Experimental validation of linear peptide binders of Epstein–Barr virus BHRF1

LiteFold Research ·

Nine of 19 LiteMol-1 peptides bound BHRF1. Each tile is the BLI sensorgram of one positive candidate, BHRF1 at 31.6 to 1000 nM.

BHRF1, the anti-apoptotic BCL-2 homologue of Epstein–Barr virus (EBV), contributes to apoptosis resistance in Burkitt lymphoma. LiteMol-1 is a masked diffusion language model that generates peptides and other molecule classes and can be conditioned on a target protein sequence. Here, we evaluated 19 linear peptides of 18–27 residues generated from the BHRF1 sequence and selected through co-folding, interface scoring and manual review. The Bim BH3 peptide served as a natural-ligand control. Tagged peptide constructs were measured against soluble BHRF1 by biolayer interferometry (BLI) to assess binding and apparent affinity in this capture format.

Nine of 19 submitted candidates (47.4%) received positive binding calls; nine were negative and one failed expression. Peptide_010 and Peptide_012 had mean reported apparent $K_D$ values of 37.6 and 49.9 nM, respectively, compared with 75.6 nM for Bim. These results support binding by selected LiteMol-1 peptides in this single-target capture assay.

Introduction

EBV encodes BHRF1, a viral homologue of the anti-apoptotic BCL-2 proteins. BHRF1 contributes to apoptosis resistance in Wp-restricted Burkitt lymphoma, and its expression promotes chemoresistance and accelerates MYC-driven lymphomagenesis in experimental systems (Fitzsimmons et al., 2020; Kelly et al., 2009).1 BHRF1 binds Bim, its key target in lymphoid cells, and a selective subset of other host pro-apoptotic proteins, including Bid, Puma and Bak. Crystal structures of BHRF1 bound to Bim (PDB 2WH6) and Bak (PDB 2XPX) BH3 peptides show an amphipathic helix in a hydrophobic groove. The groove changes conformation between unliganded and peptide-bound structures (Kvansakul et al., 2010).2

Peptides are an established therapeutic class. Their sequences can be synthesized and optimized rapidly, although proteolytic instability can limit them (Fosgerau & Hoffmann, 2015; Muttenthaler et al., 2021).3 BHRF1 is a recurring target in benchmarks of de novo binder design, including the AlphaProteo target panel (Zambaldi et al., 2024).4

Established peptide-synthesis methods facilitate sequence modification and structure–activity studies.

Sequence-based models have also been developed for de novo peptide binder design. PepMLM generates linear peptide sequences conditioned on a target protein sequence, while PepPrCLIP generates candidates by perturbing peptide embeddings and prioritizes them using contrastive learning. Both studies report experimental validation, including binding assays and cellular tests of peptide-guided protein degradation (Bhat et al., 2025; Chen et al., 2026).5 These approaches provide precedents for the sequence-based design and experimental evaluation of peptide binders. Structure-based models, which have been the dominant approach in this field, instead design binders against the target structure; examples include RFdiffusion3, an all-atom diffusion model (Butcher et al., 2025), and BindCraft, which generates binders by backpropagating through AlphaFold2 (Pacesa et al., 2025).6

We introduced LiteMol-1, a masked diffusion language model. Its introductory report evaluated the model computationally and reserved experimental validation for later work (LiteFold Research, 2026).7 This report describes one such campaign against a single target, with every candidate and its outcome reported. We asked whether peptides generated by LiteMol-1 from the BHRF1 sequence bind BHRF1 in a capture assay, and how their apparent affinities compare with the Bim BH3 peptide, a natural BHRF1 ligand with a crystal structure in complex with BHRF1 (PDB 2WH6), which served as the control (Figure 1).

The sequences, binding outcomes and BLI measurements from this campaign will be made openly available on ProteinBase.

Results

The campaign measured 19 LiteMol-1 candidates and the Bim control as tagged constructs captured on the sensor, with soluble BHRF1 as analyte (Figure 1). Only the BHRF1 sequence was supplied to LiteMol-1; no structural template or explicit BH3 motif was supplied at inference. Adaptyv Bio produced and tested the constructs and supplied the binding calls (Appendix A.2).

Figure 1 Design and experimental workflow. LiteMol-1 generated peptides from the BHRF1 sequence; Protenix v2 and Boltz-2 co-folding, ipSAE scoring and manual review contributed to prioritization, and 19 peptides plus Bim were tested by BLI at Adaptyv Bio (a). BHRF1 (gray) and Bim (violet) are shown in the prepared 2WH6-based reference (b), which does not define the physical assay reagent boundary. Tagged peptides are captured on the biosensor and soluble BHRF1 is the analyte (c). Candidate poses are predicted; tethers and sensorgrams are schematic.

Nine of 19 submitted candidates received positive binding calls

Adaptyv Bio assigned nine positive and nine negative binding calls; Peptide_005 failed expression (Tables 1–2, Figure 2). The positive fraction was 9/19 submitted candidates (47.4%), or 9/18 expression-evaluable candidates (50.0%).

Table 1 Summary of experimental outcomes. The Bim control is excluded.

