A ligand-conditioned binder design campaign targeting glyphosate generated 94 candidates, with docking scores reaching -3.04 kcal/mol and a target QED of 0.393. The top-ranked parent design shows a generator confidence score of 0.7893 and a soluble, well-packed refolded structure.
Explore more from this discipline or browse the full archive.
A ligand-conditioned design campaign against perfluorobutanesulfonamide generated 195 binder candidates, with a top docking score of -7.74 kcal/mol and a target QED of 0.806.
Send us a target surface and we will scope a feasibility review, design binders, and report the same objective metrics our research is built on.
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.
A full AffiniBind campaign against perfluorononanoic acid generated 194 binder candidates, with the top design reaching a docking score of -9.50 kcal/mol and a generator confidence of 0.9549.
Glyphosate is a small, highly polar organophosphonate with the SMILES string O=C(O)CNCP(=O)(O)O and the CCD code GPF. It is a herbicide target of persistent interest because its compact size and dense hydrogen-bonding functionality make it a demanding test case for ligand-conditioned binder design. The molecule carries four hydrogen-bond donors and three acceptors across a backbone of only four rotatable bonds, with a molecular weight of 169.07 Da and a cLogP of -1.20. Those properties place glyphosate firmly outside conventional drug-like chemical space, yet its biological relevance as a metabolite and environmental contaminant keeps it on the list of small molecules worth engineering protein binders against.
The challenge for a binder design campaign is immediate from the ligand properties. Glyphosate is small, flexible, and strongly solvated. A designed protein pocket must compete with water for each of its polar groups while also accommodating the ligand's conformational freedom. The AffiniBind pipeline treats the ligand itself as the binding interface, so every downstream design decision starts from the chemical features of this single molecule rather than from a pre-existing protein surface.
The campaign began with RDKit conformer generation and partial charge assignment for glyphosate. The ligand's computed properties are target-level facts, measured once for the molecule and used throughout the pipeline. Glyphosate has a molecular weight of 169.07 Da, a cLogP of -1.20, four hydrogen-bond donors, three hydrogen-bond acceptors, and four rotatable bonds. It records zero Lipinski violations, which is expected for a molecule this small and polar, but the more informative value is its QED drug-likeness score of 0.3934. That score reflects the molecule's distance from typical orally bioavailable small molecules: the low cLogP and high donor count pull the QED down, and the pipeline treats that as a property of the target rather than a liability to be optimized away.
The synthetic accessibility score for glyphosate is 0.7900 on the SAScore scale where 1 is easy and 10 is hard. That places the ligand itself as a readily accessible starting material, which matters for any downstream experimental campaign that will need to source or synthesize the target in quantity. The ADMET aggregate risk score for the ligand is 0.0593, with a toxicity risk score of 0.1419. These are ligand-level values, not candidate-level values, and they describe the small molecule the binders are being designed to recognize.
The pharmacophore fingerprinting stage identified 11 features for glyphosate, distributed across four feature types. The ligand presents five single-atom acceptor features and four single-atom donor features, which is consistent with the phosphonate and carboxylate oxygens and the amine and hydroxyl hydrogens. The feature set also includes one acidic group, assigned to the phosphonate or carboxylate, and one basic group, assigned to the secondary amine.
This feature distribution tells the design engine something important: glyphosate is a hydrogen-bonding-dense ligand with both acidic and basic character in a single small scaffold. A complementary protein pocket needs to satisfy five acceptor features and four donor features simultaneously, while also accommodating the ligand's zwitterionic character. The pharmacophore fingerprint, combined with Morgan and MXFP fingerprints and GNN embeddings from molfeat, becomes the conditioning signal for the binder generation stage.
The campaign generated 94 candidates using pipeline version 0.5.0. The source-tool split is clean: 20 candidates came from boltzgen alone, and 74 candidates came from boltzgen+proofread. All 94 candidates carry a generator confidence score, and 20 candidates passed the generator prefilter. The prefilter excluded zero candidates, which means the 74 boltzgen+proofread rows joined the table after the prefilter stage as variant or union rows and were never screened by that prefilter.
