Engineering De Novo Binders Against Staphylococcal Enterotoxin B with the AffiniBind Pipeline

GenX (HFPO-DA) Binder Design with AffiniBind

A ligand-conditioned binder campaign against GenX (HFPO-DA) generated 170 candidates, with a top docking score of -7.84 kcal/mol and a target QED of 0.806.

Why Staphylococcal Enterotoxin B Matters

Staphylococcal Enterotoxin B (SEB) is a bacterial superantigen that triggers massive, non-specific T-cell activation by cross-linking major histocompatibility complex class II molecules with T-cell receptors. The resulting cytokine storm can produce toxic shock syndrome, and SEB's exceptional stability, ease of aerosolization, and potency at extremely low concentrations have made it a persistent concern in both food safety and biodefense contexts. Neutralizing SEB requires a binding molecule that can physically block the superantigen's interaction surfaces while remaining stable enough for manufacturing and storage. Small-molecule approaches have struggled against SEB's relatively flat protein-protein interaction surfaces, which is precisely the problem that de novo protein binders are designed to solve.

This campaign applied the AffiniBind platform to SEB (PDB ID: STAPHYLOCOCCAL ENTEROTOXIN B) with the goal of generating a panel of candidate protein binders suitable for experimental validation. The pipeline ran from surface fingerprinting through patch identification, parallel binder generation, developability assessment, and final ranking. A total of 83 candidates were generated, with six binding patches identified and zero epitopes resolved at the structural level. The generator split was 13 candidates from bindcraft, 10 from boltzgen, and 60 from pxdesign.

Surface Fingerprinting and Patch Identification

The first stage of the AffiniBind pipeline builds a multi-scale surface fingerprint of the target. For SEB, the platform integrated ScanNet, MaSIF, MPBind, and APBS electrostatics to characterize surface geometry, chemical features, and charge distribution. These fingerprints are not used in isolation; they feed directly into patch identification, where contiguous or near-contiguous surface regions with favorable binding characteristics are segmented and ranked.

The campaign resolved six binding patches on SEB. The top three patches by residue count were:

• patch_0005: 26 residues
• patch_0004: 18 residues
• patch_0002: 14 residues

The remaining three patches were also identified but are not individually itemized in the campaign summary. Patch rank values are not available in the provided data, so the ordering above reflects residue count rather than a formal ranking score. What matters for downstream binder generation is that these patches represent the surface regions where the generator tools were directed to focus their design efforts.

Parallel Binder Generation

AffiniBind uses a parallel generation strategy rather than relying on a single design tool. For this campaign, the generator was FreeBindCraft, which couples ProteinMPNN sequence design with AlphaFold2 structure prediction. Three source tools contributed candidates: bindcraft (13 candidates), boltzgen (10 candidates), and pxdesign (60 candidates). The heavy representation from pxdesign reflects its role as the primary generator in this pipeline configuration, while bindcraft and boltzgen serve as orthogonal checks on design space coverage.

Each candidate receives a generator confidence score that reflects the generator's own assessment of structural plausibility and interface quality. For bindcraft candidates, this score is derived from the tool-specific pLDDT metric, which is coalesced into the shared confidence field for ranking purposes. The top five candidates by generator confidence score are:

The confidence spread across the top five is narrow, from 0.9180 down to 0.8728, which suggests that the generator did not identify a single dramatically superior candidate but rather a cluster of comparably plausible designs. The two bindcraft candidates share an identical confidence score of 0.8728, which is expected when the underlying pLDDT values round to the same coalesced score.

Developability Assessment

The AffiniBind developability assessment applies a twelve-dimension framework covering solubility, aggregation risk, charge state, binding affinity prediction, packing density, protease susceptibility, and manufacturing feasibility. For this campaign, the available per-candidate metrics include PRODIGY predicted binding ΔG, 3D-SAP spatial aggregation propensity, PROPKA isoelectric point, and a solubility score where reported.

The PRODIGY ΔG values span from -8.7000 kcal/mol (SEB_01_pxdesign_binder_031) to -11.0000 kcal/mol (SEB_01_bindcraft_binder_003). All five top candidates fall within a range that PRODIGY interprets as favorable for binding, though the tool's predictions are approximate and should not be read as experimental affinities. The two bindcraft candidates show the most negative ΔG values at -10.9000 and -11.0000 kcal/mol, while the pxdesign candidates cluster between -8.7000 and -10.6000 kcal/mol.

The 3D-SAP spatial scores are all negative, which is the desirable direction for this metric: more negative values indicate lower predicted aggregation propensity. SEB_01_pxdesign_binder_024 has the most favorable 3D-SAP score at -3.0555, while SEB_01_bindcraft_binder_010 has the least favorable at -1.6713. The spread here is meaningful: a difference of roughly 1.4 units on the 3D-SAP scale can translate to substantially different aggregation behavior during expression and purification.

