Ligand-Conditioned Binder Design Against 2,4-D: A WeaveSeq AffiniBind 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

2,4-Dichlorophenoxyacetic acid, commonly known as 2,4-D, is one of the most widely used synthetic auxin herbicides in agricultural practice. Its structural simplicity belies its biological potency: a chlorinated phenoxy ring linked through an ether oxygen to a carboxylic acid group. The molecule acts as a mimic of the plant hormone indole-3-acetic acid, and its persistence in soil and water has made it a recurring target for biosensor development, environmental monitoring, and bioremediation efforts. Designing a protein binder that specifically recognizes 2,4-D, rather than the broader class of phenoxyacetic acid herbicides, requires a pocket that can discriminate the two chlorine substituents at the 2 and 4 positions of the aromatic ring while also engaging the carboxylate head group.

The ligand presents a compact recognition surface. With a molecular weight of 221.04 Da, a cLogP of 2.46, and only three rotatable bonds, 2,4-D is a relatively rigid small molecule. It carries one hydrogen bond donor and two hydrogen bond acceptors, and it violates none of the Lipinski rules. These properties make it an attractive target for a ligand-conditioned design campaign: the molecule is small enough that a designed protein pocket can wrap around a substantial fraction of its surface, yet it carries enough distinct chemical features, the carboxylic acid, the ether oxygen, the two chlorine atoms, and the aromatic ring, to support specific recognition.

Ligand Preparation

The campaign began with the ligand in its SMILES representation: O=C(O)COc1ccc(Cl)cc1Cl. The CCD code for this ligand is CFA. RDKit conformer generation produced the three-dimensional starting geometry, and partial charges were assigned to prepare the ligand for downstream pharmacophore fingerprinting and docking. The ligand preparation stage is deliberately lightweight for a molecule of this size; 2,4-D has only three rotatable bonds, so the conformational space is modest, and the dominant energetic question is the orientation of the carboxylic acid relative to the phenoxy ring.

The computed ligand properties establish the baseline for the campaign. The QED drug-likeness score is 0.8520, which reflects the molecule's favorable balance of size, polarity, and lipophilicity. The synthetic accessibility score is 0.9344 on the standard SAScore scale where 1 is easy and 10 is hard, confirming that 2,4-D is a readily accessible small molecule. The ADMET aggregate risk score is 0.2780, a low overall risk, though the ADMET toxicity risk score of 0.9630 is elevated, consistent with the known herbicidal activity of this compound class. These target-level properties are computed once for the ligand and apply to the entire campaign; they are not per-candidate metrics.

Pharmacophore Analysis

The pharmacophore fingerprinting stage identified 12 distinct features across six feature classes. The breakdown is as follows:

The pharmacophore profile captures the dual character of 2,4-D. The three single-atom acceptor features map to the carboxylic acid oxygens and the ether oxygen. The single donor feature corresponds to the carboxylic acid proton. The five hydrophobe features with three-way attachment reflect the chlorinated aromatic ring and its substituents, while the lumped hydrophobe feature captures the extended hydrophobic surface of the phenoxy ring. The single acidic group feature identifies the carboxylate as the dominant ionizable element. This distribution of features, three acceptors, one donor, five hydrophobes, one aromatic ring, and one acidic group, gives the design engine a rich set of interaction anchors to condition on.

Binder Generation

The binder generation stage produced 110 candidates in total. The source-tool split is clean: 20 candidates came from the BoltzGen generator alone, and 90 candidates came from the BoltzGen plus proofread pipeline. The generator prefilter accepted all 20 of the BoltzGen candidates that carried a generator confidence score, and the remaining 90 candidates joined the table after the prefilter stage as variant or union rows. These 90 candidates were never screened by the prefilter; they are the proofread variants of the parent designs.

The per-tool metric summaries show the median generator confidence score for the 20 BoltzGen candidates is 0.6460, while the 90 BoltzGen plus proofread candidates have a median generator confidence score of 0.6759. The BoltzGen candidates also carry median developability metrics: a median solubility score of 0.7877, a median SAP3D spatial score of -1.2349, a median average packing density of 15.6092, and a median flexible loop fraction of 0.4006. These medians describe the central tendency of the scored candidates from each source tool.

The design inventory for the top candidate table is important to state plainly. The eight listed candidate rows represent three distinct designs. Seven of the eight rows are variants of two parent designs. The ranking by generator confidence score therefore reflects the parent design's score propagated to its variants, and the variant cluster should not be read as independent top candidates. The campaign as a whole generated more designs than the table shows; the table is a ranked view of the highest-confidence representatives and their proofread variants.

Top Candidates

The ranked table below lists the eight candidates that appear in the top candidate block. The table includes only the metrics that are present on every listed row, which means the table focuses on the candidate identity, source tool, ligand gate status, and structure source. The parent design row carries the full set of generator and docking metrics, and those values belong to that parent design alone.

The top-ranked candidate, BG_SM_DENOVO_CFA_E2E_boltzgen_binder_008, is the parent design from the BoltzGen generator. Its generator confidence score is 0.7884, the highest in the campaign. This parent design carries a full set of generator metrics: a gen_iptm of 0.7720, a gen_plddt of 0.7824, a gen_rmsd of 1.0438 Å, and a gen_ligand_iptm of 0.7720. The interaction metrics are particularly relevant for a ligand-conditioned design: the gen_min_interaction_pae is 1.6153, the gen_interaction_pae is 7.4399, and the gen_complex_iplddt is 0.6477. The gen_plip_hbonds_refolded count is 1.0000, indicating one hydrogen bond between the protein and ligand in the refolded structure, and the gen_delta_sasa_refolded is 313.6990 Ų, a substantial buried surface area upon complex formation.

