PFOA 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

Perfluorooctanoic acid (PFOA) is a fully fluorinated eight-carbon carboxylic acid that has become one of the most scrutinized environmental contaminants of the past two decades. Its carbon-fluorine bonds give it exceptional chemical stability, and its amphiphilic character, a charged carboxylate head with a hydrophobic perfluorinated tail, lets it associate with proteins, lipid membranes, and serum albumin in ways that complicate both its clearance and its detection. The molecule is small enough to slip into cryptic binding sites but rigid enough that a designed protein pocket must match its exact contour: six rotatable bonds, one hydrogen-bond donor, one hydrogen-bond acceptor, and a molecular weight of 414.06 Da.

The campaign described here treats PFOA as the binding interface itself. Rather than discovering a patch on a larger protein surface, the pipeline conditions binder generation directly on the ligand structure. That is a demanding test of ligand-conditioned design. PFOA offers only a single carboxylate for polar recognition and a long fluorocarbon chain that must be buried productively if a binder is to achieve specificity. The CCD code for the ligand is 8PF, and the SMILES string used throughout the campaign is O=C(O)C(F)(F)C(F)(F)C(F)(F)C(F)(F)C(F)(F)C(F)(F)C(F)(F)F.

Ligand Preparation

The ligand preparation stage began with RDKit conformer generation and partial charge assignment. PFOA is not a flexible molecule in the conventional sense: the six rotatable bonds are all carbon-carbon single bonds along the perfluorinated chain, and the fluorines create substantial steric and electronic constraints on the accessible conformers. The carboxylate head can rotate relative to the chain, but the chain itself tends to adopt extended or gently curved conformations rather than compact folds.

The computed ligand properties frame the design problem. The cLogP of 4.45 reflects the strongly hydrophobic character of the perfluorinated tail, while the single carboxylic acid contributes one donor and one acceptor. The molecule has zero Lipinski violations, which is expected for a compound of this size, but that does not make it a drug-like target in the conventional sense. The QED drug-likeness score is 0.653, and the synthetic accessibility score is 0.814 on the SAScore scale where 1 is easy and 10 is hard. The ADMET aggregate risk score is 0.395, with a toxicity risk score of 0.698. These target-level values describe the ligand itself, not the protein binders, and they matter because a binder that recognizes PFOA must accommodate a molecule with substantial hydrophobic surface area and a single polar anchor point.

Pharmacophore Analysis

The pharmacophore fingerprinting stage identified four features on PFOA. Two acceptor features come from the carboxylate oxygens, one donor feature from the carboxylic acid proton, and one acidic group feature from the ionizable carboxylate. That is a sparse pharmacophore for a molecule of this size. The entire recognition surface reduces to a single acidic group with donor and acceptor character, while the fluorinated chain contributes no classical hydrogen-bonding features at all.

This has direct consequences for binder design. A protein pocket that recognizes PFOA must derive most of its binding energy from shape complementarity to the fluorocarbon chain and from electrostatic or hydrogen-bonding interactions with the carboxylate. The pharmacophore model tells the generator where polar contacts are possible, but it also makes clear that the chain itself must be buried against hydrophobic or van der Waals surfaces. The absence of aromatic rings, amides, or other feature-rich groups means the generator cannot rely on a dense network of directional interactions.

Binder Generation

The campaign generated 115 candidates in total. Of these, 20 came from the BoltzGen generator alone, and 95 came from the BoltzGen plus proofread pipeline. The generator prefilter scored all 115 candidates, and 20 passed. The remaining 95 candidates joined the table after the prefilter stage as variant or union rows, meaning they were never screened by the prefilter. The source-tool split is therefore clean: 20 candidates from boltzgen and 95 from boltzgen+proofread.

The median generator confidence score for the 20 BoltzGen candidates is 0.6243. The median solubility score for that group is 0.7896, the median SAP3D spatial score is -1.2078, the median average packing density is 15.4224, and the median flexible loop fraction is 0.3269. The proofread group of 95 candidates has a median generator confidence score of 0.6256. The two source tools produce comparable confidence distributions, which suggests the proofread step preserves the generator's ranking signal while expanding the candidate pool.

The top-ranked candidates in the table below represent four distinct designs. Six of the eight listed rows are variants of two parent designs, and the ranking reflects the parent's score propagated to its variants. The top cluster comes from parent BG_SM_DENOVO_8PF_E2E_boltzgen_binder_008, which appears once as the parent row and five times as proofread variants. The second cluster comes from parent BG_SM_DENOVO_8PF_E2E_boltzgen_binder_019, with one parent row and one variant row.

Top Candidates

The table below compares the eight listed candidates across the metrics that are measured on every row: generator confidence score, docking score, and the key generator quality metrics.

The ranking is dominated by a single parent design. BG_SM_DENOVO_8PF_E2E_boltzgen_binder_008 and its five proofread variants occupy the top six positions, all carrying a generator confidence score of 0.8557 and a docking score of -7.38 kcal/mol. The proofread variants inherit the parent's generator metrics exactly: gen_iptm of 0.8020, gen_plddt of 0.6298, and gen_rmsd of 0.5777 Å. The second design, BG_SM_DENOVO_8PF_E2E_boltzgen_binder_019, appears at positions seven and eight with a confidence score of 0.8535 and a docking score of -7.22 kcal/mol. Its gen_iptm is slightly higher at 0.8149, and its gen_rmsd is 0.7677 Å.

