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 with the SMILES string O=S(=O)(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. It carries the CCD code P8S and is described in the campaign data as PFOS (perfluorooctanesulfonic acid). The molecule is a persistent environmental contaminant and a challenging small-molecule target for protein binder design because its perfluorinated tail is both hydrophobic and highly fluorinated, while its sulfonic acid head group presents a single strong hydrogen-bond donor and two hydrogen-bond acceptors. The ligand has a molecular weight of 500.13 Da, a computed cLogP of 4.84, one hydrogen-bond donor, two hydrogen-bond acceptors, seven rotatable bonds, and one Lipinski violation. Its QED drug-likeness score is 0.4025, and its synthetic accessibility score is 0.8044 on the standard 1-to-10 scale where 1 is easy. The ADMET aggregate risk score for the ligand is 0.4892, with a toxicity risk score of 0.7743.

These ligand-level properties matter because they define the chemical space the binder must recognize. PFOS is not a typical drug-like small molecule: the QED value of 0.4025 sits well below the usual medicinal chemistry comfort zone, and the single Lipinski violation reflects the molecule's size and lipophilicity. A binder campaign against PFOS therefore has to solve a recognition problem that is partly about charge complementarity to the sulfonate head group and partly about creating a pocket that can accommodate a long, rigid, fluorinated tail without paying an excessive desolvation penalty.

Ligand Preparation

The campaign began with RDKit conformer generation and partial charge assignment for PFOS. The ligand preparation stage produced the single molecular form described by the SMILES string and the CCD code P8S. The computed properties in the LIGAND PROPERTIES block are target-level facts: molecular weight 500.13 Da, cLogP 4.84, one hydrogen-bond donor, two hydrogen-bond acceptors, seven rotatable bonds, one Lipinski violation, QED 0.4025, synthetic accessibility 0.8044, ADMET aggregate risk 0.4892, and ADMET toxicity risk 0.7743.

The seven rotatable bonds in PFOS are distributed along the perfluorinated carbon chain, but the carbon-fluorine bonds themselves are not rotationally labile in the same way as carbon-hydrogen bonds in a typical alkyl chain. The fluorinated tail is better understood as a stiff, low-polarizability surface that resists the usual hydrophobic packing interactions. The sulfonic acid head group, by contrast, is a strong acid and will be deprotonated at physiological pH, presenting a negatively charged sulfonate group. That charge distribution is the key feature the pharmacophore fingerprinting stage was designed to capture.

Pharmacophore Analysis

The pharmacophore fingerprinting stage identified six features for PFOS. The feature counts are three Acceptor.SingleAtomAcceptor features, one Donor.SingleAtomDonor feature, and two NegIonizable.AcidicGroup features. The single donor corresponds to the sulfonic acid proton, and the two acidic group features reflect the ionizable character of the sulfonate head group. The three acceptor features are distributed across the oxygen atoms of the sulfonate group and the electron-rich fluorine environment.

This pharmacophore profile is compact but chemically decisive. A binder that recognizes PFOS must satisfy the hydrogen-bonding and charge-complementarity demands of the sulfonate head group while also providing a surface that can contact the fluorinated tail. The pharmacophore features were used to condition the single-site design stage, where the ligand itself served as the binding interface. No patch discovery was performed; the design problem was framed as building a protein pocket directly around the ligand's pharmacophore.

Binder Generation

The campaign generated 112 candidates using pipeline version 0.5.0. The source-tool split is 20 candidates from boltzgen and 92 candidates from boltzgen+proofread. All 112 candidates carry a generator confidence score. The generator prefilter accepted 20 candidates and excluded 0 candidates. The 92 candidates that joined the table after the prefilter stage are variant or union rows that were never screened by the prefilter.

The generation stage used PXDesign or BoltzGen, conditioned on the ligand structure, to produce complementary protein pockets. The 20 boltzgen candidates represent the primary generated designs, and the 92 boltzgen+proofread candidates are proofread variants of those designs. The per-tool median generator confidence score for boltzgen is 0.6695 across 20 scored candidates, and for boltzgen+proofread it is 0.6656 across 92 scored candidates. The boltzgen candidates also carry median solubility score 0.7976, median SAP3D spatial score -1.0732, median average packing density 15.7405, and median flexible loop fraction 0.4456.

