GenX, also known as HFPO-DA, is a perfluorinated ether carboxylic acid that has drawn sustained attention as a drinking-water contaminant and industrial replacement for legacy perfluorooctanoic acid. The molecule combines a terminal carboxylate with a branched perfluorinated ether tail, giving it a compact but highly fluorinated surface that resists typical hydrophobic collapse. Its SMILES string, O=C(O)[C@@](F)(OC(F)(F)C(F)(F)C(F)(F)F)C(F)(F)F, encodes a stereocenter adjacent to the carboxylate, and the CCD code IGB identifies the ligand in structural databases. For a binder-design campaign, GenX presents a demanding interface: the carboxylate offers a single strong polar anchor, while the perfluorinated tail demands a complementary pocket with unusual shape and electronic character.
The ligand-level properties computed once for GenX place it in an interesting region of chemical space. Its molecular weight is 330.05 Da, with a cLogP of 3.11, one hydrogen-bond donor, two hydrogen-bond acceptors, and four rotatable bonds. It has zero Lipinski violations and a QED drug-likeness score of 0.806. The synthetic accessibility score is 0.7167 on the standard 1-to-10 scale where 1 is easy, and the ADMET aggregate risk score is 0.3631, with a higher toxicity risk component of 0.7045. These values describe the target itself, not the protein binders, and they frame the ligand as a small, moderately lipophilic acid with a meaningful toxicity flag that a successful binder might eventually help detect or sequester.
The campaign began with RDKit conformer generation and partial charge assignment for GenX. The ligand's four rotatable bonds are concentrated in the ether linkage and the carboxylate-adjacent carbon, so the conformer ensemble samples a modest but chemically relevant range of geometries. The stereocenter at the carbon bearing the carboxylate and the fluorinated ether forces a specific three-dimensional arrangement of the acidic group relative to the perfluorinated tail. That arrangement matters for downstream pharmacophore fingerprinting and for the single-site design stage, where the ligand itself defines the binding interface.
The preparation stage produced a ligand ready for both fingerprinting and docking. The partial charges on the carboxylate oxygens and the fluorinated tail create a strongly anisotropic electrostatic surface: negative charge concentrated at the acid, with a low-polarity fluorocarbon region extending away from it. This is the surface a generated protein pocket must complement, and the pipeline's later stages use the prepared conformer set as the geometric and electronic reference for that complementarity.
The pharmacophore fingerprinting stage identified six features for GenX. The breakdown is four single-atom acceptors, one single-atom donor, and one acidic group from the negatively ionizable class. The four acceptor features map to the oxygen atoms in the ether linkage and the carboxylate, while the single donor corresponds to the carboxylate proton. The acidic group feature captures the ionizable carboxylate as a whole, which is the ligand's strongest handle for a designed binder.
This pharmacophore profile is sparse but decisive. A successful binder must satisfy the acidic group with a complementary positive or hydrogen-bonding surface, and it must also engage the acceptor-rich ether region without overcommitting to the fluorinated tail, which offers no classical pharmacophore features. The six features identified here are the chemical anchors the generator uses to condition protein design, and they explain why the top candidates cluster around designs that place polar side chains near the carboxylate while shaping a fluorocarbon-compatible cavity nearby.
The campaign generated 170 candidates using the AffiniBind small-molecule pipeline, version 0.5.0. The source-tool split is 60 candidates from boltzgen and 110 from boltzgen+proofread. The generator prefilter passed 60 candidates and excluded 0, while 110 candidates joined the table after the prefilter stage as variant or union rows. Those 110 rows were never screened by the prefilter, and they correspond to proofread variants of the boltzgen designs.
The generator confidence scores are available for all 170 candidates. The median confidence for the 60 boltzgen candidates is 0.5895, while the median for the 110 boltzgen+proofread candidates is 0.6996. The proofread variants therefore carry a higher typical confidence, which is consistent with their role as refined versions of the initial boltzgen outputs. The top-ranked candidate in the table, BG_SM_DENOVO_IGB_L80_E2E_boltzgen_binder_010, has a generator confidence score of 0.8230, and its proofread variants occupy the next five positions in the ranked list.