PROPKA isoelectric points range from 4.5000 (SEB_01_bindcraft_binder_010) to 5.4000 (SEB_01_pxdesign_binder_024). All five candidates are acidic, which is generally favorable for solubility at neutral pH and for downstream formulation. The solubility score is reported only for the two bindcraft candidates: 0.6145 for SEB_01_bindcraft_binder_010 and 0.6819 for SEB_01_bindcraft_binder_003. The pxdesign candidates do not have solubility scores in the provided data, which reflects a difference in how the source tools report this metric rather than a gap in the pipeline's assessment.

Ranking and Clustering

The final ranking in this campaign is driven by generator confidence score, with developability metrics provided as supporting context rather than as hard filters. The top five candidates are all from either pxdesign or bindcraft; no boltzgen candidate appears in the top five, which may reflect the smaller number of boltzgen candidates generated (10 total) or differences in how boltzgen's confidence scores calibrate relative to the other tools.

The ranking reveals an interesting tension. The highest-confidence candidate, SEB_01_pxdesign_binder_011, has a PRODIGY ΔG of -9.5000 kcal/mol, which is less favorable than the -11.0000 kcal/mol of SEB_01_bindcraft_binder_003. If binding affinity prediction were the sole ranking criterion, the order would shift. The AffiniBind approach deliberately keeps generator confidence as the primary ranking signal because it reflects the generator's structural plausibility assessment, which correlates with the likelihood that the designed sequence actually folds into the predicted binding-competent structure. A candidate with a slightly less favorable predicted ΔG but higher structural confidence may be more likely to express, fold, and bind in practice.

Clustering was performed as part of the ranking stage, though the specific cluster assignments and cluster counts are not itemized in the campaign summary. The clustering step is designed to ensure that the final candidate panel does not over-represent a single design mode or patch, which would reduce the diversity of the experimental screen.

Validation Plan

The campaign's ordering stage produced order files for two planned experimental validation routes. These experiments have not been run; they are available and ready to execute, with all order files and protocol documents downloadable from the target's experiments tab.

The first planned experiment is a yeast surface display (YSD) screen using IDT oPools oligo pools covering all 83 candidates, tiered by confidence. The oligos are designed for gap-repair cloning into the pCTCON2 vector, which enables display of the candidate binders on the yeast surface for subsequent selection and characterization. The tiering by confidence means that the highest-ranked candidates from this campaign will be screened first, with lower-confidence candidates available as a secondary pool.

The second planned experiment is an IVTT TR-FRET assay using 96-well IDT gBlock constructs for all 83 candidates. Each construct includes the T7 promoter, GG, SD, Kozak, ATG, and FLAG tag context, using the E. coli MFC codon table. This cell-free expression and binding assay format allows rapid screening of the full candidate panel without the need for bacterial transformation and purification, which is particularly valuable for a panel of this size.

Honest Assessment

This campaign produced a coherent panel of 83 candidate binders against SEB, with six binding patches identified and a top-five ranking that spans both pxdesign and bindcraft sources. The generator confidence scores are tightly clustered between 0.8728 and 0.9180, which is both encouraging and cautionary: encouraging because it suggests the generator found multiple plausible design solutions, cautionary because it means the ranking does not strongly discriminate among the top candidates. The developability metrics add useful texture. The 3D-SAP scores all fall in the negative range, with SEB_01_pxdesign_binder_024 showing the most favorable aggregation profile at -3.0555. The PRODIGY ΔG values are all favorable, with the bindcraft candidates showing the most negative predictions. The PROPKA pI values are all acidic, which bodes well for solubility and formulation.

What the pipeline struggled with is less visible in the summary numbers. The campaign resolved zero epitopes at the structural level, which means the patch identification stage did not converge on a single dominant binding surface. The six patches vary in size from 14 to 26 residues, and without epitope resolution it is not known whether the top candidates target the same patch or spread across multiple patches. This is not necessarily a failure: SEB's superantigen function involves multiple interaction surfaces, and a diverse panel may be more valuable than a converged one. But it does mean that the experimental validation will need to characterize binding site specificity explicitly, since the computational stage did not resolve it.

A follow-up campaign would benefit from several adjustments. First, the boltzgen candidates are underrepresented in the top five, which may reflect calibration differences rather than genuinely inferior designs; a follow-up could generate more boltzgen candidates or recalibrate its confidence scores against the pxdesign and bindcraft scales. Second, the solubility score is available only for bindcraft candidates, which limits cross-tool comparison on this dimension; a follow-up could apply a uniform solubility prediction across all candidates regardless of source tool. Third, the narrow confidence spread among the top five suggests that the generator may be saturating a particular design mode; introducing additional diversity constraints or targeting different patches explicitly could broaden the panel. The planned YSD and IVTT TR-FRET experiments are the right next step, and their results will determine whether the generator confidence ranking holds up under experimental conditions. Until those experiments are run, the computational metrics presented here remain predictions, not validations.