The fact that the proofread variants have a median confidence score very close to the primary boltzgen designs is expected: the proofread variants inherit the parent design's generator confidence score, and the ranking reflects the parent's score propagated to its variants. The campaign's top candidates are therefore not eight independent designs but rather four distinct designs, with six of the eight listed rows being variants of two parent designs.

Top Candidates

The ranked table below lists the eight candidate rows from the campaign's top candidates block. The ranking is by generator confidence score. The table includes only columns that carry a real measured value on every listed row: generator confidence score, docking score, docking CNN pose score, docking CNN affinity, number of docking modes, best mode, and the key generator metrics that are shared across all rows.

The table's structure reflects the variant clustering in the data. The first six rows are all variants of the parent design BG_SM_DENOVO_P8S_E2E_boltzgen_binder_002: the parent itself plus five proofread variants named __pf1_00 through __pf1_04. All six rows carry the same generator confidence score of 0.8370 and the same docking score of -6.74 kcal/mol. The seventh and eighth rows are the parent design BG_SM_DENOVO_P8S_E2E_boltzgen_binder_004 and one proofread variant, both with generator confidence 0.8290 and docking score -5.15 kcal/mol.

The ranking therefore reflects two distinct parent designs, not eight independent candidates. The top-ranked parent design, BG_SM_DENOVO_P8S_E2E_boltzgen_binder_002, achieves a docking score of -6.74 kcal/mol, which is the best docking score in the campaign. The second parent design, BG_SM_DENOVO_P8S_E2E_boltzgen_binder_004, reaches -5.15 kcal/mol. The difference of 1.59 kcal/mol between the two parent designs is substantial in docking terms and suggests that the top parent design forms a more favorable binding interface with PFOS.

Docking and Developability

The docking stage used AutoDock Vina with 10 modes per candidate, and the best mode was mode 1 for every listed candidate. The top parent design and its five proofread variants all achieve a docking score of -6.74 kcal/mol, with a CNN pose score of 0.4327 and a CNN affinity value of 5.3310. The second parent design and its variant achieve -5.15 kcal/mol, with a CNN pose score of 0.3031 and a CNN affinity of 5.2120.

The generator-level metrics for the top parent design are strong. The parent BG_SM_DENOVO_P8S_E2E_boltzgen_binder_002 has a generator confidence score of 0.8370, a solubility score of 0.7496 with a soluble classification, a SAP3D spatial score of -1.3100, an average packing density of 16.8732, zero protease-exposed motifs, one buried unsatisfied polar atom, and a flexible loop fraction of 0.2254. Its generated interface metrics include an iPTM of 0.8037, a pLDDT of 0.7071, an RMSD of 0.8333 Å, a ligand iPTM of 0.8037, a minimum interaction PAE of 1.4983, a complex iP-LDDT of 0.5687, an interaction PAE of 8.6641, a complex PDE of 0.6341, a PTM of 0.8420, and a target PTM of 0.7863. The refolded structure shows three hydrogen bonds to the ligand, zero salt bridges, and a delta SASA of 339.8362 Ų. The affinity prediction value is 1.9295 with a binary probability of 0.0272.

The second parent design, BG_SM_DENOVO_P8S_E2E_boltzgen_binder_004, has a generator confidence score of 0.8290, a solubility score of 0.7382 with a soluble classification, a SAP3D spatial score of -1.8072, an average packing density of 15.0571, zero protease-exposed motifs, zero buried unsatisfied polar atoms, and a flexible loop fraction of 0.2857. Its interface metrics include an iPTM of 0.7843, a pLDDT of 0.6521, an RMSD of 0.6909 Å, a ligand iPTM of 0.7843, a minimum interaction PAE of 1.6525, a complex iP-LDDT of 0.5214, an interaction PAE of 9.0007, a complex PDE of 0.6120, a PTM of 0.8637, and a target PTM of 0.7411. The refolded structure shows one hydrogen bond to the ligand, zero salt bridges, and a delta SASA of 265.7109 Ų. The affinity prediction value is 2.0507 with a binary probability of 0.0335.

