PFNA Binder Design: A Ligand-Conditioned Campaign Against Perfluorononanoic Acid

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, capped by a single carboxylic acid group, creates a chemically unusual target for protein binder design. The ligand presents a long, rigid, hydrophobic fluorocarbon chain with a single ionizable head group, a combination that demands a binding pocket capable of simultaneously satisfying the acidic carboxylate while accommodating the lipophilic fluorinated tail. PFNA appears in the campaign with the CCD code 4I6, and its structure is defined by the 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.

The ligand's computed properties frame the design problem. PFNA carries a molecular weight of 464.07 Da and a cLogP of 5.08, reflecting the extreme hydrophobicity of the perfluorinated chain. It has one hydrogen bond donor and one hydrogen bond acceptor, both contributed by the carboxylic acid group. The molecule has seven rotatable bonds, though the fluorinated backbone itself is conformationally constrained relative to a hydrocarbon chain of similar length. PFNA registers one Lipinski violation, driven by its size and lipophilicity, and its QED drug-likeness score is 0.5343. The synthetic accessibility score is 0.8122 on the standard SAScore scale where 1 is easy and 10 is hard, indicating that the ligand itself is straightforward to obtain. The ADMET aggregate risk score is 0.4067, with a toxicity risk score of 0.6761, a reminder that PFNA is a compound of toxicological concern rather than a therapeutic lead.

Ligand Preparation

The ligand preparation stage began with the PFNA SMILES string and the CCD code 4I6. RDKit conformer generation produced the three-dimensional coordinates used throughout the campaign, and partial charges were assigned to prepare the ligand for downstream docking and generative modeling. The single carboxylic acid group received its ionizable charge treatment, while the perfluorinated tail was parameterized as the neutral, highly electronegative fluorocarbon it is.

The prepared ligand then fed directly into the pharmacophore fingerprinting stage. The campaign used Morgan fingerprints and MXFP pharmacophore fingerprints, along with GNN embeddings computed through molfeat, to capture the ligand's chemical features in a form the generative models could condition on. This fingerprinting step is where the ligand's unusual chemistry gets translated into a numerical representation: the dense fluorination pattern, the single acidic group, and the overall rod-like shape all contribute to the embedding that guides binder generation.

Pharmacophore Analysis

The pharmacophore fingerprinting stage identified four features on PFNA. Two features are classified as Acceptor.SingleAtomAcceptor, one as Donor.SingleAtomDonor, and one as NegIonizable.AcidicGroup. The acidic group feature corresponds to the carboxylic acid head, which at physiological pH exists predominantly as the carboxylate anion. The single donor and the two acceptor features reflect the oxygen atoms of the carboxylic acid group: the hydroxyl oxygen can act as a hydrogen bond donor, while both the carbonyl oxygen and the ionized carboxylate oxygens can serve as hydrogen bond acceptors.

This pharmacophore profile is sparse but chemically decisive. The entire polar interaction surface of PFNA is concentrated at the carboxylic acid head. The fluorinated tail offers no hydrogen bonding capacity and no ionizable groups, only a continuous hydrophobic surface. A successful binder must therefore anchor the ligand through specific polar contacts at the carboxylate while providing a complementary hydrophobic groove for the fluorocarbon chain. The four pharmacophore features are the only polar handles the design pipeline had to work with.

Binder Generation

The campaign generated 175 candidates using the AffiniBind platform, pipeline version 0.5.0. The single-site design strategy treated the ligand itself as the binding interface, with no patch discovery step. The generative models were conditioned directly on the prepared PFNA structure and its pharmacophore fingerprints, producing protein backbones intended to form a complementary pocket around the ligand.

The source-tool split shows two distinct generation paths. Sixty candidates came from the BoltzGen generator alone, while 115 candidates came from a BoltzGen plus proofread workflow. The proofread step applies an additional refinement pass to the generated structures, and the median generator confidence score for proofread candidates is 0.8755, compared to 0.8092 for the BoltzGen-only set. All 175 candidates carry a generator confidence score. The generator prefilter accepted 60 candidates and excluded zero, while the 115 proofread variants joined the table after the prefilter stage as variant or union rows.

