-NGI SciLifeLab facility). The raw NGS data were demultiplexed, sequencing adapters were trimmed, and the sequences with an average quality of 20 or more were then translated to peptide sequences. A table was generated containing all peptide sequences and their total read counts. No minimum count filtering was enforced, but only those peptides that were expected for the bait used in the selection were kept for the analysis (Table S3). The NGS sequence processing was performed using custom Python scripts.

4.7 ELM instance specificity determinant dataset

A dataset of PSSMs encoding motif class specificity determinants was created from the motif instances in the ELM database (Kumar et al., 2024). For each SLiM class, peptides were extracted and aligned using the ELM-defined class consensus; alignments were converted to a PSSM using the PSSMSearch web application (Krystkowiak et al., 2018) with default parameters and the frequency PSSM scoring method, resulting in 234 PSSMs. Each bait-peptide pair screened in the DMS-BM analysis was annotated with a corresponding ELM class. The dataset is available online (DOI: 10.5281/zenodo.15297110).

4.8 Specificity determinant comparison

The probability of the observed importance-weighted similarity score between two columns was calculated using a randomization approach based on the comparison of random PSSM columns. A sample of 100,000 randomly selected column pairs between the two PSSM datasets was compared, and the distribution of importance-weighted similarity scores was calculated. The likelihood of seeing the observed importance-weighted similarity score by chance, p^ISW, was defined based on the distribution of the importance-weighted similarity score of the randomly paired PSSM columns. The probability of the window, p^window, was calculated as the product of the pairwise column p^ISW probabilities from the window. The p^window score was normalized to correct for the number of comparisons performed for the window using uniform product distribution correction to define the p^{window_corrected} probability. After all comparison windows were scored, the highest-scoring pair of windows was returned as the aligned specificity determinants, and the p^{window_corrected} was used as the similarity score between the PSSMs.

4.9 Completeness

4.10 Establishing the optimal combination of selection days

All normalized and non-normalized PSSMs split by the day of the selections (1, 2, 3, and 4) and combined day of the selections (1/2, 2/3, 3/4, 1/2/3, and 2/3/4) were compared with the dataset of 234 ELM class specificity determinant PSSMs. A similarity score p-value and the similarity score-derived rank of each comparison were calculated. The comparisons with the bait-peptide pair PSSM with the expected ELM class were classified as True Positive, and all the other comparisons were classified as False Positives. For both the p-value and the rank data, an ROC analysis was performed, and the area under the curve (AUC) was calculated to measure the quality of the selections on different days and combinations of days. The metrics were calculated using the roc_curve and auc functions from the sklearn.metrics library in Python 3.9.7, respectively.

4.11 Replicate comparison

The specificity determinants derived from overlapping and distinct peptide replicates were compared for each peptide in the DMS-BM bait–peptide pair set and a comparison p-value was calculated from each comparison. The p-values for the comparison of the bait–peptide pairs were grouped based on the following criteria: (i) “Same bait/Overlapping peptide” for the bait-peptide pair with the same bait and overlapping peptides, (ii) “Same bait–Different peptide” for the bait–peptide pair with the same bait and non-overlapping peptides, and “Other” for the remaining peptides. “Same Bait–Same Peptide” were excluded from the analysis. The groups were then plotted as boxplots using the seaborn 0.11.1 library in Python 3.9.7. The difference in the means of the groups was compared with a pairwise Mann–Whitney test using the stats.mannwhitneyu function from the scipy 1.7.0 library in Python 3.9.7.

4.12 Expression and purification of proteins for affinity measurements

His-GST-tagged human bait proteins were expressed in 4 L E. coli BL21 (DE3) cultures for FP-based affinity measurements. Lysates were cleared by centrifugation (16,000 RCF, 4°C, 1 h) and the supernatants were mixed with Ni Sepharose High Performance resin (Cytiva) (1 mL beads/pellet) followed by incubation (4°C, agitation, 1 h). The beads were washed with washing buffer (20 mM NaPO₄, 0.5 M NaCl, 30 mM imidazole, pH = 7.5). The His-GST tags were cleaved by incubation for 16–18 h at 4°C in 200 μL 0.5 mg HRV 3C protease and 1 mL primary buffer (20 mM NaPO₄, 0.5 M NaCl, pH = 7.5). Cleaved proteins were collected. The samples were dialysed to 50 mM potassium phosphate buffer (pH = 7.5) for 16–18 h at 4°C. Protein purity and quality were confirmed through SDS-PAGE and thermal shift assay (Tycho NT.6, NanoTemper). His-GST tagged domains of SARS-CoV-2 proteins were expressed and harvested as described above. After the centrifugation step, the lysate was mixed with Pierce glutathione agarose (ThermoFisher) and incubated at 4°C under agitation for 30 min. The gel was washed with wash buffer 2 (50 mM Tris, 300 mM NaCl, 2 mM DTT, pH 7.8) and the protein of interest was eluted with elution buffer (wash buffer 2 supplemented with 10 mM reduced GSH). The His-GST tag was cleaved using HRV 3C protease (inhouse; 18 h at 4°C) and the cleaved tag was removed by reverse immobilized metal affinity chromatography. Purified proteins were subjected to size exclusion chromatography (HiLoad 16/600 Superdex 75 pg.; Cytiva) to remove any residual impurities, concentrated, flash frozen, and stored at −80°C until further use.

4.13 FP-monitored affinity measurements

The results of the FP competition experiments were fitted to a sigmoidal dose–response (variable response; GraphPad Prism).

4.14 Infections experiment

VeroE6 cells were infected with SARS-CoV-2 (MOI:0.5) for 1 h at 37°C and 5% CO₂, then the inoculum was removed and replaced with a medium containing the indicated concentration of peptide. Eight hours post infection, cells were lysed and RNA was isolated using NucleoSpin RNA Plus XS (Macherey Nagel) according to the manufacturer's instructions. cDNA was synthesized using the High-capacity cDNA Reverse Transcription kit (Thermo Fisher). SARS-CoV-2 RNA was quantified using qPCRBIO probe mix Hi-ROX (PCR Biosystems) and the following primers and probes: GTCATGTGTGGCGGTTCACT, CAACACTATTAGCATAAGCAGTTGT, and FAM-CAGGTGGAACCTCATCAGGAGATGC-BHQ. GAPDH was used as a reference gene, detected by the RT qPCR Primer Assay (NM_001195426, Qiagen) and the qPCRBIO SyGreen mix Hi-ROX (PCR Biosystems). qPCR experiments were run on a StepOnePlus real-time PCR system (Applied Biosystems).

4.15 Viability test

Cells were treated with the indicated peptides for 8 h, then cellular viability was determined using CellTiter-Glo® Luminescent Cell Viability Assay (Promega) on a Varioskan LUX Multimode Microplate Reader (ThermoFisher Scientific).