The developability metrics for the two parent designs show a clear contrast. The top parent has a higher average packing density (16.8732 versus 15.0571) and a lower flexible loop fraction (0.2254 versus 0.2857), which together suggest a more rigid and better-packed binding pocket. The top parent also has one buried unsatisfied polar atom, while the second parent has zero. The SAP3D spatial score is -1.3100 for the top parent and -1.8072 for the second parent, with more negative values indicating better spatial packing quality. The solubility scores are both in the soluble range, at 0.7496 and 0.7382 respectively.

The ligand-level ADMET assessment is target-level, not per-candidate. PFOS itself has an ADMET aggregate risk score of 0.4892 and a toxicity risk score of 0.7743. These values reflect the ligand's inherent toxicological profile, which is high for toxicity risk. The binder candidates are not themselves drug-like small molecules; they are proteins, and their developability is assessed through the solubility, packing, and spatial metrics reported above.

Validation Plan

The campaign's ordering stage produced order files for two planned experiments. The first is a yeast surface display screen using IDT oPools oligo pools covering all 112 candidates, tiered by confidence, for gap-repair cloning into the pCTCON2 vector. The second is an IVTT TR-FRET assay using 96-well IDT gBlock constructs for 20 candidates, with the T7 promoter plus GG plus SD plus Kozak plus ATG plus FLAG construct context using the E. coli MFC codon table.

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 order files and protocol documents are downloadable from the experiments tab.

Honest Assessment

This campaign produced a credible first-pass binder design for PFOS, with a clear top candidate and a well-defined variant cluster. The best docking score of -6.74 kcal/mol from the parent design BG_SM_DENOVO_P8S_E2E_boltzgen_binder_002 is a reasonable starting point for a small-molecule binder, and the fact that all five proofread variants of that parent retain the same docking score and generator confidence of 0.8370 suggests the design is stable under the proofreading perturbations. The second parent design, at -5.15 kcal/mol, is meaningfully weaker, and the campaign's ranking correctly places it below the top cluster.

What worked well is the integration of ligand preparation, pharmacophore fingerprinting, and single-site design. The six pharmacophore features for PFOS, dominated by the sulfonate head group's acceptor and acidic character, provided a focused conditioning signal. The generator produced 20 primary designs that all passed the prefilter, and the proofread expansion to 92 variants gave the campaign enough diversity to explore local sequence space around the top parents. The developability metrics for the top parent are encouraging: soluble classification, zero protease-exposed motifs, a SAP3D score of -1.3100, and a packing density of 16.8732 all point to a well-folded candidate.

The campaign's main limitation is the modest docking score ceiling. A docking score of -6.74 kcal/mol is not a high-affinity binder by small-molecule standards, and the CNN pose score of 0.4327 for the top parent indicates that the docked pose is only moderately confident. The affinity prediction probability of 0.0272 is low, and the affinity prediction value of 1.9295 suggests the generator's own affinity model is not strongly convinced of tight binding. The second parent design is weaker still, with a CNN pose score of 0.3031 and a docking score of -5.15 kcal/mol. The ligand's own ADMET toxicity risk of 0.7743 is a reminder that PFOS is a challenging target from a safety perspective, though the binder candidates themselves are proteins and are assessed through developability metrics rather than small-molecule ADMET.

A follow-up campaign would target the docking score directly. The top parent's interface metrics show room for improvement: the complex iP-LDDT of 0.5687 and the interaction PAE of 8.6641 suggest the binding interface is not as confidently modeled as the overall protein core. The buried unsatisfied polar count of 1.0 in the top parent is a specific liability that could be addressed by redesigning the pocket around the sulfonate head group. The flexible loop fraction of 0.2254 is acceptable but could be reduced to stiffen the binding site. The second parent design, with its lower packing density of 15.0571 and higher flexible loop fraction of 0.2857, would benefit from a more aggressive proofreading strategy or a different conditioning signal. The planned yeast surface display screen and IVTT TR-FRET assay are the right next steps, and the order files are ready for another lab to reproduce the campaign's candidates and test them experimentally.