Staphylococcal Enterotoxin B: A Computational Binder Campaign on a Superantigen Target

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 This Target Matters

Staphylococcal enterotoxin B (SEB) is a bacterial superantigen that triggers massive, nonspecific T-cell activation by cross-linking major histocompatibility complex class II molecules with T-cell receptors outside the conventional peptide-binding groove. The result is a cytokine storm that can progress to toxic shock. SEB is also classified as a select agent because of its stability, its potency at very low doses, and its history of consideration as a bioweapon. A high-affinity protein binder against SEB would have immediate value as a detection reagent, as a potential neutralizing modality, and as a structural probe for understanding how superantigen engagement differs from conventional antigen recognition.

This campaign used the AffiniBind platform, version 0.5.0, to design de novo protein binders against SEB. The target structure is deposited under PDB ID STAPHYLOCOCCAL ENTEROTOXIN B. The pipeline ran from surface fingerprinting through patch identification, parallel binder generation, developability assessment, and finally generator confidence ranking with clustering. Thirty-three candidates were generated. Six binding patches were identified. One epitope was resolved. The generator was FreeBindCraft, which couples ProteinMPNN sequence design with AlphaFold2 structure validation.

The engineering-first philosophy matters here. Every stage of this campaign produced an auditable data contract, and candidates advanced through explicit gates. The computational predictions are designed to reduce experimental risk, not replace experimentation. The validation experiments described at the end of this article are planned and available for reproduction; they have not been run.

Surface Fingerprinting

The first stage of the AffiniBind pipeline builds a multidimensional picture of the target surface. For SEB, the platform integrated four complementary surface analysis tools: ScanNet for geometric and evolutionary pocket detection, MaSIF for surface patch featurization based on geometric and chemical fingerprints, MPBind for binding propensity estimation, and APBS for electrostatic potential mapping. Each tool contributes a different view of where a designed binder might productively engage the target.

The output of this stage is not a single score but a set of surface fingerprints that feed into patch identification. The pipeline does not treat any one tool as authoritative. Instead, the fingerprints are combined to identify contiguous surface regions that show agreement across multiple modalities: geometric accessibility, chemical complementarity potential, and electrostatic character.

For SEB, this stage identified six binding patches. The patch ranking is reported by patch identifier and residue count. Patch 0002 contains 14 residues. Patch 0005 contains 26 residues. Patch 0003 contains 25 residues. The remaining three patches are not individually itemized in the campaign summary, but their existence is accounted for in the total of six.

The patch sizes are informative. Patch 0005, at 26 residues, is the largest identified patch and represents a substantial contiguous surface area. Patch 0003, at 25 residues, is nearly as large. Patch 0002, at 14 residues, is more compact. In de novo binder design, patch size matters because it constrains the minimum contact area a binder must achieve to form a stable interface. A 14-residue patch can support a smaller, more focused binder interface, while a 26-residue patch offers more surface for distributed contacts but also demands a larger designed binding surface to achieve shape complementarity.

Patch Identification and Epitope Resolution

From the six identified patches, the pipeline resolved one epitope. The campaign summary does not specify which of the six patches corresponds to the resolved epitope, nor does it provide the residue-level composition of the epitope beyond the patch sizes already listed. What is known is that the pipeline converged on a single epitope as the primary target for binder generation.

This convergence is notable. In many campaigns, multiple patches proceed to binder generation in parallel, and the ranking stage then selects among binders targeting different epitopes. Here, the pipeline resolved one epitope from the six candidates, which suggests that the surface fingerprinting and patch identification stages produced a clear prioritization signal. The resolved epitope became the anchor for all downstream binder generation.

The epitope resolution step is where the pipeline transitions from descriptive surface analysis to a concrete design target. The resolved epitope defines the contact surface that generated binders are expected to engage. Without residue-level detail in the campaign summary, the specific molecular features of the resolved epitope are not known from this data contract. What is known is that the pipeline committed to a single epitope and generated all 33 candidates against it.

