PFNA Binder Design with AffiniBind: A Perfluorinated Acid 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.

The Target Ligand

Perfluorononanoic acid (PFNA) is a nine-carbon perfluorinated carboxylic acid that belongs to the broader class of per- and polyfluoroalkyl substances. Its fully fluorinated tail creates an unusually hydrophobic and chemically inert surface, while the terminal carboxylic acid provides the only polar handle for molecular recognition. That combination makes PFNA a demanding target for protein binder design: the ligand offers a single hydrogen-bond donor and a single acceptor, both concentrated at the carboxyl head, leaving the fluorinated chain to be accommodated largely through shape complementarity and desolvation.

The campaign targeted PFNA as the binding interface itself. In the single-site design mode, the ligand is the interface; there is no separate patch discovery step. The pipeline therefore had to generate protein pockets that could wrap around a 464.07 Da ligand with a cLogP of 5.08 and seven rotatable bonds, while tolerating one Lipinski violation. The target-level QED of 0.534 and ADMET aggregate risk score of 0.407 frame the ligand as a challenging but tractable small molecule, with the toxicity risk score of 0.676 reflecting the known persistence and bioaccumulation concerns associated with long-chain perfluorinated acids.

Ligand Preparation

The ligand preparation stage began with the PFNA SMILES string:

O=C(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

RDKit conformer generation produced the three-dimensional starting geometry, and partial charges were assigned for downstream docking and interaction scoring. The CCD code for PFNA is 4I6, which anchors the ligand identity in the structural data ecosystem.

The computed ligand properties define the chemical space the binder design had to address. PFNA carries a molecular weight of 464.07 Da, placing it at the upper edge of typical fragment-like targets and well into small-molecule territory. The cLogP of 5.08 indicates strong lipophilicity, consistent with the perfluorinated tail. With one hydrogen-bond donor and one hydrogen-bond acceptor, both on the carboxylic acid, the ligand presents a minimal polar pharmacophore. The seven rotatable bonds are all carbon-carbon bonds along the fluorinated chain, giving the ligand conformational flexibility that the conformer generation step had to sample. The synthetic accessibility score of 0.812 on the SAScore scale (where 1 is easy and 10 is hard) reflects that PFNA itself is a readily accessible compound, though that score describes the ligand, not the protein binders designed against it.

Pharmacophore Analysis

The pharmacophore fingerprinting stage identified four features on PFNA. Two acceptor features arise from the single-atom acceptor character of the carboxylic acid oxygens, one donor feature from the acidic proton, and one negatively ionizable feature from the acidic group. This is a compact pharmacophore: the entire polar recognition surface of PFNA is concentrated at the carboxyl head.

The pharmacophore features were used to condition the single-site design stage. Because the ligand is the interface, the pharmacophore fingerprinting directly informed which protein side-chain chemistries could complement the carboxyl head: hydrogen-bond donors to engage the acceptor oxygens, hydrogen-bond acceptors to engage the acidic proton, and a positively charged or polar residue to recognize the deprotonated acid at physiological pH.

Binder Generation

The binder generation stage produced 107 candidates using two source tools. The primary generator, boltzgen, contributed 20 candidates that passed the generator prefilter. The proofread pipeline, boltzgen+proofread, contributed 87 candidates that joined the table after the prefilter stage as variant or union rows. All 107 candidates carried a generator confidence score.

The prefilter behavior is notable: 20 candidates passed the generator prefilter, and 0 candidates were excluded by it. The 87 proofread candidates joined the table after the prefilter stage, meaning they were never screened by that particular filter. This source-tool split matters for interpreting the ranking: the boltzgen candidates represent the generator's direct output after prefiltering, while the boltzgen+proofread candidates represent a refinement pass that expanded the candidate pool.

The median generator confidence score for the 20 boltzgen candidates was 0.7744. The 87 boltzgen+proofread candidates had a median generator confidence score of 0.7768. The two source tools produced comparable confidence distributions, with the proofread pipeline slightly higher at the median.

Top Candidates

The ranked table below lists the eight candidate rows from the campaign. These eight rows represent three distinct designs: one parent design with five proofread variants, one parent design with two proofread variants, and one standalone boltzgen candidate. The seven variant rows are proofread refinements of the two parent designs.

The table reads as a story of two parent designs. The top six rows are all the same parent design, BG_SM_DENOVO_4I6_E2E_boltzgen_binder_001, with five proofread variants carrying the __pf1_00 through __pf1_04 suffixes. The parent and all five variants share a generator confidence score of 0.9289 and a docking score of -9.00 kcal/mol. The GNINA CNN pose score of 0.5940 and CNN affinity of 4.5370 are also identical across this cluster, reflecting that the proofread variants preserved the parent's binding mode.

The seventh and eighth rows are variants of a second parent design, BG_SM_DENOVO_4I6_E2E_boltzgen_binder_007, with the __pf1_01 and __pf1_04 suffixes. This cluster carries a generator confidence score of 0.8868 and a docking score of -8.76 kcal/mol. The GNINA CNN pose score of 0.7896 is higher than the first cluster's 0.5940, and the CNN affinity of 5.8610 is also higher, suggesting that while the second design docks slightly less favorably by Vina score, the CNN model sees a more native-like pose.

The ranking by generator confidence score places the first parent design at the top, but the docking and CNN metrics tell a more nuanced story. The first cluster wins on Vina docking score (-9.00 versus -8.76 kcal/mol), while the second cluster wins on CNN pose confidence (0.7896 versus 0.5940) and CNN affinity (5.8610 versus 4.5370). A follow-up campaign would need to weigh these trade-offs carefully.