Measure Count Detail
Candidates submitted 19 Peptide_001 to Peptide_019
Positive binding calls 9 47.4% of submitted
Negative binding calls 9 BLI classification from Adaptyv Bio
Expression failure 1 Peptide_005
Positives with poor-fit flags 2 Peptide_001 and Peptide_002
Positives below 100 nM mean apparent $K_D$ 2 Peptide_010 (37.6 nM), Peptide_012 (49.9 nM); Bim 75.6 nM

Mean reported apparent $K_D$ values of the nine positive candidates span 37.6–719.9 nM. Peptide_001 and Peptide_002 retain provider poor-fit flags (*). The lower-amplitude traces of Peptide_001, Peptide_002, Peptide_006, Peptide_015 and Peptide_017 contain rapid transitions or drift (Figure S1).

Figure 2 Apparent affinities of the positive candidates and Bim. Individual selected-fit estimates (dots) and their means (horizontal lines, in nM) are shown on a logarithmic axis, ordered by mean. Bim is violet, Peptide_010 and Peptide_012 are red, and the other positive candidates are blue.

Table 2 Candidates, BLI classifications and apparent affinities. Sequences are in Appendix Table S1. Mean $K_D$ is defined in Methods. Meas. counts measurement entries, including entries without curves. * Poor-fit flag.

Candidate Length (aa) Expression Binding result Mean $K_D$ (nM) Meas.
Bim 26 High Control (positive) 75.6 2
Peptide_010 20 High Positive 37.6 2
Peptide_012 19 Medium Positive 49.9 4
Peptide_003 20 High Positive 199.9 2
Peptide_009 20 High Positive 296.2 2
Peptide_001 18 High Positive* 399.4* 2
Peptide_015 20 High Positive 553.0 2
Peptide_002 20 High Positive* 598.4* 2
Peptide_017 20 Medium Positive 682.1 2
Peptide_006 18 High Positive 719.9 2
Peptide_004 19 Medium Negative — 2
Peptide_007 27 Medium Negative — 2
Peptide_008 19 High Negative — 2
Peptide_011 19 High Negative — 2
Peptide_013 19 High Negative — 2
Peptide_014 20 High Negative — 2
Peptide_016 18 Medium Negative — 2
Peptide_018 19 Medium Negative — 2
Peptide_019 18 High Negative — 2
Peptide_005 19 None Expression failure — 2

Peptide_010 and Peptide_012 have lower apparent KD than Bim

Peptide_010 and Peptide_012 had mean reported apparent $K_D$ values of 37.6 and 49.9 nM, compared with 75.6 nM for Bim (Table 3). Individual estimates ranged from 35.3–39.9 nM for Peptide_010 (two measurements in run 1), 47.5–52.1 nM for Peptide_012 (four measurements across three runs), and 72.2–79.1 nM for Bim (two measurements in run 1).

All three showed concentration-dependent binding and measurable dissociation (Figure 3). Figure S5 contains all eight lead and control records.

Figure 3 Structures and BLI measurements of Peptide_010 (a), Peptide_012 (b) and Bim (c). Two measurements are shown per peptide; Peptide_012 has four in total. Lead poses are Protenix v2 predictions; Bim is a reference complex. Blue traces: concentration series; gray: 0 nM reference; black: provider-selected fits. Dissociation begins near 300 s. * marks the 1000 nM trace excluded from fitting (measurement 24). The processed 100 nM curve in measurement 19 has a 35.4 s sampling gap; missing observations are not connected. Response scales differ between peptides. All eight measurements are in Figure S5.

Selected fits attribute lower apparent KD to faster association

The selected fits satisfy $K_D = k_{\mathrm{off}}/k_{\mathrm{on}}$. Relative to Bim, Peptide_010 and Peptide_012 have higher fitted $k_{\mathrm{on}}$ and $k_{\mathrm{off}}$, with the larger relative increase in $k_{\mathrm{on}}$ accounting algebraically for their lower apparent $K_D$ (Table 3). Bim has the smallest fitted $k_{\mathrm{off}}$.

Table 3 Apparent affinity and fitted kinetics of the leads and Bim. $K_D$ is shown as range (arithmetic mean); kon and koff are geometric means of selected-fit values. Ratios compare geometric means with Bim.

Peptide Meas. $K_D$, nM $k_{\mathrm{on}}$ (M⁻¹ s⁻¹) $k_{\mathrm{off}}$ (s⁻¹) $k_{\mathrm{on}}$ ratio $k_{\mathrm{off}}$ ratio
Bim (control) 2 72.2–79.1 (75.6) 2.05e+04 0.00155 1.00 1.00
Peptide_010 2 35.3–39.9 (37.6) 8.21e+04 0.00308 4.01 1.99
Peptide_012 4 47.5–52.1 (49.9) 2.48e+05 0.0124 12.12 8.00

$K_D$, kon and koff are from Adaptyv Bio-selected 1:1 binding fits (Appendix A.3). Ratios are relative to Bim. The ratio of geometric mean affinities equals the koff ratio divided by the kon ratio.