The design inventory for the eight listed candidate rows is three distinct designs, with seven rows as variants of two parent designs. The parent BG_SM_DENOVO_IGB_L80_E2E_boltzgen_binder_010 accounts for six of the eight rows: the parent itself and five proofread variants named __pf1_00 through __pf1_04. The second parent, BG_SM_DENOVO_IGB_L80_E2E_boltzgen_binder_059, contributes two proofread variants, __pf1_02 and __pf1_03. The ranking reflects the parent's score propagated to its variants, so the top six rows are one design cluster, not six independent top candidates.
The ranked table below lists the eight candidate rows from the campaign's top-candidate output. The columns are limited to metrics that carry a real value on every listed row, so the table presents the candidate identifier, source tool, ligand gate status, and structure source. The parent row also carries its generator confidence score, docking score, and developability metrics, which are described in the prose that follows the table.
| Candidate | Source tool | Ligand gate status | Structure source |
|---|
| BG_SM_DENOVO_IGB_L80_E2E_boltzgen_binder_010 | boltzgen | not_run | refold |
| BG_SM_DENOVO_IGB_L80_E2E_boltzgen_binder_010__pf1_02 | boltzgen+proofread | not_run | refold |
| BG_SM_DENOVO_IGB_L80_E2E_boltzgen_binder_010__pf1_03 | boltzgen+proofread | not_run | refold |
| BG_SM_DENOVO_IGB_L80_E2E_boltzgen_binder_010__pf1_04 | boltzgen+proofread | not_run | refold |
The parent design BG_SM_DENOVO_IGB_L80_E2E_boltzgen_binder_010 carries a generator confidence score of 0.8230, the highest in the listed set. Its docking score is -5.56 kcal/mol, with a GNINA CNN pose score of 0.3439 and a GNINA CNN affinity value of 3.5820. The docking run produced 10 modes, and the best mode was mode 1. The parent's developability metrics include a solubility score of 0.7339 with a soluble classification, a SAP3D spatial score of -1.2197, an average packing density of 16.2105, zero protease-exposed motifs, zero buried unsatisfied polar atoms, and a flexible loop fraction of 0.3289. The generator also reported a gen_iptm of 0.7737, a gen_plddt of 0.6448, a gen_rmsd of 0.6334, a gen_ligand_iptm of 0.7737, a gen_complex_iplddt of 0.5741, a gen_interaction_pae of 7.9477, a gen_complex_pde of 0.9081, a gen_ptm of 0.8111, a gen_target_ptm of 0.8627, a gen_protein_iptm of 0.0000, two PLIP hydrogen bonds in the refolded structure, zero PLIP salt bridges, a delta SASA of 318.4286 Ų, a predicted affinity value of 2.4028, a binary affinity probability of 0.2003, zero raw chain breaks, a max consecutive C-N-A value of 1.39, and zero unplaced atoms.
The proofread variants of this parent, __pf1_00 through __pf1_04, each carry the source tool label boltzgen+proofread, a ligand gate status of not_run, and a structure source of refold. The second parent's variants, BG_SM_DENOVO_IGB_L80_E2E_boltzgen_binder_059__pf1_02 and __pf1_03, share the same source tool, ligand gate status, and structure source. The ranking of these variant rows reflects the parent's confidence score propagated to its variants, so the top six rows represent one design cluster rather than six independent candidates.
The campaign's best docking score across all candidates is -7.84 kcal/mol. The parent design in the top-candidate table docks at -5.56 kcal/mol, which is a moderate score for a small-molecule binder and suggests that the generated pocket engages GenX with a mix of polar and shape complementarity. The GNINA CNN pose score of 0.3439 and CNN affinity value of 3.5820 for the parent indicate that the pose is plausible but not overwhelmingly favored by the convolutional scoring function. The docking run's 10 modes and best mode of 1 show that the top-ranked pose was also the first mode identified.