Parallel Binder Generation

Binder generation used FreeBindCraft, which combines ProteinMPNN for sequence design with AlphaFold2 for structure prediction and validation. The generator produced 33 candidates across three source-tool categories. Fifteen candidates came from bindcraft alone. Eight candidates came from bindcraft with proofreading, a variant that applies an additional validation pass to the generated sequences. Ten candidates came from boltzgen.

The source-tool split matters for interpreting the candidate ranking. The top five candidates by generator confidence score all derive from a single parent binder, SEB_02_bindcraft_binder_005. The top-ranked candidate is the parent itself, from the bindcraft source tool. The next four candidates are proofread variants of that parent, designated with the suffix pattern __pf1_01 through __pf1_09. This clustering is not accidental. The proofread variants share the same generator confidence score as the parent, 0.8820, which indicates that the proofreading step preserved the core structural prediction while exploring sequence variations around it.

The generator confidence score is the primary ranking metric in this campaign. It is built from the generator's own confidence metrics, including predicted local distance difference test values where applicable. For bindcraft candidates, the generator-confidence pLDDT is reported under the tool-specific column gen_binder_plddt, and the top candidates block coalesces that value into gen_plddt for those rows. The score of 0.8820 for the top five candidates represents a high-confidence structural prediction from the generator.

The boltzgen candidates, ten in total, do not appear in the top five. This does not mean they failed. The campaign summary reports only the top five candidates by generator confidence score. The boltzgen candidates may have lower confidence scores, or they may have been deprioritized for other reasons. Without the full ranking table, the distribution of scores across all 33 candidates is not known from this data contract.

Developability Assessment

The AffiniBind platform assesses every candidate across twelve developability dimensions. These dimensions cover solubility, aggregation risk, charge state, binding affinity, packing density, protease susceptibility, and manufacturing feasibility. The specific tools and metrics represented in the top candidates block include PRODIGY for binding affinity prediction, a solubility score, 3D-SAP for spatial aggregation propensity, and PROPKA for isoelectric point estimation.

The top five candidates show a clear pattern in the available developability metrics. The parent binder, SEB_02_bindcraft_binder_005, has the most complete developability profile. Its PRODIGY binding affinity prediction is -9.5000 kcal/mol. Its solubility score is 0.6486. Its 3D-SAP spatial score is -1.5337. Its PROPKA isoelectric point is 5.8000.

The first proofread variant, SEB_02_bindcraft_binder_005__pf1_03, shows a slightly less favorable PRODIGY prediction at -8.8000 kcal/mol, a 3D-SAP spatial score of -1.4887, and a PROPKA isoelectric point of 5.7000. Its solubility score is not reported in the top candidates block.

The remaining three proofread variants, __pf1_09, __pf1_01, and __pf1_08, have no developability metrics reported in the top candidates block. Their PRODIGY, solubility, 3D-SAP, and PROPKA values are all absent from the data contract.

The developability metrics that are available tell a coherent story. The parent binder has a strongly favorable PRODIGY prediction at -9.5000 kcal/mol, which suggests a stable predicted binding interface. The 3D-SAP spatial score of -1.5337 is negative, which in the 3D-SAP framework indicates lower spatial aggregation propensity relative to a reference distribution. The solubility score of 0.6486 is on a scale where higher values indicate better predicted solubility. The PROPKA isoelectric point of 5.8000 places the binder in the mildly acidic range, which is relevant for formulation and purification conditions.

The first proofread variant shows a modest reduction in predicted binding affinity, from -9.5000 to -8.8000 kcal/mol, and a slight shift in 3D-SAP from -1.5337 to -1.4887. The isoelectric point shifts from 5.8000 to 5.7000. These are small changes, consistent with a proofreading step that explores sequence variations around the parent while preserving the core structural prediction.

The table makes the clustering pattern explicit. All five top candidates share the same generator confidence score. The parent has the most complete developability profile. The first proofread variant has partial developability data. The remaining three proofread variants have generator confidence scores but no developability metrics in the top-level summary.