BLI signal size did not track apparent KD

In this panel, the response at a single concentration did not consistently track reported apparent affinity. At 100 nM BHRF1, Peptide_010 reached 2.88–3.07 nm and Peptide_012 0.25–1.04 nm, despite mean apparent $K_D$ values of 37.6 and 49.9 nM. Peptide_012 response varied more than fourfold while its $K_D$ estimates remained 47.5–52.1 nM. Response depends on active surface capacity as well as occupancy; differences in loading or active fraction could contribute. In its higher-response measurement 26, both fitted $k_{\mathrm{on}}$ and $k_{\mathrm{off}}$ were lower.

Table 4 Responses at 100 nM BHRF1 for the positives and Bim.

Candidate Records End of association (nm) Mean $K_D$ (nM)
Bim (control) 2 1.211–1.283 75.6
Peptide_010 2 2.876–3.070 37.6
Peptide_012 4 0.246–1.044 49.9
Peptide_003 2 0.873–0.925 199.9
Peptide_009 2 0.867–0.899 296.2
Peptide_001 2 0.128–0.143 399.4*
Peptide_015 2 0.148–0.151 553.0
Peptide_002 2 0.191–0.209 598.4*
Peptide_017 2 0.125–0.134 682.1
Peptide_006 2 0.074–0.098 719.9

* Provider poor-fit flag. The Peptide_010 measurement 19 endpoint is based on one observation after a 35.4 s gap in the processed 100 nM curve; other endpoint windows contain 25–26 observations.

Negative candidates show little or no binding response

The nine non-binders showed low responses at 100 nM BHRF1. At 1000 nM, eight of nine remained below their matched mean blank endpoints (Table 5, Figure 4).

Table 5 Responses of the nine negative candidates. Ranges span the two measurement records of each candidate; negative values are shown as recorded.

Candidate Tag Expression 100 nM end (nm) 1000 nM end (nm) 1000 nM end minus matched blank (nm)
Peptide_004 C Medium -0.024 to -0.006 0.016 to 0.032 -0.038 to -0.022
Peptide_007 N Medium -0.012 to 0.011 0.002 to 0.034 -0.052 to -0.020
Peptide_008 C High 0.000 to 0.010 -0.003 to 0.014 -0.058 to -0.041
Peptide_011 C High -0.038 to -0.022 -0.068 to -0.056 -0.122 to -0.111
Peptide_013 N High -0.030 to -0.006 -0.021 to -0.013 -0.075 to -0.068
Peptide_014 N High -0.161 to -0.152 -0.028 to -0.020 -0.083 to -0.074
Peptide_016 N Medium 0.004 to 0.029 0.049 to 0.188 0.033 to 0.154
Peptide_018 C Medium -0.014 to -0.004 0.003 to 0.006 -0.028 to -0.013
Peptide_019 C High 0.005 to 0.010 0.022 to 0.023 -0.033 to -0.032

Endpoint minus blank uses the run and concentration-matched mean blank endpoint. Run 1 blanks are the mean of two reads per concentration; runs 2–4 have one blank read per concentration.

Peptide_016 had 1000 nM responses of 0.049 and 0.188 nm, both above their matched blanks, but retained the provider's negative call and has no affinity estimate. Peptide_005 failed expression and has no curves; its binding remains unknown.

All nine positive candidates had requested N-terminal tags: 9/13 N-tagged candidates were positive and four were negative. Of six C-tagged candidates, five were negative and one failed expression. It is therefore unclear whether the lack of binding among C-tagged candidates reflects the tag position, the sequences themselves, or both.

Figure 4 End-of-association responses. Individual measurements at 100 nM (a) and 1000 nM (b) BHRF1; circles, positive calls; squares, negative calls; diamonds, Bim. Gray bands show blank endpoint ranges. All exported endpoints are plotted, including concentrations excluded from kinetic fitting. Peptide_005 (expression failure) is not plotted.

Evaluation of interface confidence

Retrospective Boltz-2 and Protenix v2 confidence scores were compared with provider calls for 18 expression-evaluable candidates (nine positive, nine negative; Figure 5). These candidates had already passed structural and manual prioritization, so score distributions are restricted and the AUROCs describe this selected panel rather than general predictor performance. In the 13 N-tagged candidates (nine positive, four negative), Protenix ipSAE-max AUROC was 0.750 versus 0.889 in the full panel, while Boltz-2 ipSAE-max was 0.861 versus 0.630 (Table S6).

Figure 5 Retrospective confidence scores in the selected 18-candidate panel. AUROC with descriptive 95% sequence-family bootstrap intervals from 10,000 draws (a); dashed line, chance (0.5). Scores for nine positive and nine negative calls (b); one marker per peptide, using the median of three seeds; bars, group medians. Axes differ by column. Dashed lines at ipTM 0.90 and binder pLDDT 80 are reference values, not selection cutoffs. ipSAE used 10 Å PAE and 15 Å distance cutoffs.

Methods

Before wet-lab submission, candidates were prioritized using Boltz-2 and Protenix co-folding (ByteDance AML AI4Science Team et al., 2025; Passaro et al., 2025) and ipSAE interface confidence (Dunbrack, 2025), together with predicted-interface inspection and manual review.8 Sequence-only input refers to generation; downstream selection used structural information.