The developability profile of the parent design is also strong. The solubility score is 0.8151, classified as soluble. The SAP3D spatial score is -1.2708, the average packing density is 15.3220, and the flexible loop fraction is 0.4068. The design carries one buried unsatisfied polar count and zero protease-exposed motifs. The gen_affinity_pred_value is 1.0267, with a binary affinity probability of 0.1827.

The six variants of this parent design, named __pf1_00 through __pf1_04 and __pf1_02, come from the BoltzGen plus proofread pipeline. They share the parent's ligand gate status of not_run and structure source of refold. The two variants of the second parent design, BG_SM_DENOVO_CFA_E2E_boltzgen_binder_001__pf1_02 and __pf1_04, round out the table. These seven variant rows represent the proofread outputs of the two parent designs, and their presence in the top table reflects the parent designs' high generator confidence scores.

Docking and Developability

The docking stage used AutoDock Vina with the GNINA CNN scoring function. The parent design BG_SM_DENOVO_CFA_E2E_boltzgen_binder_008 has a docking score of -4.58 kcal/mol, with a GNINA CNN pose score of 0.6452 and a GNINA CNN affinity value of 3.6080. The docking run produced 10 modes, and the best mode was mode 1. This docking score of -4.58 kcal/mol is the best docking score among the candidates listed in the top table, and it is close to the campaign's overall best docking score of -5.04 kcal/mol reported in the campaign summary.

The docking score of -4.58 kcal/mol for the top-ranked parent design indicates a moderate predicted binding affinity. For a small molecule of 221 Da, a docking score in the -4 to -5 kcal/mol range corresponds to a predicted dissociation constant in the micromolar to high-nanomolar range, depending on the scoring function calibration. The GNINA CNN pose score of 0.6452 suggests the pose is reasonably well-defined, though not at the level of confidence seen in high-affinity protein-ligand complexes. The gen_affinity_pred_value of 1.0267 and the binary affinity probability of 0.1827 from the generator are consistent with a design that is predicted to bind, but with room for affinity maturation.

The developability metrics for the parent design are encouraging. The solubility score of 0.8151 places the design in the soluble class, which is a critical property for downstream experimental validation. The SAP3D spatial score of -1.2708 is negative, which is the favorable direction for this metric. The average packing density of 15.3220 and the flexible loop fraction of 0.4068 describe a protein with a moderate degree of structural order. The single buried unsatisfied polar count is a minor concern, but not a disqualifying one. The absence of protease-exposed motifs is a positive developability signal.

The ligand-level ADMET properties apply to the target, not to the protein binders. The QED drug-likeness of 0.8520, the ADMET aggregate risk score of 0.2780, and the synthetic accessibility of 0.9344 describe 2,4-D itself. These values are relevant for the intended application of the binder: if the goal is to detect or sequester 2,4-D in an environmental or biological context, the ligand's own drug-likeness and ADMET profile matter for the assay design, not for the protein binder's developability.

Validation Plan

The campaign's ordering stage produced order files for two planned experimental validation routes. These experiments are planned and available for reproduction; they have not been run, and no experimental results exist yet.

The first planned experiment is a yeast surface display screen. The order files specify IDT oPools oligo pools covering all 110 candidates, tiered by generator confidence score. The oligos are designed for gap-repair cloning into the pCTCON2 vector, a standard yeast display construct. This screen would provide the first experimental readout of binding for the full candidate set.

The second planned experiment is an IVTT TR-FRET assay. The order files specify 96-well IDT gBlock constructs for 20 candidates, using the T7 promoter with the GG + SD + Kozak + ATG + FLAG construct context and the E. coli MFC codon table. This assay would provide a quantitative binding readout for the top subset of candidates.

The order files and protocol documents are downloadable from the target's experiments tab. These resources are available for another lab to reproduce the planned validation work.

Honest Assessment

This campaign against 2,4-D produced a workable set of candidate protein binders, with the top parent design reaching a generator confidence score of 0.7884 and a docking score of -4.58 kcal/mol. The pipeline executed cleanly through ligand preparation, pharmacophore fingerprinting, binder generation, docking, and developability assessment. The pharmacophore analysis captured the key features of 2,4-D: the acidic carboxylate, the ether oxygen, the chlorinated aromatic ring, and the hydrophobic surface. The binder generation produced 110 candidates, with 20 from BoltzGen and 90 from the proofread pipeline, and the top candidates cluster around two parent designs.

What worked well is the integration of the ligand-conditioned design. The top parent design shows a reasonable predicted binding pose, with one hydrogen bond to the ligand and a buried surface area of 313.70 Ų upon complex formation. The developability profile is solid: soluble, well-packed, and free of protease-exposed motifs. The generator confidence score of 0.7884 is the highest in the campaign, and the docking score of -4.58 kcal/mol is consistent with a micromolar-range binder.

What the pipeline struggled with is the affinity ceiling. The best docking score in the campaign is -5.04 kcal/mol, which is moderate for a small-molecule binder. The binary affinity probability of 0.1827 for the top parent design suggests the generator itself is not fully confident in the binding prediction. The GNINA CNN pose score of 0.6452 is acceptable but not high. A follow-up campaign would target affinity maturation: starting from the top parent design, introducing mutations in the pocket to strengthen the hydrogen bond network with the carboxylate, adding a second hydrogen bond to the ether oxygen, and filling the hydrophobic cleft around the chlorinated ring. The proofread variants of the parent design are available as starting points for this maturation work.

The planned validation experiments are ready to run. The yeast surface display screen would provide the first experimental binding data for all 110 candidates, and the IVTT TR-FRET assay would quantify binding for the top 20. These experiments are the critical next step: the computational metrics are promising but moderate, and experimental validation will determine whether the designed binders achieve the specificity and affinity required for practical use against 2,4-D.