Endosulfan 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

Endosulfan is an organochlorine insecticide whose persistence and toxicity have made it a molecule of sustained environmental and toxicological interest. The ligand presents a compact, heavily chlorinated scaffold: a molecular weight of 406.93 Da, a computed cLogP of 3.69, zero hydrogen-bond donors, three hydrogen-bond acceptors, and zero rotatable bonds. The rigid, hydrophobic character of the molecule is reflected in its pharmacophore decomposition, which identifies three single-atom acceptors, seven three-way-attached hydrophobes, and one lumped hydrophobe of type RH5_5. These features define a binding surface that is dominated by halogenated carbon frameworks and a constrained cyclic sulfite ester, with limited polar functionality available for directional hydrogen bonding.

The ligand-level drug-likeness metrics place endosulfan in an unusual position for a small-molecule target. Its QED score is 0.5754, its synthetic accessibility score is 0.4244 on the standard SAScore scale where 1 is easy and 10 is hard, and it carries zero Lipinski violations. The ADMET aggregate risk score is 0.4893, while the ADMET toxicity risk score is 0.9695, a value that underscores the toxicological burden associated with this scaffold. These target-level properties are computed once for the ligand and frame the design problem: the binder must recognize a compact, lipophilic, and conformationally constrained molecule whose most prominent chemical features are chlorinated carbon centers and a cyclic sulfite oxygen system.

Ligand Preparation

The campaign began with the ligand SMILES string O=[S@]1OC[C@@H]2[C@H](CO1)[C@]1(Cl)C(Cl)=C(Cl)[C@@]2(Cl)C1(Cl)Cl, corresponding to CCD code S68. RDKit conformer generation and partial charge assignment produced the three-dimensional representation used throughout the pipeline. The ligand's zero rotatable bonds mean that the conformational search is constrained to ring puckering and the orientation of the cyclic sulfite ester, rather than torsional sampling across flexible linkers. This rigidity is a practical advantage for single-site design: the ligand presents a relatively fixed pharmacophore arrangement, and the binder generation stage can condition on a well-defined molecular surface rather than an ensemble of rapidly interconverting conformers.

The prepared ligand carries three hydrogen-bond acceptors, all of which appear in the pharmacophore fingerprint as single-atom acceptor features. The remaining eleven pharmacophore features are hydrophobic in character, with seven three-way-attached hydrophobes and one lumped hydrophobe. This distribution tells the design engine that the dominant recognition problem is shape complementarity against a chlorinated hydrocarbon cage, with only a small polar patch available for specific hydrogen-bonding interactions.

Pharmacophore Analysis

The pharmacophore fingerprint for endosulfan resolves into eleven features across three classes. The acceptor class contributes three single-atom acceptors, consistent with the three hydrogen-bond acceptors computed in the ligand properties. The hydrophobic class is split between seven three-way-attached hydrophobes and one lumped hydrophobe of type RH5_5. The three-way-attached hydrophobes correspond to the heavily substituted carbon centers that carry chlorine atoms and ring junctions, while the lumped hydrophobe captures the contiguous nonpolar surface formed by the fused bicyclic core.

This feature distribution has direct consequences for binder design. A pharmacophore with seven hydrophobic attachment points and only three acceptors biases the search toward pockets that bury substantial nonpolar surface area while positioning a small number of polar side chains to satisfy the sulfite oxygens. The generator must avoid over-reliance on hydrogen-bond networks, because the ligand simply does not offer enough acceptor or donor functionality to support an extended polar interface. Instead, the scoring emphasis falls on shape complementarity, van der Waals packing, and the burial of the chlorinated surface.

Binder Generation

The campaign generated 113 candidates using the AffiniBind small-molecule pipeline, version 0.5.0. The source-tool split is clean: 20 candidates came from the boltzgen generator, and 93 came from boltzgen with proofreading applied. All 113 candidates carry a generator confidence score. The generator prefilter passed 20 candidates and excluded zero, while 93 candidates joined the table after the prefilter stage as variant or union rows that were never screened by the prefilter. These 93 post-prefilter rows are the proofread variants of the boltzgen parents.