The retrospective evaluation used Boltz-2 and Protenix v2 predictions (Zhang et al., 2026).9 For each seed, ipSAE min and max are the minimum and maximum of the two directional interface scores; each metric was then summarized by its median across three seeds. The cutoffs were 10 Å PAE and 15 Å distance. ipTM is the predictor-native score; binder pLDDT is the mean over peptide residues (0–100). The reference lines in Figure 5 were not filters applied in this retrospective analysis.

AUROC compared nine positive and nine negative calls, excluding Bim and the expression failure. Families were connected components at ≥80% identity, defined as identical aligned residues divided by the longer sequence length. Global alignments used match +2, mismatch −1, gap opening −3 and extension −0.5, taking the first optimal alignment. There were 14 families in the evaluable panel. We drew 10,000 bootstrap samples of 14 families with replacement, retaining every member and its multiplicity; 9,979 samples contained both classes. The 2.5th and 97.5th percentiles define descriptive intervals, without adjustment for multiple comparisons. All eight metrics used the same draws (random seed 20261113). The N-tagged sensitivity subset contained 13 candidates and nine families.

Peptide identifiers were matched to source records by core sequence. BLI used captured tagged constructs and soluble BHRF1; the requested 1–160 boundary is not confirmed for the supplied reagent (Appendix A.1). Apparent $K_D$ means are arithmetic means of measurement-level estimates; multiplicative kinetic comparisons use geometric means. Endpoints use exported observations. Provider-selected models, binding calls and concentration exclusions were preserved; no new affinity fitting was performed.

Discussion

This campaign identified selected LiteMol-1 peptides that bind BHRF1 in the tagged capture assay. Peptide_010 and Peptide_012 had lower provider-reported apparent $K_D$ than Bim, but the comparison is conditional on the surface format, selected models and limited repeat measurements. Local response capacities, drift and structured fit deviations limit interpretation of precise rate constants and affinity rankings.

The peptides reported here are biochemical leads. Moving from biochemical leads towards therapeutic candidates would require evidence on selectivity, binding site and cellular activity, and on general peptide liabilities such as stability and delivery, none of which is addressed here.

The reported hit fraction describes sequence generation followed by structural and manual filtering via computational scores. It does not measure an unfiltered generator success rate. The peptides include related sequences and BH3-like patterns (Figure S8); similarity neither establishes a novel binding mode nor reveals training-set exposure. Tag placement is confounded with sequence, and the same peptide was not tested in both orientations. These limitations also affect interpretation of the retrospective predictor comparison. No matched panel from another generator was evaluated experimentally or through an equivalent unfiltered computational pipeline; comparative generator performance is therefore unresolved.

Appendix A: Methods and supplementary data

Tagged constructs and the BHRF1 target

Reported core sequences exclude tags and linkers. N-terminal tag placement was requested for 13 candidates and Bim; six candidates had C-terminal placement (Table S1). The rationale for individual assignments could not be established. Adaptyv Bio's standard construct includes split-GFP11 for expression quantification and Twin-Strep for capture, joined by Gly/Ser linkers (Adaptyv Bio, n.d.-c).10 Exact assembled sequences and linker arrangements for this campaign remain unconfirmed. Figure S7 shows the two requested orientations schematically.

Soluble BHRF1 was used as analyte. Provider correspondence indicated that the transmembrane region would be excluded, but the supplied lot sequence was not confirmed. The selected GenScript product is described as partial BHRF1 with an N-terminal 10×His tag (Adaptyv Bio, n.d.-d).11 Residues 1–160 were requested. The target His tag is distinct from the peptide capture tag.

BLI acquisition and Adaptyv Bio outcome definitions

Binding was measured by BLI using captured tagged peptides and soluble BHRF1, following the provider's assay format (Adaptyv Bio, n.d.-b).12 Exported concentrations were 0, 31.6, 100, 316.2 and 1000 nM, with approximately 300 s each for association and dissociation. Peptide_010 and Bim each have two measurements in run 1; Peptide_012 has four across three runs.

Binding calls were supplied by Adaptyv Bio. Its documentation describes True as a clear binder above controls, False as expression with signal at or below controls, and Unknown as no/low expression or marginal signal (Adaptyv Bio, n.d.-a).13 These general definitions do not supply a quantitative rule that reproduces each campaign decision, including Peptide_016. Nine candidates received True calls and nine False calls; Peptide_005 failed expression and received Unknown. Bim is a positive control excluded from candidate counts. Poor-fit flags for Peptide_001 and Peptide_002 are preserved as provider annotations.

Selected kinetic models and descriptive response calculations

The 22 provider-selected fits cover nine positive candidates and Bim. Each is a 1:1 model with $k_{\mathrm{on}}$ and $k_{\mathrm{off}}$ shared across included concentrations, curve-specific $R_{\max}$ and a shared linear drift term; refractive-index and initial-offset terms were fixed at zero. The displayed fitted response includes both binding and drift, so its slope is not solely a bound-complex decay rate. Per-curve capacities allow loading or active-fraction differences but also make estimates conditional on that parameterization. Common-capacity and steady-state models were not independently fitted.