The per-tool medians for the BoltzGen-only set provide a baseline for the generated ensemble. Across 60 scored candidates, the median solubility score is 0.7507, the median SAP3D spatial score is -1.2950, the median average packing density is 15.2698, and the median flexible loop fraction is 0.3009. These values describe a set of designs that are generally soluble, reasonably well packed, and carry a moderate fraction of flexible loop residues.

Top Candidates

The ranked candidate table lists eight rows representing four distinct designs. Six of the rows are variants of two parent designs, produced by the proofread workflow. The ranking reflects the parent design's generator confidence score propagated to its variants, so the variant cluster around a high-confidence parent appears as a block of top-ranked rows.

The top-ranked parent design, BG_SM_DENOVO_4I6_L80B_E2E_boltzgen_binder_041, carries a generator confidence score of 0.9214 and a docking score of -6.88 kcal/mol. Its solubility score is 0.7024, placing it in the soluble class, and its SAP3D spatial score is -2.4681. The average packing density is 14.8148, and the flexible loop fraction is 0.2778. This design has one protease-exposed motif and zero buried unsatisfied polar atoms. The generator's internal metrics for this design include an iPTM of 0.8766, a pLDDT of 0.6676, and a ligand iPTM of 0.8766. The minimum interaction PAE is 0.8428, and the interaction PAE is 6.2621. The complex ipLDDT is 0.6055, and the complex PDE is 0.4896. The refolded structure shows one hydrogen bond to the ligand and zero salt bridges, with a delta SASA of 334.6653 Ų upon binding. The affinity prediction value is 1.6690, with a binary affinity probability of 0.0498. The docking run produced ten modes, with the best mode at -6.88 kcal/mol and a GNINA CNN pose score of 0.2938.

The second parent design, BG_SM_DENOVO_4I6_L80B_E2E_boltzgen_binder_021, has a generator confidence score of 0.9156 and a docking score of -6.61 kcal/mol. Its solubility score is 0.7438, also in the soluble class, and its SAP3D spatial score is -1.3565. The average packing density is 16.2895, and the flexible loop fraction is 0.2368. This design also has one protease-exposed motif and zero buried unsatisfied polar atoms. The generator metrics include an iPTM of 0.8716, a pLDDT of 0.6447, and a ligand iPTM of 0.8716. The minimum interaction PAE is 1.0359, and the interaction PAE is 7.2222. The complex ipLDDT is 0.5674, and the complex PDE is 0.5369. The refolded structure shows one hydrogen bond to the ligand and zero salt bridges, with a delta SASA of 325.0665 Ų. The affinity prediction value is 1.3467, with a binary affinity probability of 0.1077. The docking run produced ten modes, with the best mode at -6.61 kcal/mol and a GNINA CNN pose score of 0.2930.

The six variant rows all come from the proofread workflow and carry the boltzgen+proofread source tool label. Each variant is named with its parent design as a prefix, followed by a proofread identifier such as __pf1_02 or __pf1_00. The variants of binder_041 are __pf1_02 and __pf1_00, while the variants of binder_021 are __pf1_04, __pf1_03, __pf1_00, and __pf1_01. These variants share the refold structure source and the ligand gate status of not_run with their parents.

Docking and Developability

Docking validation was performed with AutoDock Vina, and the two parent designs that carry docking scores both show favorable binding energies. The top-ranked parent, binder_041, docks at -6.88 kcal/mol, while binder_021 docks at -6.61 kcal/mol. Both values are in the range expected for a small molecule of this size engaging a designed protein pocket, though they are not exceptionally tight. The GNINA CNN pose scores are 0.2938 and 0.2930 respectively, and the CNN affinity values are 4.1380 and 4.3770. These CNN-based scores provide an orthogonal assessment of pose quality and predicted affinity, and they are consistent with the Vina docking energies in ranking the two parents in the same order.

The developability metrics for the two parent designs are encouraging. Both are classified as soluble, with solubility scores of 0.7024 and 0.7438. The SAP3D spatial scores are -2.4681 and -1.3565, both in the negative range that indicates favorable spatial packing. The average packing densities are 14.8148 and 16.2895, and the flexible loop fractions are 0.2778 and 0.2368. Both designs have exactly one protease-exposed motif and zero buried unsatisfied polar atoms, suggesting that the polar network in each pocket is fully satisfied.