Ranking and Clustering

The final stage of the AffiniBind pipeline ranks candidates by generator confidence score and clusters them by sequence or structural similarity. The clustering is evident in the top five: one parent and four proofread variants of that parent. The proofread variants are designated by the suffix __pf1_XX, where the number indicates the proofreading iteration or variant index.

The ranking is dominated by a single design lineage. This is a double-edged result. On one hand, it indicates that the generator converged on a high-confidence structural solution for the resolved epitope, and the proofreading step explored variations around that solution without degrading the confidence score. On the other hand, the top five candidates are all variations on the same parent, which means the effective diversity of the top-ranked set is low. The boltzgen candidates, which represent a different generation approach, do not appear in the top five.

The campaign generated 33 candidates total. The top five represent approximately 15 percent of the total. The remaining 28 candidates are not itemized in the top candidates block, so their generator confidence scores and developability metrics are not known from this data contract. The source-tool split indicates that 15 candidates came from bindcraft, 8 from bindcraft with proofreading, and 10 from boltzgen. The top five are drawn entirely from the bindcraft and bindcraft+proofread categories.

Validation Plan

The campaign's ordering stage produced order files for two planned experimental validation workflows. These experiments are planned and available for reproduction. They have not been run. No Kd values, titration curves, or other experimental results exist for this campaign.

The first planned experiment is a yeast surface display screen. The order files specify IDT oPools oligo pools covering 25 candidates, tiered by confidence, for gap-repair cloning into the pCTCON2 vector. Yeast surface display is a standard method for screening binder libraries against a target of interest, and the tiering by confidence means the highest-ranked candidates are prioritized for synthesis and screening.

The second planned experiment is an in vitro transcription-translation time-resolved fluorescence resonance energy transfer assay. The order files specify 96-well IDT gBlock constructs covering 15 candidates, with the T7 promoter plus GG plus SD plus Kozak plus ATG plus FLAG construct context, using the E. coli MFC codon table. IVTT TR-FRET allows rapid, cell-free expression of candidate binders and assessment of target binding without the need for protein purification.

Both experiments are available for another lab to reproduce. The order files and protocol documents are downloadable from the target's experiments tab: /targets/protein/staphylococcal-enterotoxin-b/experiments.

Honest Assessment

This campaign produced a focused set of high-confidence binder candidates against staphylococcal enterotoxin B, with a clear top-ranked lineage and a coherent developability profile for the parent binder. The pipeline worked as designed: surface fingerprinting identified six patches, patch identification resolved one epitope, parallel binder generation produced 33 candidates across three source tools, and the ranking stage converged on a single parent with four proofread variants. The parent binder's PRODIGY prediction of -9.5000 kcal/mol is strongly favorable, and its 3D-SAP spatial score of -1.5337 and solubility score of 0.6486 suggest a developable candidate.

What the pipeline struggled with is diversity. The top five candidates are all variations on a single parent, which means the effective diversity of the top-ranked set is low. The boltzgen candidates, which represent a different generation approach, did not break into the top five. This is not necessarily a failure; the generator may have found a genuinely strong solution for the resolved epitope, and the proofread variants may offer useful sequence diversity within that structural solution. But it does mean that the campaign's experimental risk is concentrated in a single design lineage.

A follow-up campaign would target diversity more explicitly. The six identified patches offer multiple potential epitopes, and the current campaign resolved only one. Targeting additional epitopes would spread experimental risk across independent design solutions. Adjusting the generator's sampling parameters or incorporating additional generation tools could also increase the diversity of the top-ranked set. The developability assessment for three of the five top candidates is not reported in the top-level summary, which limits the ability to compare proofread variants on developability grounds. A follow-up campaign would benefit from carrying forward the full developability panel for all top-ranked candidates, so that ranking decisions can incorporate developability metrics alongside generator confidence.

The validation plan is ready to run. The YSD screen and IVTT TR-FRET assay are both planned and available for reproduction, with order files and protocols downloadable from the experiments tab. The next step is experimental validation. The computational pipeline has done its job: it produced ranked candidates with auditable data contracts at every stage. Whether the top-ranked binder actually binds SEB with high affinity is a question that only the planned experiments can answer.