The eight listed top candidates represent four distinct designs, with six rows being variants of two parent designs. The ranking by generator confidence score places the parent design BG_SM_DENOVO_S68_E2E_boltzgen_binder_018 at the top with a confidence of 0.8764, followed by five of its proofread variants, then one variant of the second parent BG_SM_DENOVO_S68_E2E_boltzgen_binder_008, and finally that second parent itself with a confidence of 0.8700. The variant cluster structure matters for interpretation: the top six rows are refinements of a single parent, and their shared lineage means the ranking reflects that parent's score propagated through its proofread descendants rather than six independent discoveries.

The per-tool medians provide a broader view of the generator output. For the 20 boltzgen candidates, the median generator confidence score is 0.6971, the median solubility score is 0.7331, the median SAP3D spatial score is -1.5374, the median average packing density is 15.4539, and the median flexible loop fraction is 0.2582. For the 93 boltzgen plus proofread candidates, the median generator confidence score is 0.6978. The proofreading step therefore preserves the central tendency of the confidence distribution while expanding the candidate pool nearly fivefold.

Top Candidates

The two parent designs carry the full set of per-candidate metrics, and their values anchor the comparison. The top parent, BG_SM_DENOVO_S68_E2E_boltzgen_binder_018, achieves a generator confidence score of 0.8764, a docking score of -4.75 kcal/mol, a solubility score of 0.5373 with a soluble classification, a SAP3D spatial score of -1.7564, an average packing density of 14.9167, and a flexible loop fraction of 0.1528. The second parent, BG_SM_DENOVO_S68_E2E_boltzgen_binder_008, achieves a generator confidence score of 0.8700, a docking score of -6.73 kcal/mol, a solubility score of 0.7721 with a soluble classification, a SAP3D spatial score of -2.0797, an average packing density of 15.0448, and a flexible loop fraction of 0.3134.

The proofread variants of these parents carry their own lineage identifiers and source-tool labels, but their per-candidate metric lines list only the parent relationship, source tool, ligand gate status, and structure source. The variant rows therefore contribute to the design inventory and the source-tool split without introducing independent values for the developability or docking metrics. The correct reading of the ranked list is that the campaign produced two fully characterized parent designs, each with a cluster of proofread descendants, and the ranking order reflects the parent confidence scores propagated through those clusters.

The table compares the two parent designs across the metrics that both carry. The confidence gap between them is narrow, 0.8764 versus 0.8700, but the docking scores diverge substantially: the second parent binds with a score of -6.73 kcal/mol, a 1.98 kcal/mol improvement over the top-ranked parent's -4.75 kcal/mol. This inversion between generator confidence and docking score is the central tension of the campaign. The generator's internal ranking favors the first parent, while the docking stage identifies the second parent as the stronger binder by a meaningful margin.

The developability metrics reinforce the second parent's advantage. Its solubility score of 0.7721 exceeds the first parent's 0.5373, and its SAP3D spatial score of -2.0797 is more favorable than -1.7564. The average packing density values are close, 15.0448 versus 14.9167, suggesting both designs achieve comparable core packing. The flexible loop fraction differs more noticeably: 0.3134 for the second parent versus 0.1528 for the first, indicating a more loop-rich surface in the second design.

Docking and Developability

The docking stage used AutoDock Vina with ten modes per candidate, and the best mode is reported for each parent. The top-ranked parent by generator confidence docks at -4.75 kcal/mol with a GNINA CNN pose score of 0.8839 and a CNN affinity value of 6.7530. The second parent docks at -6.73 kcal/mol with a GNINA CNN pose score of 0.5172 and a CNN affinity value of 6.5780. The pose score for the first parent is higher, suggesting the CNN model assigns greater confidence to the predicted binding orientation, but the docking score itself favors the second parent by nearly 2 kcal/mol.

The campaign's best docking score across all candidates is -7.63 kcal/mol, which exceeds both parent values shown in the ranked table. This best score belongs to a candidate outside the eight listed rows, and it establishes the upper bound of what the generator produced. The best QED score is 0.575, matching the ligand-level QED of 0.5754 to three significant figures, and the best ADMET risk is 0.489, again consistent with the ligand-level aggregate risk of 0.4893. These target-level values are the correct references for drug-likeness and ADMET assessment, because the ligand properties are computed once for endosulfan and apply to the target rather than to individual protein candidates.