The provider excluded 100 nM in Peptide_001 measurement 3 and 1000 nM in Peptide_012 measurements 24–26; measurement 23 retained 1000 nM. The campaign records identify the excluded curves. Affinity comparisons have not been tested by refitting a uniform concentration range.

$K_D = k_{\mathrm{off}}/k_{\mathrm{on}}$ was verified for all selected fits. Provider-reported within-fit 95% intervals are listed in Table S4 and retained in the data package. Table S5 reports curve-specific $R_{\max}$ and drift.

All 225 exported reads are marked "baseline": subtraction of the same result's zero-concentration reference in the same run. The package describes smoothing that reference with a 31-point, second-order Savitzky–Golay filter. Concentration-matched blank subtraction was not applied to these published curves; blanks were summarized separately. Association endpoints are medians 5–1 s before association ends. Dissociation loss is (early − late)/early, using medians 1–5 s after dissociation starts and over the final 5 s. These are descriptive signal changes, not fractional occupancies or independent estimates of $k_{\mathrm{off}}$. The processed 100 nM curve in measurement 19 has a 35.4 s gap and only one endpoint-window observation; its cause and effect on fitting are not established.

Data provenance and figure assembly

Analyses preserve the source package, file hashes, sequence-based identities and original measurement/fit identifiers. Binding calls, missing samples and concentration exclusions were retained; no source curves were modified. Candidate complexes in Figures 3 and S6 were predicted using Protenix v2 and visualized using UCSF ChimeraX (Pettersen et al., 2021).14

Generation and historical prioritization

The initial pool comprised approximately 500–600 generated and co-folded structures with Boltz-2 and Protenix v2.

Tables S1–S6 and Figures S1–S8 provide the supplementary evidence.

Table S1 Complete candidate and control panel

Peptide Core sequence aa Tag Outcome Mean $K_D$ (nM) Meas.
Bim DMRPEIWIAQELRRIGDEFNAYYARR 26 N Control 75.6 2
Peptide_001 TFYKEVAKHLEKIGDSLD 18 N Positive* 399.4* 2
Peptide_002 RPIIWIGQEISKVGDRENAY 20 N Positive* 598.4* 2
Peptide_003 DPKEIARRLEKIGEEEDLRL 20 N Positive 199.9 2
Peptide_004 YLAHVNEEVRRVGDELNQR 19 C Negative — 2
Peptide_005 HQEIAQRLSRIGDEEDLRL 19 C Expr. fail — 2
Peptide_006 SVEVEVARHLERIGDELD 18 N Positive 719.9 2
Peptide_007 EAAAHLDFVHSEIRRVGEELNHAAAEQ 27 N Negative — 2
Peptide_008 LHNRVAQEVRRIGDELANE 19 C Negative — 2
Peptide_009 DHREIAQRLEKIGDEEDLRL 20 N Positive 296.2 2
Peptide_010 DYREIAERLAAIGDEEDLRL 20 N Positive 37.6 2
Peptide_011 SLDSVTHEIHRVGDEWAKT 19 C Negative — 2
Peptide_012 HREIAERLARIGDEEDLKL 19 N Positive 49.9 4
Peptide_013 ALENFAAEVRRLGDEIHQR 19 N Negative — 2
Peptide_014 GMKTVKNDLQAVARLYNREL 20 N Negative — 2
Peptide_015 LPQEVEQQVQEAIQEIKAVT 20 N Positive 553.0 2
Peptide_016 SEVAQRLERIGEEEDLRL 18 N Negative — 2
Peptide_017 DPVEIARRLEKIGEEEDLKY 20 N Positive 682.1 2
Peptide_018 YRAAVREEIRRIGDELEAA 19 C Negative — 2
Peptide_019 DLREISKKLEQVGDNMQK 18 C Negative — 2

Sequences exclude tags/linkers. N/C indicates tag placement. * Provider poor-fit flag.

Table S2 Individual selected fits and concentration exclusions

Peptide Meas./ run Fit App. $K_D$ (nM) $k_{\mathrm{on}}$ (M⁻¹ s⁻¹) $k_{\mathrm{off}}$ (s⁻¹) RMSE (%) Excluded (nM)
Bim 1/1 3 79.12 2.06e+04 0.00163 2.04 None
Bim 2/1 6 72.17 2.03e+04 0.00147 2.00 None
Peptide_001* 3/1 9 471.68 1.58e+05 0.0746 5.13 100
Peptide_001* 4/1 12 327.12 1.66e+05 0.0542 6.65 None
Peptide_002* 5/1 15 462.03 1.36e+05 0.0628 6.01 None
Peptide_002* 6/1 18 734.73 1.11e+05 0.0814 3.99 None
Peptide_003 7/1 21 206.33 9.34e+04 0.0193 2.58 None
Peptide_003 8/1 24 193.54 9.27e+04 0.0179 2.76 None
Peptide_006 11/1 27 773.59 1.37e+05 0.106 4.58 None
Peptide_006 12/1 30 666.15 1.48e+05 0.0987 4.16 None
Peptide_009 17/1 33 286.40 7.22e+04 0.0207 2.20 None
Peptide_009 18/1 36 306.09 6.98e+04 0.0214 1.83 None
Peptide_010 19/1 39 35.29 8.52e+04 0.00301 2.52 None
Peptide_010 20/1 42 39.94 7.92e+04 0.00316 2.53 None
Peptide_012 23/1 45 48.74 2.81e+05 0.0137 3.62 None
Peptide_012 24/1 48 52.15 2.7e+05 0.0141 4.20 1000
Peptide_012 25/2 51 47.51 2.85e+05 0.0135 3.45 1000
Peptide_012 26/3 54 51.32 1.76e+05 0.00904 3.74 1000
Peptide_015 31/1 57 783.48 9.92e+04 0.0777 3.30 None
Peptide_015 32/1 60 322.49 1.83e+05 0.059 3.57 None
Peptide_017 35/3 63 705.96 1.41e+05 0.0998 3.52 None
Peptide_017 36/4 66 658.33 8.27e+04 0.0544 2.99 None