The generator's internal quality metrics add another layer of confidence. The iPTM values of 0.8766 and 0.8716 indicate strong predicted interface quality. The pLDDT values of 0.6676 and 0.6447 are moderate, reflecting the difficulty of designing a pocket around a highly fluorinated ligand. The ligand iPTM values match the overall iPTM values, confirming that the interface quality is driven by the ligand contacts. The minimum interaction PAE values of 0.8428 and 1.0359 are low, indicating high confidence in the relative position of the ligand and the pocket residues. The refolded structures each show one hydrogen bond to the ligand and zero salt bridges, consistent with the pharmacophore analysis that identified the carboxylic acid head as the sole polar interaction site.

The ligand-level properties frame the drug-likeness and ADMET context for the entire campaign. PFNA itself has a QED of 0.5343, which is moderate and reflects the molecule's size and lipophilicity. The ADMET aggregate risk score is 0.4067, with a toxicity risk score of 0.6761. These values are target-level facts, computed once for the ligand, and they apply to every binder design in the campaign. The synthetic accessibility score of 0.8122 indicates that PFNA is easy to synthesize, which is useful for experimental validation.

Validation Plan

The campaign's ordering stage produced order files for two planned experimental workflows, both available for reproduction from the target's experiments tab. The order files and protocol documents are downloadable there.

The first planned experiment is a yeast surface display screen. The order files specify IDT oPools oligo pools covering all 175 candidates, tiered by confidence, for gap-repair cloning into the pCTCON2 vector. This screen is designed to test the full candidate set in a high-throughput format, with the confidence tiering allowing the highest-ranked designs to be prioritized for downstream characterization.

The second planned experiment is an IVTT TR-FRET assay. The order files specify 96-well IDT gBlock constructs covering 60 candidates, with the T7 promoter plus GG plus SD plus Kozak plus ATG plus FLAG construct context using the E. coli MFC codon table. This assay is designed to test a focused subset of the candidates in a cell-free format that can provide quantitative binding data.

Honest Assessment

This campaign against PFNA produced a workable set of binder candidates, but the results reveal the tension between generative confidence and docking validation. The top-ranked parent design, binder_041, carries a generator confidence score of 0.9214 and docks at -6.88 kcal/mol. The second parent, binder_021, has a confidence of 0.9156 and docks at -6.61 kcal/mol. The docking scores are respectable but not exceptional for a ligand of this size, and the CNN affinity values of 4.1380 and 4.3770 suggest that the predicted binding is moderate rather than tight. The generator's internal metrics tell a more optimistic story: the iPTM values above 0.87, the low minimum interaction PAE values, and the fully satisfied polar networks all indicate that the models believe these pockets are well formed. The single hydrogen bond to the ligand in each refolded structure is consistent with the pharmacophore analysis, which identified the carboxylic acid head as the only polar handle on PFNA. The challenge is that one hydrogen bond and a hydrophobic groove may not be enough to achieve high-affinity, high-specificity binding against a ligand whose fluorinated tail is chemically inert.

The proofread workflow clearly adds value at the level of generator confidence, with a median of 0.8755 compared to 0.8092 for the BoltzGen-only set. But the proofread variants in the top table carry only their source tool and structure source labels, which limits the direct comparison of their developability metrics against the parent designs. A follow-up campaign would benefit from running the full developability and docking pipeline on the proofread variants, so that the refinement step's impact on solubility, packing, and docking can be quantified rather than inferred from confidence scores alone.

The ligand itself is the campaign's central constraint. PFNA's QED of 0.5343 and ADMET toxicity risk of 0.6761 are reminders that this is a recognition target, not a therapeutic lead. The design goal is a binder that can selectively recognize PFNA in complex mixtures, and the current candidates represent a reasonable first pass at that goal. A follow-up campaign would target higher docking scores, ideally below -8 kcal/mol, and would seek to increase the number of polar contacts to the ligand beyond the single hydrogen bond observed in the refolded structures. The planned yeast surface display screen and IVTT TR-FRET assay are the right next steps, and the order files are available for any lab ready to test these designs experimentally.