PFOS Binder Design with AffiniBind: A Ligand-Conditioned Campaign

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.

The Target Ligand

Perfluorooctanesulfonic acid (PFOS) is a fully fluorinated eight-carbon sulfonic acid that has become one of the most persistent environmental contaminants ever measured. Its carbon-fluorine backbone resists metabolic and environmental degradation, while its terminal sulfonic acid group gives it surfactant properties and strong binding to serum proteins. The molecule carries a molecular weight of 500.13 Da, a computed cLogP of 4.84, and one hydrogen-bond donor against two acceptors. Seven rotatable bonds give the perfluorinated tail some conformational freedom, though the fluorinated chain is considerably stiffer than a hydrocarbon of equivalent length.

The ligand-level drug-likeness metrics tell a blunt story. PFOS has a QED score of 0.4025, one Lipinski violation, and a synthetic accessibility score of 0.8044 on the standard 1-to-10 scale where 1 is easiest. The ADMET aggregate risk score is 0.4892, driven substantially by a toxicity risk score of 0.7743. These are target-level facts, computed once for the ligand, and they frame the design problem: PFOS is not a drug candidate, it is a toxicant that a designed protein binder might sequester, detect, or help remediate. The campaign's goal was to generate protein binders that recognize PFOS as a small-molecule target, using the ligand itself as the binding interface.

Ligand Preparation

The ligand preparation stage began with the PFOS SMILES string: O=S(=O)(C(F)(F)C(F)(F)C(F)(F)C(F)(F)C(F)(F)C(F)(F)C(F)(F)C(F)(F)F. The CCD code for this ligand is P8S. RDKit conformer generation and partial charge assignment produced the three-dimensional ligand representation used throughout the campaign. The sulfonic acid head group was modeled in its acidic form, consistent with the pharmacophore analysis that identified two acidic-group features.

The prepared ligand served as the single-site design target. In this pipeline variant, the ligand is the binding interface: no patch discovery stage runs, and the generator is conditioned directly on the ligand structure to produce a complementary protein pocket. The ligand's seven rotatable bonds and the conformational preferences of the perfluorinated tail were encoded in the conformer ensemble passed to the generator.

Pharmacophore Analysis

The pharmacophore fingerprinting stage identified six features on PFOS. Three features are single-atom acceptors, one is a single-atom donor, and two are acidic-group features from the ionizable sulfonic acid. This feature set captures the ligand's hydrogen-bonding and charge-complementarity requirements: a designed binder must satisfy the sulfonic acid's strong negative charge while also engaging the fluorinated tail through shape complementarity and van der Waals contacts.

The pharmacophore features were combined with Morgan and MXFP molecular fingerprints and GNN embeddings from molfeat to produce the ligand representation that conditioned binder generation. The six-feature pharmacophore is compact but chemically decisive: the two acidic-group features dominate the electrostatic design problem, while the three acceptor features and single donor feature define the hydrogen-bonding geometry a successful pocket must present.

Binder Generation

The campaign generated 185 candidates using pipeline version 0.5.0. The source-tool split is 80 candidates from boltzgen and 105 from boltzgen+proofread. All 185 candidates carry a generator confidence score. The generator prefilter accepted 80 candidates and excluded 0; the 105 candidates that joined the table after the prefilter stage are variant or union rows that were never screened by the prefilter.

The per-tool medians show the proofread step's effect. The 80 boltzgen candidates have a median generator confidence score of 0.6513, a median solubility score of 0.8024, a median SAP3D spatial score of -1.2056, a median average packing density of 15.7463, and a median flexible loop fraction of 0.4123. The 105 boltzgen+proofread candidates have a median generator confidence score of 0.7574. The proofread step lifted the median confidence by roughly 0.106, a meaningful shift for a ranking metric.

The top-ranked candidates in the table below represent 4 distinct designs among the 8 listed rows. Six of the eight rows are variants of two parent designs: five variants of BG_SM_DENOVO_P8S_L80_E2E_boltzgen_binder_073 and one variant of BG_SM_DENOVO_P8S_L80_E2E_boltzgen_binder_044. The ranking reflects the parent's score propagated to its variants, so the table should be read as a design-level ranking, not as eight independent candidates.

Top Candidates

The table below lists the two parent designs that carry full metric profiles. The variant rows are described in prose following the table, using only the values listed on their own lines.

The top parent design, BG_SM_DENOVO_P8S_L80_E2E_boltzgen_binder_073, carries a generator confidence score of 0.8812, a solubility score of 0.8037 with a soluble classification, a SAP3D spatial score of -1.4982, an average packing density of 16.2368, and a flexible loop fraction of 0.2237. Its docking score is -5.25 kcal/mol with a GNINA CNN pose confidence of 0.4992 and a CNN affinity value of 5.4430. The design has one protease-exposed motif, zero buried unsatisfied polar atoms, and zero chain breaks. Its generated interface metrics include an iPTM of 0.8259, a ligand iPTM of 0.8259, a minimum interaction PAE of 1.5016, and an interaction PAE of 8.6145. The refolded structure shows a delta SASA of 257.4915 Ų, indicating substantial burial of surface area upon complex formation.