The generator-level metrics for the two parents provide additional structural detail. The first parent carries a gen_iptm of 0.8241, a gen_plddt of 0.6799, a gen_rmsd of 0.5972, and a gen_ligand_iptm of 0.8241. Its gen_complex_iplddt is 0.5492, its gen_interaction_pae is 7.8910, and its gen_complex_pde is 0.5427. The second parent carries a gen_iptm of 0.8212, a gen_plddt of 0.6425, a gen_rmsd of 0.6390, and a gen_ligand_iptm of 0.8212. Its gen_complex_iplddt is 0.5781, its gen_interaction_pae is 6.0555, and its gen_complex_pde is 0.5836. The second parent's lower interaction PAE of 6.0555 versus 7.8910 indicates a more confident interfacial packing prediction, consistent with its stronger docking score.

The PLIP interaction counts from the refolded structures show one hydrogen bond for the second parent and zero for the first, with zero salt bridges for both. The delta SASA upon refolding is 404.0145 Ų for the second parent and 234.1435 Ų for the first, indicating that the second parent buries substantially more surface area upon complex formation. Both parents show zero raw chain breaks, zero unplaced atoms, and no chain-break or unplaced-atom flags, with maximum consecutive C-N-Cα values of 1.38 and 1.36 respectively.

Validation Plan

The campaign's ordering stage produced order files for two planned experimental routes, 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 113 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 20 candidates, with the T7 promoter plus GG plus SD plus Kozak plus ATG plus FLAG construct context under 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.

The YSD screen is the broader of the two planned routes, covering the full 113-candidate pool with confidence-based tiering. The IVTT TR-FRET assay focuses on a 20-candidate subset, matching the number of boltzgen candidates that passed the generator prefilter. The order files and protocol documents are downloadable from the experiments tab, making the validation plan reproducible by another laboratory.

Honest Assessment

This campaign produced a workable set of ligand-conditioned binder candidates against a challenging target. Endosulfan's pharmacophore is dominated by seven hydrophobic three-way-attached features and only three single-atom acceptors, which pushed the generator toward shape-complementary pockets with limited polar interaction networks. The pipeline responded with 113 candidates, a clean source-tool split between 20 boltzgen parents and 93 proofread variants, and two fully characterized parent designs whose docking scores span a meaningful range from -4.75 to -6.73 kcal/mol, with the campaign best reaching -7.63 kcal/mol.

The central tension is the disagreement between generator confidence and docking score. The top-ranked parent by confidence, at 0.8764, docks at only -4.75 kcal/mol, while the second parent, at 0.8700, docks at -6.73 kcal/mol and also shows better solubility, better SAP3D spatial score, lower interaction PAE, and greater buried surface area upon refolding. The generator's internal ranking and the docking stage are pulling in different directions, and the developability metrics side with the docking stage. A follow-up campaign should weight the docking score and the interaction PAE more heavily in the final ranking, because those metrics capture the interfacial quality that the generator confidence score appears to underweight for this hydrophobic target.

The proofread variant cluster is the other structural feature worth addressing. Six of the eight listed rows are variants of a single parent, and the ranking reflects that parent's score propagated through its descendants. The proofreading step expanded the pool from 20 to 113 candidates without shifting the median confidence score, which moved only from 0.6971 to 0.6978. A follow-up campaign would benefit from diversifying the parent pool before proofreading, rather than generating many variants of a small number of parents, because the current design inventory is concentrated in two lineages and the best docking score of -7.63 kcal/mol sits outside the listed top candidates entirely.

The ligand itself remains the hardest constraint. Its QED of 0.5754 and ADMET toxicity risk of 0.9695 mean that any binder designed against endosulfan is engaging a molecule with substantial toxicological liability. The binder design problem is therefore not about improving the ligand's drug-likeness, which is fixed at the target level, but about achieving specific recognition of a compact, rigid, heavily chlorinated scaffold. The planned YSD screen across all 113 candidates and the IVTT TR-FRET assay on the 20-candidate subset are the correct next steps to test whether the docking and developability advantages of the second parent translate into measurable binding in vitro. The order files and protocols are available for another lab to reproduce, and the honest read of the computational data is that the second parent design, not the top-ranked confidence hit, is the candidate most worth advancing to experimental validation.