The two parent designs are close in confidence, separated by only 0.0022, but they differ in their interaction metrics. The binder_008 parent has a gen_interaction_pae of 7.2316 and a gen_complex_pde of 0.6715, while binder_019 has a gen_interaction_pae of 7.3400 and a gen_complex_pde of 0.6217. The binder_008 design also shows a lower gen_affinity_pred_value of 1.3130 compared to 2.0374 for binder_019, and a higher gen_affinity_probability_binary of 0.3841 versus 0.2079. These differences suggest the top-ranked design places the ligand in a slightly more confident predicted binding mode, even though both designs bury substantial surface area: gen_delta_sasa_refolded is 294.86 Ų for binder_008 and 371.41 Ų for binder_019.

Docking and Developability

The docking stage used AutoDock Vina with ten modes per candidate, and the best mode was mode 1 for every listed candidate. The best docking score across the entire campaign is -7.93 kcal/mol, which is stronger than the -7.38 kcal/mol and -7.22 kcal/mol scores carried by the top listed candidates. The docking scores for the binder_008 cluster are all -7.38 kcal/mol, and the binder_019 cluster scores are -7.22 kcal/mol. The GNINA CNN pose scores are 0.5239 for binder_008 and 0.3771 for binder_019, with CNN affinity values of 4.5660 and 4.0580 respectively.

The developability metrics for the two parent designs are informative. The binder_008 parent has a solubility score of 0.6301 and is classified as soluble. Its SAP3D spatial score is -2.1294, its average packing density is 14.6296, and its flexible loop fraction is 0.2407. The binder_019 parent has a higher solubility score of 0.8678, also classified as soluble, with a SAP3D spatial score of -1.3713, an average packing density of 16.2857, and a flexible loop fraction of 0.3714. Both parents have zero buried unsatisfied polar atoms and zero chain breaks. The binder_008 parent has three protease-exposed motifs, while binder_019 has one.

The generator quality metrics reinforce the ranking. Both parent designs have gen_raw_chain_breaks of 0.0000 and gen_has_unplaced_atoms of False, with gen_unplaced_atoms_frac of 0.0000. The gen_max_consecutive_c_n_a values are 1.3900 for binder_008 and 1.4000 for binder_019, both indicating clean backbone geometry. The gen_plip_hbonds_refolded and gen_plip_saltbridge_refolded values are 0.0000 for both designs, which is consistent with the sparse pharmacophore: the generator does not predict classical hydrogen bonds or salt bridges to the ligand in the refolded structures, relying instead on shape complementarity and van der Waals burial.

Validation Plan

The campaign's ordering stage produced order files for two planned experiments, both available for reproduction from the target's experiments tab. The first is a yeast surface display screen using IDT oPools oligo pools covering all 115 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 ready to run; they have not been executed, and no binding affinities or titration data exist yet. The order files and protocol documents are downloadable from the experiments tab.

Honest Assessment

This campaign produced a coherent set of ligand-conditioned binders for PFOA, with a clear top design cluster and a defensible ranking. The best docking score of -7.93 kcal/mol across the campaign is respectable for a small, feature-poor ligand, and the top listed candidates at -7.38 kcal/mol and -7.22 kcal/mol are consistent with the generator confidence ordering. The pipeline's strength here is the tight integration between the pharmacophore model and the generator: the four-feature pharmacophore correctly identifies the carboxylate as the only polar anchor, and the generator responds by burying the fluorocarbon chain against hydrophobic surfaces while placing the carboxylate in a position that the docking stage scores favorably.

What worked less well is the diversity of the top-ranked designs. The eight listed rows collapse to four distinct designs, and six of those rows are variants of a single parent. The proofread step expands the candidate count from 20 to 115, but the top of the ranking is dominated by one design family. That is not necessarily a failure: a high-confidence design with multiple proofread variants is a reasonable outcome for a difficult target. But it does mean the campaign's effective design diversity at the top is narrower than the raw candidate count suggests. A follow-up campaign would benefit from explicit diversity constraints during generation, perhaps by penalizing sequence or structural similarity among high-confidence candidates, or by increasing the number of independent BoltzGen seeds before the proofread expansion.

The ligand itself imposes hard limits. PFOA has one donor, one acceptor, and one acidic group, and the rest of the molecule is a fluorocarbon chain with no classical hydrogen-bonding features. The generator's PLIP analysis finds zero hydrogen bonds and zero salt bridges in the refolded structures, which means the predicted binding relies almost entirely on shape complementarity and van der Waals burial. That is a plausible mechanism for a perfluorinated ligand, but it also means the docking scores and confidence metrics are doing more work than they would for a feature-rich small molecule. The buried surface areas of 294.86 Ų and 371.41 Ų are substantial, and the GNINA CNN pose scores of 0.5239 and 0.3771 suggest the top design places the ligand in a more confident pose than the second design, but the absence of directional polar contacts leaves less room for error in the predicted geometry.

The developability metrics are encouraging. Both parent designs are classified as soluble, with solubility scores of 0.6301 and 0.8678, and both have zero buried unsatisfied polar atoms. The SAP3D spatial scores of -2.1294 and -1.3713 are within a range that suggests acceptable spatial packing, and the average packing densities of 14.63 and 16.29 are reasonable for designed binders. The flexible loop fractions of 0.2407 and 0.3714 differ notably, with binder_019 carrying more flexible loop content, which may explain its slightly lower docking score despite its larger buried surface area.

The planned validation experiments are the right next step. The yeast surface display screen covers all 115 candidates and will provide a direct read on whether the confidence ranking translates into actual binding. The IVTT TR-FRET assay on 20 candidates will give a quantitative affinity read for the top designs. The key question for the follow-up campaign is whether the binder_008 cluster's predicted binding mode, with its lower buried surface area but higher confidence and better docking score, holds up experimentally against the binder_019 design's larger interface. If the experimental data confirm the ranking, the pipeline's emphasis on generator confidence over raw buried area is validated. If not, the follow-up should reweight the ranking toward docking score and buried surface area, and should push the generator to produce more independent designs rather than expanding a small number of parents through proofreading.