Paraquat 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

Paraquat is a bipyridinium dication with the SMILES string C[n+]1ccc(-c2cc[n+](C)cc2)cc1 and the CCD code KHJ. It is a small, rigid, doubly charged molecule with a molecular weight of 186.26 Da, a computed cLogP of 1.00, zero hydrogen-bond donors, zero hydrogen-bond acceptors, and a single rotatable bond. The molecule violates none of the Lipinski rules, and its QED drug-likeness score is 0.587. Its synthetic accessibility score is 0.7978 on the standard SAScore scale where 1 is easy and 10 is hard, placing it firmly in the readily synthesizable range.

The ADMET aggregate risk score for paraquat is 0.2533, but the toxicity component is the dominant concern at 0.9147. That split matters for binder design: the ligand itself is not a drug candidate, but a target for recognition. A protein binder that can selectively engage paraquat has potential utility in environmental sensing, sequestration, or diagnostic applications where the goal is capture rather than therapeutic delivery. The design problem is therefore not to improve the ligand's drug-likeness, but to build a complementary protein surface that recognizes the dication's shape, aromatic system, and positive charge distribution.

Ligand Preparation

The campaign began with RDKit conformer generation and partial charge assignment for paraquat. The ligand's rigidity simplifies this stage considerably. With only one rotatable bond and a planar bipyridinium core, the conformer ensemble is compact, and the partial charges on the two quaternary nitrogen atoms define the dominant electrostatic feature that any binder must satisfy.

The prepared ligand then entered the pharmacophore fingerprinting stage. Morgan and MXFP fingerprints captured the aromatic and hydrophobic character of the bipyridinium scaffold, while GNN embeddings from molfeat provided a learned representation of the ligand's chemical environment. These representations condition the downstream generator, giving it a compact description of what the protein pocket must recognize.

Pharmacophore Analysis

The pharmacophore fingerprint identified 10 features across three classes. The aromatic character is captured by 2 Aromatic.Arom6 features, corresponding to the two pyridinium rings. The hydrophobic scaffold contributes 6 Hydrophobe.ThreeWayAttach features, reflecting the branched attachment points around the central bond and the methyl substituents. The two quaternary nitrogen atoms appear as 2 PosIonizable.PosN features.

This feature distribution tells a clear design story. Paraquat is not a large hydrophobic ligand; it is a compact dication with a flat aromatic face and two positive charges separated by roughly the length of the bipyridinium core. A successful binder must present complementary negative charge or polarizable groups to satisfy the PosIonizable.PosN features, while providing an aromatic or hydrophobic pocket to stack against the Aromatic.Arom6 rings. The six Hydrophobe.ThreeWayAttach features suggest that the pocket should also make multiple van der Waals contacts around the ligand periphery.

Binder Generation

The campaign generated 105 candidates using the AffiniBind pipeline version 0.5.0. The source-tool split is 20 candidates from boltzgen and 85 from boltzgen+proofread. All 105 candidates carry a generator confidence score, and 20 candidates passed the generator prefilter. The prefilter excluded zero candidates. The remaining 85 candidates joined the table after the prefilter stage as variant or union rows, meaning they were never screened by the prefilter and represent proofread variants of the primary designs.

The median generator confidence score for the 20 boltzgen candidates is 0.5978, with a median solubility score of 0.8071, a median SAP3D spatial score of -1.2692, a median average packing density of 15.5038, and a median flexible loop fraction of 0.3798. The 85 boltzgen+proofread candidates have a median generator confidence score of 0.6015. The proofread variants cluster tightly around their parent designs, and the ranking reflects the parent's score propagated to its variants.

Top Candidates

The ranked table below lists the eight candidate rows that carry the full set of measured values. The table includes the two parent designs and their proofread variants, with every cell drawn from that candidate's own data line.

The top-ranked parent design, BG_SM_DENOVO_KHJ_E2E_boltzgen_binder_015, carries a generator confidence score of 0.8084 and a docking score of -5.23 kcal/mol. Its solubility score is 0.8173, placing it in the soluble class, and its SAP3D spatial score is -1.0591. The average packing density is 16.3750, and the flexible loop fraction is 0.4750. The design has zero protease-exposed motifs and one buried unsatisfied polar atom. The generator's internal metrics reinforce the ranking: an interface predicted TM-score of 0.7536, a ligand iPTM of 0.7536, a minimum interaction PAE of 2.0481, and a complex ipLDDT of 0.5375. The PLIP analysis of the refolded structure identifies two salt bridges and a delta SASA of 226.7294 Ų, consistent with a pocket that wraps around the dication and buries substantial surface area.