* Poor-fit candidate flag. RMSE is relative to maximum signal.

Table S3 Lead and positive-panel responses at 100 nM

Peptide Meas. Assoc. end (nm) Dissoc. loss (%) Blank / sample (%) End-window points Max gap (s)
Bim 1 1.283 32.3 1.40 26 1.0
Bim 2 1.211 33.5 1.49 25 0.4
Peptide_001 3 0.143 38.7 12.59 26 0.2
Peptide_001 4 0.128 120.4 14.06 26 0.8
Peptide_002 5 0.209 52.0 8.61 26 0.8
Peptide_002 6 0.191 85.2 9.42 26 0.8
Peptide_003 7 0.925 78.3 1.95 26 0.6
Peptide_003 8 0.873 85.3 2.06 26 0.8
Peptide_006 11 0.098 34.4 18.37 26 0.6
Peptide_006 12 0.074 33.9 24.32 26 0.2
Peptide_009 17 0.899 84.3 2.00 26 0.6
Peptide_009 18 0.867 86.7 2.08 26 0.6
Peptide_010 19 3.070 39.5 0.59 1 35.4
Peptide_010 20 2.876 40.5 0.63 26 0.8
Peptide_012 23 0.277 70.6 6.50 26 0.4
Peptide_012 24 0.246 81.0 7.32 26 0.6
Peptide_012 25 0.295 78.3 1.02 26 0.6
Peptide_012 26 1.044 62.0 0.38 25 0.8
Peptide_015 31 0.151 60.0 11.92 26 0.6
Peptide_015 32 0.148 75.6 12.16 26 0.6
Peptide_017 35 0.134 86.4 2.99 26 0.6
Peptide_017 36 0.125 79.4 8.00 26 0.4

Definitions are in Appendix A.3. Loss above 100% reflects negative late response. Record 19 has one endpoint after a 35.4 s gap; Blank/sample is the absolute matched mean blank endpoint divided by positive sample endpoint.

Table S4 Provider-reported within-fit 95% intervals

Peptide Meas. $K_D$ (nM) $k_{\mathrm{on}}$ (M⁻¹ s⁻¹) $k_{\mathrm{off}}$ (s⁻¹)
Bim 1 77.85–80.41 2.05e+04–2.08e+04 1.61e-03–1.66e-03
Bim 2 71.01–73.33 2.01e+04–2.05e+04 1.45e-03–1.49e-03
Peptide_001 3 424.54–525.55 1.45e+05–1.72e+05 7.16e-02–7.77e-02
Peptide_001 4 289.07–371.11 1.50e+05–1.83e+05 5.15e-02–5.72e-02
Peptide_002 5 412.36–519.33 1.24e+05–1.49e+05 6.01e-02–6.57e-02
Peptide_002 6 676.55–799.87 1.03e+05–1.19e+05 7.93e-02–8.37e-02
Peptide_003 7 202.40–210.35 9.19e+04–9.50e+04 1.91e-02–1.94e-02
Peptide_003 8 189.71–197.46 9.11e+04–9.42e+04 1.78e-02–1.81e-02
Peptide_006 11 682.48–882.47 1.23e+05–1.52e+05 1.02e-01–1.11e-01
Peptide_006 12 599.89–742.60 1.36e+05–1.61e+05 9.50e-02–1.02e-01
Peptide_009 17 281.50–291.39 7.12e+04–7.33e+04 2.06e-02–2.08e-02
Peptide_009 18 301.69–310.56 6.89e+04–7.06e+04 2.12e-02–2.15e-02
Peptide_010 19 34.48–36.10 8.45e+04–8.59e+04 2.93e-03–3.08e-03
Peptide_010 20 39.23–40.65 7.86e+04–7.98e+04 3.11e-03–3.22e-03
Peptide_012 23 47.44–50.07 2.74e+05–2.87e+05 1.35e-02–1.38e-02
Peptide_012 24 50.56–53.79 2.63e+05–2.77e+05 1.39e-02–1.43e-02
Peptide_012 25 46.21–48.85 2.79e+05–2.91e+05 1.34e-02–1.37e-02
Peptide_012 26 50.08–52.59 1.73e+05–1.79e+05 8.91e-03–9.17e-03
Peptide_015 31 728.96–843.40 9.30e+04–1.06e+05 7.62e-02–7.93e-02
Peptide_015 32 302.55–343.93 1.73e+05–1.93e+05 5.77e-02–6.04e-02
Peptide_017 35 653.73–764.04 1.32e+05–1.51e+05 9.77e-02–1.02e-01
Peptide_017 36 626.18–692.67 7.91e+04–8.64e+04 5.36e-02–5.53e-02

Intervals are copied from the selected provider fits, with display rounding only. Point estimates and excluded concentrations are in Table S2. These intervals are conditional on the selected model; they do not include biological or systematic assay uncertainty.