The developability metrics for the parent design are encouraging. The solubility score of 0.7339 with a soluble classification suggests the designed protein is not aggregation-prone. The SAP3D spatial score of -1.2197 is negative, which is the favorable direction for spatial packing quality. The average packing density of 16.2105 is within a typical range for well-packed cores. Zero protease-exposed motifs and zero buried unsatisfied polar atoms are both favorable indicators for stability and expression. The flexible loop fraction of 0.3289 indicates that about a third of the design is in flexible loops, which is a reasonable value for a de novo binder.
The generator's structural confidence metrics for the parent are mixed. The gen_plddt of 0.6448 and gen_complex_iplddt of 0.5741 are moderate, while the gen_iptm of 0.7737 and gen_ptm of 0.8111 are stronger. The gen_rmsd of 0.6334 Å suggests the refolded structure is close to the generated model. The gen_interaction_pae of 7.9477 and gen_min_interaction_pae of 1.9885 indicate that the ligand is reasonably well resolved relative to the protein in the complex prediction. The gen_affinity_pred_value of 2.4028 and binary affinity probability of 0.2003 are modest, which is consistent with the docking score and suggests that the parent design is a starting point rather than a finished high-affinity binder.
The per-tool medians provide context for the parent's values. The 60 boltzgen candidates have a median generator confidence of 0.5895, median solubility score of 0.7994, median SAP3D spatial score of -1.0766, median average packing density of 15.6070, and median flexible loop fraction of 0.4205. The parent's confidence of 0.8230 is well above the boltzgen median, and its packing density of 16.2105 is slightly higher than the median, while its flexible loop fraction of 0.3289 is lower than the median. The 110 boltzgen+proofread candidates have a median generator confidence of 0.6996, which is higher than the boltzgen median and reflects the proofreading step's effect on confidence.
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 170 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 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. These experiments are planned and ready to run; they have not been executed, and no Kd values, titration curves, or other experimental results exist for this campaign.
This campaign against GenX (HFPO-DA) produced a workable first-pass binder design with a clear pharmacophore rationale and a defensible developability profile. The ligand's six pharmacophore features, dominated by four acceptors and one acidic group, gave the generator a concrete set of chemical anchors, and the top parent design responded with a pocket that docks GenX at -5.56 kcal/mol while maintaining a soluble, well-packed, and loop-balanced scaffold. The best docking score across the full campaign, -7.84 kcal/mol, shows that the generator can find stronger poses than the top-ranked parent, but the ranking by generator confidence did not surface that pose in the top eight rows. That is a real tension in the pipeline: confidence and docking score are not perfectly aligned, and a follow-up campaign should rank by docking score or by a composite of docking and confidence rather than by confidence alone.
The proofread variants dominate the ranked table, but their individual metric surfaces are thin. The parent carries the docking score, developability metrics, and generator structural confidence values, while the variant rows carry only source tool, ligand gate status, and structure source. That makes it hard to say whether proofreading improved or degraded any specific property. The median confidence for boltzgen+proofread candidates is 0.6996, higher than the boltzgen median of 0.5895, which suggests the proofreading step does raise confidence on average. But the absence of per-variant docking and developability values means the campaign cannot yet claim that proofreading improves binding or stability, only that it changes the confidence distribution.
The ligand itself is a challenging target. Its QED of 0.806 and synthetic accessibility of 0.7167 are favorable, but the ADMET toxicity risk of 0.7045 is a reminder that GenX is a toxicant, not a drug lead. The binder design pipeline is not trying to make GenX drug-like; it is trying to make a protein that recognizes GenX. The target-level ADMET values are therefore context, not a success metric for the binders. The real success metrics are docking score, generator confidence, and the developability panel, and on those the parent design is a credible but not outstanding candidate.
A follow-up campaign would target three things. First, rank by docking score or a docking-confidence composite to surface the -7.84 kcal/mol pose and its design cluster. Second, run the ligand gate stage on the proofread variants so that each variant carries its own docking and developability values, which would let the campaign compare parent and variant directly. Third, expand the design inventory beyond the two parent clusters that dominate the top eight rows, since the full campaign generated 170 candidates and the listed rows represent only three distinct designs. The planned YSD screen and IVTT TR-FRET assay are the right next experiments, and they are ready to run, but they will need the improved ranking and per-variant metrics to make the best use of the 170-candidate pool.