The second parent design, BG_SM_DENOVO_P8S_L80_E2E_boltzgen_binder_044, has a generator confidence score of 0.8580, a solubility score of 0.7180 with a soluble classification, a SAP3D spatial score of -1.8246, an average packing density of 14.9296, and a flexible loop fraction of 0.3239. Its generated interface metrics include an iPTM of 0.8297, a ligand iPTM of 0.8297, a minimum interaction PAE of 1.4584, and an interaction PAE of 7.2349. The refolded structure shows a delta SASA of 433.0145 Ų. This design has zero protease-exposed motifs and zero buried unsatisfied polar atoms.

The six variant rows are proofread versions of these two parents. Five variants of BG_SM_DENOVO_P8S_L80_E2E_boltzgen_binder_073 (__pf1_01, __pf1_02, __pf1_03, __pf1_00, __pf1_04) and one variant of BG_SM_DENOVO_P8S_L80_E2E_boltzgen_binder_044 (__pf1_03) were generated by the boltzgen+proofread source tool. Each variant carries a ligand_gate_status of not_run and a structure_source of refold. The proofread step's median confidence of 0.7574 across all 105 proofread candidates reflects the distribution these variants belong to.

Docking and Developability

The campaign's best docking score is -10.04 kcal/mol, a value that appears in the full campaign ranking. The top parent design's docking score of -5.25 kcal/mol places it in a moderate affinity range for a designed binder, while the best campaign score suggests that other candidates in the 185-member set achieved stronger predicted binding. The GNINA CNN pose confidence of 0.4992 for the top parent indicates a moderately confident pose prediction, and the CNN affinity value of 5.4430 corresponds to a predicted binding affinity in the micromolar range.

The developability metrics for the two parent designs show distinct profiles. The top parent has a higher average packing density (16.2368 versus 14.9296) and a lower flexible loop fraction (0.2237 versus 0.3239), suggesting a more rigid, better-packed core. Its SAP3D spatial score of -1.4982 is less negative than the second parent's -1.8246, indicating fewer spatial aggregation-prone regions. Both designs have zero buried unsatisfied polar atoms and zero chain breaks, and both are classified as soluble.

The ligand-level ADMET aggregate risk score of 0.4892 and toxicity risk score of 0.7743 apply to PFOS itself, not to the protein binders. These values underscore the target's toxicological profile and the motivation for designing a binder that could sequester or detect it. The ligand's QED of 0.4025 and one Lipinski violation are target-level facts that do not bear on the protein candidates' developability.

Validation Plan

The campaign's ordering stage produced order files for two planned experiments, both available for reproduction from the target's experiments tab. These experiments have not been run; the order files and protocol documents are downloadable and ready for another lab to execute.

The first planned experiment is a yeast surface display screen using IDT oPools oligo pools covering all 185 candidates, tiered by confidence, for gap-repair cloning into the pCTCON2 vector. This screen would provide experimental binding data for the full candidate set.

The second planned experiment is an IVTT TR-FRET assay using 96-well IDT gBlock constructs covering 80 candidates. The construct context includes the T7 promoter, GG, SD, Kozak, ATG, and FLAG tag, using the E. coli MFC codon table. This assay would provide quantitative binding measurements for the top-scoring subset.

Honest Assessment

This campaign produced a credible first-pass binder design set for PFOS, with a top generator confidence score of 0.8812 and a best docking score of -10.04 kcal/mol. The pipeline's single-site design strategy, conditioning directly on the ligand structure, generated 185 candidates across two source tools, with the proofread step lifting median confidence from 0.6513 to 0.7574. The pharmacophore analysis correctly identified the sulfonic acid's two acidic-group features as the dominant electrostatic design constraint, and the generated designs show zero buried unsatisfied polar atoms, suggesting the generator successfully placed complementary polar groups.

What worked: the boltzgen generator produced designs with high solubility scores (0.8037 and 0.7180 for the two parents), low SAP3D spatial scores, and zero chain breaks. The refolded structures show substantial delta SASA values (257.49 and 433.01 Ų), indicating real interface burial. The proofread step's confidence lift is consistent and meaningful.

What the pipeline struggled with: the docking scores for the top-ranked parents are moderate (-5.25 kcal/mol for the top parent), well above the campaign best of -10.04 kcal/mol. This suggests that generator confidence and docking affinity are not perfectly correlated, and the final ranking may prioritize designs that are developable but not maximally tight binders. The flexible loop fractions (0.2237 and 0.3239) are moderate, and the second parent's lower packing density (14.9296) may indicate a less stable core.

A follow-up campaign would target tighter docking scores while preserving the developability profile. The gap between the top parent's -5.25 kcal/mol and the campaign best of -10.04 kcal/mol suggests that stronger binders exist in the 185-candidate set but were not ranked at the top by generator confidence. Re-ranking with a docking-weighted objective, or running a second round of proofreading on the highest-docking candidates, would be a direct next step. The planned YSD screen and IVTT TR-FRET assay, once executed, will provide the experimental binding data needed to calibrate the computational ranking and identify which metric, generator confidence or docking score, better predicts true PFOS binding.