GenX (HFPO-DA) Binder Design: 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

GenX, chemically known as HFPO-DA (hexafluoropropylene oxide dimer acid), is a perfluorinated ether carboxylic acid that has drawn sustained attention as a replacement chemistry for legacy perfluoroalkyl substances. The ligand carries a single carboxylic acid group attached to a chiral perfluorinated ether backbone, with the SMILES string O=C(O)[C@@](F)(OC(F)(F)C(F)(F)C(F)(F)F)C(F)(F)F and the CCD code IGB. Its molecular weight sits at 330.05 Da, with a computed cLogP of 3.11, one hydrogen-bond donor, two hydrogen-bond acceptors, and four rotatable bonds. The molecule violates none of the Lipinski rules, and its QED drug-likeness score is 0.806, placing it in a favorable region for small-molecule recognition.

The structural challenge for binder design is immediate from the ligand's composition. The perfluorinated tail is electron-rich in the sense of fluorine's electronegativity but chemically inert and hydrophobic, while the carboxylic acid head offers a single ionizable anchor point. A designed protein pocket must therefore balance a polar contact at the acid group with a substantial nonpolar surface to accommodate the fluorinated chain. The ligand's synthetic accessibility score of 0.7167 on the SAScore scale, where 1 is easy and 10 is hard, indicates that GenX itself is straightforward to obtain for experimental validation, which matters when planning downstream assays.

Ligand Preparation

The campaign began with RDKit conformer generation and partial charge assignment for the GenX ligand. The ligand preparation stage produced the molecular descriptors used throughout the pipeline: molecular weight 330.05 Da, cLogP 3.11, one hydrogen-bond donor, two hydrogen-bond acceptors, four rotatable bonds, and zero Lipinski violations. The QED drug-likeness score computed at this stage was 0.8060, and the ADMET aggregate risk score was 0.3631, with a toxicity-specific risk score of 0.7045.

These ligand-level properties are target facts, not candidate facts. They describe the small molecule the binders must recognize, and they frame the design problem. A QED of 0.8060 means the ligand itself is a reasonable drug-like small molecule, which is relevant if the eventual application involves detecting or sequestering GenX in biological or environmental samples. The ADMET aggregate risk of 0.3631 is moderate, driven in part by the toxicity risk score of 0.7045, a reminder that the target is a perfluorinated compound with known persistence and bioaccumulation concerns.

Pharmacophore Analysis

The pharmacophore fingerprinting stage identified six features on the GenX ligand. The feature-type breakdown is specific: four single-atom acceptors, one single-atom donor, and one acidic group from the NegIonizable class. The four acceptor features map to the fluorine and ether oxygen atoms distributed along the perfluorinated tail and the carboxylic acid oxygens. The single donor corresponds to the carboxylic acid proton, and the acidic group captures the ionizable carboxylate character.

This pharmacophore profile is compact but chemically coherent. The acidic group is the strongest single recognition handle, offering the possibility of a salt bridge or a buried hydrogen-bonding network in a designed pocket. The four acceptor features are individually weak, since fluorine atoms are poor hydrogen-bond acceptors in aqueous environments, but collectively they define the shape and electrostatic surface of the fluorinated chain. A binder that recognizes GenX selectively will likely need to exploit the acid group as the primary anchor while using shape complementarity to discriminate the perfluorinated tail from shorter or non-fluorinated analogs.

Binder Generation

The campaign generated 120 candidates using the AffiniBind pipeline version 0.5.0. The source-tool split is clean: 20 candidates came from boltzgen alone, and 100 candidates came from boltzgen followed by proofreading, designated as boltzgen+proofread. All 120 candidates carried a generator confidence score, and 20 candidates passed the generator prefilter. The prefilter excluded zero candidates, meaning all 20 boltzgen candidates that entered the prefilter stage advanced. The remaining 100 candidates joined the table after the prefilter stage as variant or union rows, never screened by the prefilter.

The generator confidence scores show a meaningful spread. The median confidence score for the 20 boltzgen candidates was 0.5904, and the median for the 100 boltzgen+proofread candidates was also 0.5904. The identical medians reflect the fact that the proofread variants inherit the confidence score of their parent boltzgen design, so the central tendency is driven by the parent distribution. The top-ranked candidates in the table carry confidence scores of 0.8659 and 0.7548, well above the median, indicating that the ranking surface is not flat.

Top Candidates

The ranked table below shows the eight candidate rows that carry the highest generator confidence scores in this campaign. The ranking reflects the parent design's confidence score propagated to its proofread variants, so the top six rows all descend from a single parent design, BG_SM_DENOVO_IGB_E2E_boltzgen_binder_005, with a confidence score of 0.8659. The remaining two rows descend from BG_SM_DENOVO_IGB_E2E_boltzgen_binder_006, with a confidence score of 0.7548.

The table's structure is dominated by the parent-variant relationship. The top six rows all share the same confidence score, docking score, iPTM, pLDDT, and RMSD because they are the same parent design with proofread refinements. The parent BG_SM_DENOVO_IGB_E2E_boltzgen_binder_005 and its five proofread variants all dock at -6.71 kcal/mol, with an iPTM of 0.8268, a pLDDT of 0.6088, and an RMSD of 0.7610 Å. The second design, BG_SM_DENOVO_IGB_E2E_boltzgen_binder_006, and its proofread variant dock at -5.51 kcal/mol, with an iPTM of 0.7918, a pLDDT of 0.5316, and an RMSD of 1.7126 Å.