Docking and Developability

The docking validation used AutoDock Vina with 10 modes per candidate, and the best mode was mode 1 for all eight listed candidates. The best docking score across the campaign was -9.76 kcal/mol, which is stronger than the -9.00 kcal/mol achieved by the top-ranked parent design in the table. The best QED score across the campaign was 0.534, matching the target ligand's own QED of 0.5343. The best ADMET risk score was 0.407, again matching the target-level ADMET aggregate risk of 0.4067.

The developability metrics for the top-ranked parent design, BG_SM_DENOVO_4I6_E2E_boltzgen_binder_001, are strong. The solubility score of 0.8176 places it in the soluble class. The SAP3D spatial score of -1.2544 indicates favorable spatial packing. The average packing density of 15.2941 and flexible loop fraction of 0.2353 suggest a well-ordered core with limited flexible regions. The protease-exposed motif count of 2.0000 and buried unsatisfied polar count of 0.0000 are both favorable for stability and expression.

The generator-level metrics for this parent design reinforce its quality. The gen_iptm of 0.8852 and gen_plddt of 0.6680 indicate a confident fold prediction. The gen_rmsd of 0.7027 Å suggests the refolded structure is close to the generated model. The gen_ligand_iptm of 0.8852 and gen_min_interaction_pae of 0.9538 point to a well-defined ligand-binding interface. The gen_complex_iplddt of 0.6137 and gen_interaction_pae of 6.1041 describe the overall complex quality. The gen_affinity_pred_value of 1.0200 and gen_affinity_probability_binary of 0.0886 provide a quantitative affinity estimate, though the low binary probability suggests the model is not confident in a high-affinity classification.

The second parent design, BG_SM_DENOVO_4I6_E2E_boltzgen_binder_007, shows slightly different generator metrics. Its gen_iptm of 0.8630 and gen_plddt of 0.6243 are lower than the first design. The gen_rmsd of 0.9528 Å is higher, indicating more structural deviation upon refolding. The gen_ligand_iptm of 0.8630 and gen_min_interaction_pae of 1.0464 are also less favorable. However, the gen_affinity_pred_value of 1.5186 is higher than the first design's 1.0200, and the gen_affinity_probability_binary of 0.0499 is lower, suggesting the model sees a different affinity profile.

The ADMET assessment for PFNA itself is target-level, not per-candidate. The ADMET aggregate risk score of 0.4067 and toxicity risk score of 0.6761 reflect the ligand's inherent properties. The QED drug-likeness of 0.5343 and synthetic accessibility of 0.8122 describe the ligand, not the protein binders. These values matter for the ultimate application: a binder that recognizes PFNA would be used in a detection or sequestration context where the ligand's own ADMET profile is less relevant than the binder's 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.

The first planned experiment is a yeast surface display screen using IDT oPools oligo pools covering all 107 candidates, tiered by confidence, for gap-repair cloning into the pCTCON2 vector. This screen is designed to test binding of the full candidate pool in a high-throughput format.

The second planned experiment is an IVTT TR-FRET assay using 96-well IDT gBlock constructs for 20 candidates. The constructs use the T7 promoter with a GG + SD + Kozak + ATG + FLAG construct context, following the E. coli MFC codon table. This assay is designed to measure binding affinity in a cell-free format for the top candidates.

Honest Assessment

This campaign against PFNA produced a focused set of candidates with a clear top design, but the results reveal the tension between generator confidence and docking-based ranking. The top-ranked parent design, BG_SM_DENOVO_4I6_E2E_boltzgen_binder_001, achieved a generator confidence score of 0.9289 and a docking score of -9.00 kcal/mol, with strong developability metrics including a solubility score of 0.8176 and a SAP3D spatial score of -1.2544. Its five proofread variants all preserved these values, indicating that the proofread pipeline did not disrupt the parent's favorable properties. The second parent design, BG_SM_DENOVO_4I6_E2E_boltzgen_binder_007, docked slightly worse at -8.76 kcal/mol but earned a higher GNINA CNN pose score of 0.7896 and CNN affinity of 5.8610, suggesting the CNN model saw a more native-like binding mode despite the lower Vina score. The campaign's best docking score of -9.76 kcal/mol was achieved somewhere in the full 107-candidate pool, but that candidate did not appear in the top-ranked table, which was ordered by generator confidence rather than docking score. This mismatch between confidence-based ranking and docking-based ranking is the central tension a follow-up campaign would need to resolve. The pharmacophore analysis identified only four features, all concentrated at the carboxyl head, which constrained the design space: the binder had to recognize a small polar head while accommodating a long fluorinated tail through nonpolar contacts. The target-level QED of 0.534 and ADMET aggregate risk of 0.407 reflect PFNA's inherent properties, and the toxicity risk of 0.676 underscores why a binder against this ligand matters for detection and sequestration applications. A follow-up campaign would benefit from re-ranking the full candidate pool by docking score rather than generator confidence, testing the second parent design's variants more extensively given their higher CNN pose confidence, and exploring whether the proofread pipeline can be tuned to improve the docking scores of the top confidence-ranked designs. The planned YSD screen of all 107 candidates and the IVTT TR-FRET assay of 20 candidates are the right next steps, but the choice of which 20 candidates to test should be informed by both the confidence ranking and the docking ranking, not confidence alone.