Table S5 Selected-fit response capacities and drift

Peptide Meas. $R_{\max}$ 31.6 nM $R_{\max}$ 100 nM $R_{\max}$ 316.2 nM $R_{\max}$ 1000 nM Drift (nm/s)
Bim 1 4.476 3.400 2.722 2.806 1.48e-04
Bim 2 4.171 3.243 2.701 2.881 2.92e-05
Peptide_001 3 1.870 Excluded 0.722 1.211 4.09e-04
Peptide_001 4 1.020 0.414 0.677 1.203 1.81e-04
Peptide_002 5 0.834 0.783 0.745 1.198 2.17e-04
Peptide_002 6 1.761 1.275 1.084 1.471 1.47e-04
Peptide_003 7 3.509 2.521 2.482 2.672 4.24e-04
Peptide_003 8 3.318 2.338 2.415 2.673 3.21e-04
Peptide_006 11 0.557 0.573 0.507 0.677 1.04e-04
Peptide_006 12 0.390 0.420 0.426 0.652 1.06e-04
Peptide_009 17 3.315 3.170 2.526 2.975 3.30e-04
Peptide_009 18 3.488 3.254 2.537 3.006 2.53e-04
Peptide_010 19 5.286 3.977 3.831 3.759 7.99e-04
Peptide_010 20 4.509 3.747 3.841 3.779 9.42e-04
Peptide_012 23 0.514 0.365 0.312 0.281 1.08e-04
Peptide_012 24 0.550 0.357 0.311 Excluded 8.03e-05
Peptide_012 25 0.407 0.399 0.364 Excluded 8.86e-05
Peptide_012 26 1.153 1.382 1.331 Excluded 4.70e-04
Peptide_015 31 1.798 1.142 0.791 0.857 1.17e-04
Peptide_015 32 0.684 0.471 0.418 0.580 7.67e-05
Peptide_017 35 0.713 0.910 0.762 0.592 3.95e-05
Peptide_017 36 0.438 0.830 0.999 1.113 3.97e-05

$R_{\max}$ values are fitted capacities in nm for the indicated BHRF1 concentrations. "Excluded" means that trace was not in the selected fit; no $R_{\max}$ is supplied for it. The 0 nM reference is not fitted as a binding curve. One drift slope is shared across included curves in each measurement.

Table S6 Predictor scores and tag-orientation sensitivity

Predictor Metric All 18 AUROC 95% family interval N-tagged 13 AUROC
Protenix v2 ipSAE max 0.889 0.667–1.000 0.750
Protenix v2 ipSAE min 0.840 0.576–1.000 0.806
Protenix v2 ipTM 0.852 0.576–1.000 0.806
Protenix v2 Binder pLDDT 0.778 0.485–1.000 0.806
Boltz-2 ipSAE max 0.630 0.356–1.000 0.861
Boltz-2 ipSAE min 0.765 0.500–1.000 0.778
Boltz-2 ipTM 0.827 0.542–1.000 0.778
Boltz-2 Binder pLDDT 0.593 0.200–0.917 0.639

All 18: nine positive and nine negative calls in 14 sequence families. N-tagged 13: nine positive and four negative calls in nine families. The full-panel intervals use 10,000 family draws, of which 9,979 retained both classes.

Supplementary figures

Supplementary Figures S1–S8 follow: positive-call sensorgrams (S1), negative calls and expression failure (S2), blanks (S3), selected-fit parameters (S4), all lead/control measurements (S5), other predicted positive-candidate complexes (S6), requested tag orientations (S7), and sequence context/families (S8).

Figure S1 Remaining positive-call sensorgrams. All seven other positive candidates and Peptide_012 records 25–26. Peptide_001 and Peptide_002 retain poor-fit flags. Excluded concentrations are shown without fits: 100 nM for Peptide_001 record 3 and 1000 nM for Peptide_012 records 25–26. Panels use individual response scales.

Figure S2 Negative-call sensorgrams. Both records for all nine negative candidates share one response scale; negative shifts are retained. Peptide_005 failed expression and has no exported curves.

Figure S3 Blank traces. All 25 blank-buffer traces, grouped by blank set and run and colored by BHRF1 concentration. Dashed lines distinguish second reads at the same concentration.

Figure S4 Selected-fit parameters. Apparent KD (a), kon (b) and koff (c) for 22 selected fits covering Bim and nine positive candidates. Each symbol represents one fit; † marks poor-fit candidates. Fit quality and concentration exclusions are listed in Table S2.