The confidence gap between the two parent designs is 0.1111, and the docking gap is 1.20 kcal/mol in favor of the higher-confidence design. The iPTM values follow the same ordering: 0.8268 for the top design versus 0.7918 for the second. The pLDDT values are more modest for both, at 0.6088 and 0.5316 respectively, which suggests that while the interface predictions are reasonable, the overall structural confidence in the generated backbones is moderate. The RMSD values of 0.7610 Å and 1.7126 Å indicate that the top design is closer to its predicted geometry than the second design.

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 campaign was -7.43 kcal/mol, which exceeds the -6.71 kcal/mol of the top-listed candidate and the -5.51 kcal/mol of the second design. The docking scores reported for the listed candidates are -6.71 kcal/mol for the binder_005 family and -5.51 kcal/mol for the binder_006 family, both with a GNINA CNN pose score of 0.7978 and 0.5672 respectively, and a GNINA CNN affinity value of 4.5910 and 3.8590 respectively.

The developability metrics for the two parent designs are informative. The binder_005 parent carries a solubility score of 0.7242 with a soluble classification, a SAP3D spatial score of -2.3410, an average packing density of 14.2319, three protease-exposed motif counts, zero buried unsatisfied polar counts, and a flexible loop fraction of 0.1884. The binder_006 parent carries a solubility score of 0.7417 with a soluble classification, a SAP3D spatial score of -1.1406, an average packing density of 15.4177, zero protease-exposed motif counts, zero buried unsatisfied polar counts, and a flexible loop fraction of 0.3418.

The per-tool medians provide context for the parent designs. Across the 20 boltzgen candidates, the median solubility score was 0.7932, the median SAP3D spatial score was -1.0816, the median average packing density was 15.4735, and the median flexible loop fraction was 0.4051. The binder_005 parent sits below the median on flexible loop fraction and above the median on SAP3D spatial score, while binder_006 sits above the median on packing density and below the median on flexible loop fraction. These are single-point comparisons, but they suggest the top design is not an outlier on developability metrics.

The ligand-level ADMET aggregate risk score of 0.3631 and toxicity risk score of 0.7045 are target properties, not candidate properties. They describe the small molecule GenX itself, and they matter for any application where the binder-ligand complex might be used in a biological context. The synthetic accessibility of the ligand, at 0.7167 on the SAScore scale, means the target is easy to source for experimental validation.

Validation Plan

The campaign's ordering stage produced order files for two planned experimental routes. Both are available for reproduction and have not been run.

The first planned experiment is a yeast surface display screen using IDT oPools oligo pools covering all 120 candidates, tiered by confidence, for gap-repair cloning into the pCTCON2 vector. This is the standard route for screening a large candidate pool in a single experiment, and the tiering by confidence means the highest-ranked designs can be enriched or analyzed separately.

The second planned experiment is an IVTT TR-FRET assay using 96-well IDT gBlock constructs for 20 candidates. The construct context includes the T7 promoter, GG, SD, Kozak, ATG, and FLAG tag, using the E. coli MFC codon table. This assay is designed for a smaller, focused set of candidates, consistent with the 20 candidates that passed the generator prefilter.

The order files and protocol documents are downloadable from the target's experiments tab. These planned experiments provide a clear path from the computational campaign to experimental binding data, and the fact that the order files are already generated means another lab can reproduce the screen or the TR-FRET assay directly.

Honest Assessment

This campaign produced a coherent set of ligand-conditioned binder candidates for GenX, with a clear ranking structure and a plausible path to experimental validation. The strongest signal is the consistency between the generator confidence score, the docking score, and the interface metrics for the top parent design. The binder_005 family carries a confidence score of 0.8659, a docking score of -6.71 kcal/mol, an iPTM of 0.8268, and a ligand iPTM of 0.8268, all pointing in the same direction. The second design, binder_006, is weaker on every one of these axes: confidence 0.7548, docking -5.51 kcal/mol, iPTM 0.7918. The ranking is not ambiguous.

What the pipeline struggled with is more subtle. The pLDDT values for both parent designs are moderate, at 0.6088 and 0.5316, which means the generated backbones are not predicted with high structural confidence even when the interface metrics are favorable. The RMSD values of 0.7610 Å and 1.7126 Å suggest that the refolded structures deviate from the generated geometries, with the second design deviating more than twice as far as the first. The flexible loop fraction of 0.3418 for binder_006 is high, and while the top design's 0.1884 is better, both designs carry a degree of conformational flexibility that could complicate experimental characterization.

The campaign's best docking score of -7.43 kcal/mol is better than the top-listed candidate's -6.71 kcal/mol, which means the table's ranking by generator confidence does not perfectly align with docking performance. A follow-up campaign would benefit from re-ranking the full 120-candidate set by docking score rather than by generator confidence alone, and from examining whether the candidates with docking scores near -7.43 kcal/mol carry the same favorable developability metrics as the top confidence-ranked designs. The proofread variants did not improve the docking scores or the interface metrics relative to their parents, which raises the question of what the proofreading stage is contributing beyond structural refinement. A targeted follow-up would focus on the parent designs with the best docking scores, generate proofread variants only for those, and prioritize the yeast surface display screen as the first experimental filter, since it can handle all 120 candidates in a single tiered experiment. The IVTT TR-FRET assay, with its 20-candidate capacity, is well matched to the 20 candidates that passed the generator prefilter, and running it after the YSD screen would provide orthogonal binding data on the most promising subset.