The second parent design, BG_SM_DENOVO_KHJ_E2E_boltzgen_binder_004, has a generator confidence score of 0.7747 and a docking score of -4.53 kcal/mol. Its solubility score is 0.7844, also in the soluble class, with a SAP3D spatial score of -0.9311. The average packing density is 15.6721, and the flexible loop fraction is 0.4426. This design carries three protease-exposed motifs and zero buried unsatisfied polar atoms. Its interface iPTM is 0.7101, with a minimum interaction PAE of 1.9224 and a complex ipLDDT of 0.6219. The refolded structure shows two salt bridges and a delta SASA of 239.0652 Ų. The affinity prediction value is 0.9901, with a binary affinity probability of 0.3620.

The six proofread variants of these two parents are named with the __pf1_XX suffix. Five variants derive from binder_015 and one from binder_004. These variants carry the boltzgen+proofread source label and share the refold structure source. Their generator confidence scores cluster around the parent values, and the campaign's median for the proofread group is 0.6015.

Docking and Developability

Docking validation used AutoDock Vina with 10 modes per candidate. The best docking score across the campaign is -5.23 kcal/mol, achieved by BG_SM_DENOVO_KHJ_E2E_boltzgen_binder_015. The GNINA CNN pose score for this candidate is 0.8481, with a CNN affinity of 4.2470. The second parent design docks at -4.53 kcal/mol, with a GNINA CNN pose score of 0.8073 and a CNN affinity of 4.5090.

The developability metrics for the two parent designs are strong. Both are classified as soluble, with solubility scores of 0.8173 and 0.7844. The SAP3D spatial scores of -1.0591 and -0.9311 indicate acceptable spatial packing quality. The average packing densities of 16.3750 and 15.6721 fall in the expected range for well-folded small binders. The flexible loop fractions of 0.4750 and 0.4426 suggest that roughly half of each design is in loop regions, which is typical for de novo binders of this size.

The ligand-level ADMET assessment applies to paraquat itself, not to the protein candidates. The aggregate ADMET risk score is 0.2533, with a toxicity risk of 0.9147. The QED drug-likeness of 0.587 and the synthetic accessibility of 0.7978 are target-level facts. These values frame the ligand as a tractable small-molecule target for binder recognition, even though its toxicity profile would preclude therapeutic use of the ligand itself.

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 105 candidates. The candidates are tiered by generator confidence score, and the oligos are designed for gap-repair cloning into the pCTCON2 vector. This screen is ready to run and will provide the first experimental read on binding.

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 available for another lab to reproduce and will provide solution-phase binding data.

Both experiments are planned and available, not yet executed. No Kd values, titration curves, or other experimental results exist from this campaign.

Honest Assessment

This campaign produced a focused set of paraquat binder candidates with a clear top design. The best candidate, BG_SM_DENOVO_KHJ_E2E_boltzgen_binder_015, combines a generator confidence score of 0.8084 with a docking score of -5.23 kcal/mol, two salt bridges in the refolded structure, and a buried surface area of 226.73 Ų. The second parent design is close behind, with a docking score of -4.53 kcal/mol and a slightly higher buried surface area of 239.07 Ų. Both designs are soluble, well-packed, and carry the electrostatic features expected for recognizing a dication.

What worked well is the consistency between the generator's internal metrics and the docking validation. The top-ranked design by generator confidence also has the best docking score, and both parent designs show salt bridges in their refolded structures, which is exactly what the pharmacophore analysis predicted for a ligand with two PosIonizable.PosN features. The proofread variants cluster tightly around their parents, suggesting that the proofreading step did not introduce large structural perturbations.

The pipeline's main limitation is the modest docking score range. The best score of -5.23 kcal/mol is reasonable for a small, rigid dication, but it does not suggest extremely tight binding. The GNINA CNN pose scores of 0.8481 and 0.8073 indicate that the docked poses are plausible, but the CNN affinity values of 4.2470 and 4.5090 are in a mid-range that would benefit from experimental validation. The flexible loop fractions near 0.45 to 0.48 mean that roughly half of each design is in loops, which could affect stability or binding kinetics.

A follow-up campaign would target higher affinity by focusing on the electrostatic complementarity. The two quaternary nitrogen atoms in paraquat demand precise placement of negative charges or polarizable groups, and the current designs achieve this with two salt bridges each. Increasing the number of electrostatic contacts, or adding an aromatic stacking platform to engage the bipyridinium face, could push the docking scores below -6 kcal/mol. The planned YSD and TR-FRET experiments will provide the first experimental read on whether the computational ranking holds up in the lab, and those results should guide the next round of design.