Figure S5 All eight BLI records for the leads and Bim. Two records each for Peptide_010 and Bim, and four for Peptide_012. Blue/gray traces show observations; black lines show selected fits. Asterisks mark exclusion of 1000 nM from fitting. Sampling gaps remain unconnected. Response scales are shared within each peptide, without normalization.

Figure S6 Predicted complexes of the seven remaining BLI-positive peptides. Colored peptides are shown against gray BHRF1. Complexes were predicted with Protenix v2; their poses are not experimentally confirmed. Peptide_001 and Peptide_002 retain poor-fit affinity flags. Corresponding sensorgrams are in Figure S1.

Figure S7 Tagged peptide capture in the two requested orientations. N and C label peptide-construct polarity. The capture-tag module lies before or after the peptide core; soluble BHRF1 is the analyte. Provider-standard components include split-GFP11 and Twin-Strep with Gly/Ser linkers, but exact module order, linker lengths and geometry are schematic. The BHRF1 product is described as N-terminally 10×His-tagged; this target tag is distinct from the peptide capture handle.

Figure S8 Sequence context and family assignments. Pairwise global alignments of each candidate to Bim are merged at Bim reference positions; dashes are gaps and shading marks identity to Bim. Families use candidate-to-candidate global identity at ≥80%, with the longer sequence as denominator and connected-component grouping. F01 contains six submitted candidates, five positive and one expression failure; all other families are singletons.

Footnotes

  1. Fitzsimmons et al. (2020). EBV BCL-2 homologue BHRF1 drives chemoresistance and lymphomagenesis by inhibiting multiple cellular pro-apoptotic proteins. Cell Death & Differentiation, 27(5), 1554–1568. DOI
    Kelly et al. (2009). An Epstein-Barr Virus Anti-Apoptotic Protein Constitutively Expressed in Transformed Cells and Implicated in Burkitt Lymphomagenesis: The Wp/BHRF1 Link. PLoS Pathogens, 5(3), e1000341. DOI ↩

  2. Kvansakul et al. (2010). Structural Basis for Apoptosis Inhibition by Epstein-Barr Virus BHRF1. PLOS Pathogens, 6(12), e1001236. DOI ↩

  3. Fosgerau & Hoffmann (2015). Peptide therapeutics: Current status and future directions. Drug Discovery Today, 20(1), 122–128. DOI
    Muttenthaler et al. (2021). Trends in peptide drug discovery. Nature Reviews Drug Discovery, 20(4), 309–325. DOI ↩

  4. Zambaldi et al. (2024). De novo design of high-affinity protein binders with AlphaProteo (Preprint 2409.08022). arXiv. DOI ↩

  5. Bhat et al. (2025). De novo design of peptide binders to conformationally diverse targets with contrastive language modeling. Science Advances, 11(4), eadr8638. DOI
    Chen et al. (2026). Target sequence-conditioned design of peptide binders using masked language modeling. Nature Biotechnology, 44(6), 1002–1010. DOI ↩

  6. Butcher et al. (2025). De novo Design of All-atom Biomolecular Interactions with RFdiffusion3 [Preprint]. bioRxiv. DOI
    Pacesa et al. (2025). One-shot design of functional protein binders with BindCraft. Nature, 646(8084), 483–492. DOI ↩

  7. LiteFold Research. (2026, August 17). LiteMol-1: A Multi Molecule Foundation Diffusion Language Model for Agents. LiteFold. https://www.litefold.ai/research/litemol1 ↩

  8. ByteDance AML AI4Science Team et al. (2025). Protenix—Advancing Structure Prediction Through a Comprehensive AlphaFold3 Reproduction [Preprint]. bioRxiv. DOI
    Passaro et al. (2025). Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction [Preprint]. bioRxiv. DOI
    Dunbrack (2025). Rēs ipSAE loquuntur: What's wrong with AlphaFold's ipTM score and how to fix it [Preprint]. bioRxiv. DOI ↩

  9. Zhang et al. (2026). Protenix-v2: Broadening the Reach of Structure Prediction and Biomolecular Design [Preprint]. bioRxiv. DOI ↩

  10. Adaptyv Bio. (n.d.-c). Construct & tag design. Retrieved November 9, 2026, from https://docs.adaptyvbio.com/wiki/technologies/construct-tag-design ↩

  11. Adaptyv Bio. (n.d.-d). Recombinant HHV-4 BHRF1, N-10xHis-tagged, partial (DP0058). Retrieved November 9, 2026, from https://targets.adaptyvbio.com/product/Custom/DP0058 ↩

  12. Adaptyv Bio. (n.d.-b). BLI methods. Retrieved November 9, 2026, from https://docs.adaptyvbio.com/wiki/methods/bli-methods ↩

  13. Adaptyv Bio. (n.d.-a). Binding. Retrieved November 9, 2026, from https://docs.adaptyvbio.com/docs/experiment-types/binding ↩

  14. Pettersen et al. (2021). UCSF ChimeraX: Structure visualization for researchers, educators, and developers. Protein Science, 30(1), 70–82. DOI ↩