The per-tool medians show how the two source groups compare. The 20 boltzgen candidates have a median generator confidence score of 0.6486, a median solubility score of 0.8206, a median SAP3D spatial score of -1.1076, a median average packing density of 16.1619, and a median flexible loop fraction of 0.4130. The 74 boltzgen+proofread candidates have a median generator confidence score of 0.6439. The proofread variants cluster tightly around their parent designs, and their confidence scores reflect the parent's score propagated through the proofreading step.
The single-site design constraint is central here. Glyphosate is the binding interface, so the pipeline does not perform patch discovery. The generator is conditioned directly on the ligand structure and its pharmacophore fingerprint, and the output is a protein pocket built around the ligand rather than a ligand docked into a pre-existing cavity.
The ranked table below lists the eight candidate rows that appear in the campaign's top-candidate output. These eight rows represent three distinct designs: one parent design with six proofread variants, and a second parent design with two proofread variants. The ranking reflects the parent's generator confidence score propagated to its variants, so the variant cluster around BG_SM_DENOVO_GPF_E2E_boltzgen_binder_002 occupies the top six rows, followed by the two variants of BG_SM_DENOVO_GPF_E2E_boltzgen_binder_015.
| Rank | Candidate | Source Tool | Generator Confidence | Structure Source |
|---|---|---|---|---|
| 1 | BG_SM_DENOVO_GPF_E2E_boltzgen_binder_002 | boltzgen | 0.7893 | refold |
The top-ranked parent design, BG_SM_DENOVO_GPF_E2E_boltzgen_binder_002, carries a generator confidence score of 0.7893. Its six proofread variants all carry a generator confidence score of 0.6439, which matches the median for the boltzgen+proofread source group. The second parent design, BG_SM_DENOVO_GPF_E2E_boltzgen_binder_015, appears through its two proofread variants, both also at 0.6439. The ranking therefore reflects the parent's score propagated to its variants, and the variant cluster should be read as one design family rather than six independent top candidates.
The parent design's own metrics are the richest in the table. Its solubility score is 0.8616, classified as soluble, and its SAP3D spatial score is -0.6681. The average packing density is 15.1852, the protease-exposed motif count is 1.0000, and the buried unsatisfied polar count is 0.0000. The flexible loop fraction is 0.4259. These values describe a refolded structure that is soluble, reasonably packed, and free of buried unsatisfied polar atoms, with a moderate fraction of flexible loop.
The docking stage used AutoDock Vina with 10 modes, and the best mode for the top parent design is mode 1. The docking score for BG_SM_DENOVO_GPF_E2E_boltzgen_binder_002 is -1.38 kcal/mol, with a GNINA CNN pose score of 0.7253 and a GNINA CNN affinity value of 3.6810. The campaign's best docking score across all candidates is -3.04 kcal/mol, which is a meaningful improvement over the top parent's -1.38 kcal/mol and suggests that some lower-ranked designs may form tighter contacts with the ligand.
The generator-level metrics for the top parent add structural context. The gen_iptm is 0.7610, the gen_plddt is 0.7020, and the gen_rmsd is 0.8960 Å. The gen_ligand_iptm is 0.7610, matching the overall iptm, and the gen_min_interaction_pae is 1.1695. The gen_complex_iplddt is 0.5635, the gen_interaction_pae is 6.7661, and the gen_complex_pde is 0.6030. The gen_ptm is 0.8079, and the gen_target_ptm is 0.9407.
The refolded structure shows two hydrogen bonds to the ligand, zero salt bridges, and a delta SASA of 193.949 Ų upon binding. The affinity prediction value is 1.9116, with a binary affinity probability of 0.3582. The raw chain-break count is 0.0000, the max consecutive C-N-CA value is 1.4500, and the unplaced atom count is 0.0000 with a fraction of 0.0000. These values describe a refolded structure that is continuous, fully placed, and engaged with the ligand through two hydrogen bonds.
The developability metrics for the top parent are encouraging. The solubility score of 0.8616 with a soluble classification, the SAP3D spatial score of -0.6681, and the average packing density of 15.1852 all point to a design that should express and fold without major aggregation or packing defects. The protease-exposed motif count of 1.0000 is low, and the buried unsatisfied polar count of 0.0000 means every buried polar atom is satisfied by a hydrogen-bond partner.
The campaign's ordering stage produced order files for two planned experiments, both available for reproduction from the target's experiments tab. These experiments have not been run, and no Kd values, titration curves, or other experimental results exist yet.
The first planned experiment is a yeast surface display screen using IDT oPools oligo pools covering all 94 candidates, tiered by confidence, for gap-repair cloning into the pCTCON2 vector. This screen is designed to rank the full candidate set by surface expression and ligand binding, with the confidence tiering allowing the highest-ranked designs to be tested first.
The second planned experiment is an IVTT TR-FRET assay using 96-well IDT gBlock constructs for 20 candidates. The constructs use the T7 promoter with a GG + SD + Kozak + ATG + FLAG construct context and the E. coli MFC codon table. This assay is designed to measure binding in a cell-free format, providing a direct readout of affinity for the top subset of candidates.
Both experiments are ready to run, and the order files and protocol documents are downloadable from the experiments tab. The tiering by confidence means the top parent design and its variants would be among the first candidates tested in the YSD screen, while the 20-candidate TR-FRET assay would focus on the highest-confidence subset.
This campaign set out to design protein binders against glyphosate, a small, highly polar ligand with dense hydrogen-bonding functionality. The pipeline produced 94 candidates across two source tools, with a clean split between 20 boltzgen designs and 74 boltzgen+proofread variants. The top parent design achieved a generator confidence score of 0.7893, a soluble refold with zero buried unsatisfied polars, and two hydrogen bonds to the ligand. The campaign's best docking score reached -3.04 kcal/mol, and the target ligand's QED of 0.3934 and ADMET aggregate risk of 0.0593 are target-level facts that frame the design problem.
What worked: the pipeline generated a coherent set of designs around a difficult small-molecule target, and the top parent's developability metrics are genuinely encouraging. The solubility score of 0.8616, the SAP3D spatial score of -0.6681, and the average packing density of 15.1852 all suggest a design that should express and fold cleanly. The refolded structure is continuous, fully placed, and engaged with the ligand through two hydrogen bonds, with a delta SASA of 193.949 Ų indicating substantial burial of the ligand upon binding.
What the pipeline struggled with: the docking scores are modest across the board. The top parent's -1.38 kcal/mol is weak for a designed binder, and even the campaign's best score of -3.04 kcal/mol is below what would typically be considered a high-confidence hit. The GNINA CNN pose score of 0.7253 for the top parent suggests the pose is reasonable but not exceptional, and the binary affinity probability of 0.3582 is below the threshold that would flag a strong binder. The generator confidence scores cluster tightly around 0.64 to 0.79, which is workable but not outstanding.
A follow-up campaign would target several specific improvements. First, the docking scores need to be pushed lower, ideally below -6 kcal/mol, by conditioning the generator more strongly on the ligand's pharmacophore features and by increasing the sampling depth around the top parent design. Second, the proofread variants all carry the same generator confidence score of 0.6439, which suggests the proofreading step is not differentiating among variants effectively; a more aggressive proofreading protocol or a different variant-ranking metric could help. Third, the flexible loop fraction of 0.4259 in the top parent is high, and reducing that fraction while maintaining the soluble refold would likely improve binding by reducing entropic penalties. Fourth, the planned YSD and TR-FRET experiments should be run as soon as possible, because the computational metrics alone are not sufficient to rank these designs with confidence. The experimental data would provide the ground truth needed to retrain or fine-tune the generator for this ligand class.
The campaign's honest takeaway is that glyphosate is a hard target, and the pipeline produced a reasonable first-pass set of designs with solid developability but modest predicted affinity. The path forward is clear: run the planned experiments, use the results to refine the generator, and focus the next round on pushing docking scores lower while preserving the solubility and packing quality that the top parent already demonstrates.