Saxitoxin 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

Saxitoxin is a potent neurotoxin that blocks voltage-gated sodium channels, and it matters as a target because rapid, specific recognition of this small molecule has direct applications in environmental monitoring, seafood safety screening, and diagnostic tool development. The molecule is a guanidinium-rich alkaloid with a compact, highly polar architecture. Its chemical identity in this campaign is captured by the SMILES string NC(=O)OC[C@@H]1N=C(N)N2CCC(O)(O)[C@@]23NC(N)=N[C@@H]13, and it carries the CCD code 9SL.

The ligand's computed properties explain both its challenge and its appeal as a binder target. Saxitoxin has a molecular weight of 299.29 Da, a cLogP of -3.85, six hydrogen-bond donors, ten hydrogen-bond acceptors, and only two rotatable bonds. That polar, rigid character produces one Lipinski violation and a QED drug-likeness score of 0.2796. The synthetic accessibility score is 0.5244 on the standard SAScore scale where 1 is easy and 10 is hard, which places saxitoxin well within the synthetically tractable range for a natural product derivative. The ADMET aggregate risk score is 0.1919, with a toxicity risk component of 0.4040. These target-level values are the correct reference points for the campaign's drug-likeness and ADMET assessment, and they frame the design problem: the binder must complement a small, highly functionalized, water-soluble ligand rather than a hydrophobic drug-like pocket.

Ligand Preparation

The campaign began with RDKit conformer generation and partial charge assignment for saxitoxin. The ligand's two rotatable bonds keep its conformational ensemble narrow, which is an advantage for single-site design because the binding interface is defined almost entirely by the ligand's rigid guanidinium and carbamate groups. The preparation stage produced the geometry and charge model that fed directly into pharmacophore fingerprinting and the subsequent binder generation step.

The ligand's polarity profile is extreme. With ten hydrogen-bond acceptors and six donors packed into a 299 Da scaffold, saxitoxin presents a dense network of directional interaction opportunities. The two guanidinium groups, captured in the pharmacophore analysis as positively ionizable features, dominate the electrostatic surface. The carbamate oxygen and the hydrated ketone-like center add further acceptor density. This is not a ligand that rewards a simple hydrophobic enclosure strategy; it rewards precise polar complementarity.

Pharmacophore Analysis

The pharmacophore fingerprinting stage identified 14 features across four distinct types. The breakdown is: 4 single-atom acceptors, 7 single-atom donors, 1 three-way-attach hydrophobe, and 2 guanidine positive-ionizable features. The donor count exceeds the acceptor count, which reflects the molecule's guanidinium-rich character. The two guanidine features are the most distinctive elements of the pharmacophore, and they anchor the electrostatic recognition strategy that the binder generator was conditioned to satisfy.

The single hydrophobe feature, a three-way-attach motif, is the only nonpolar handle on the ligand. That scarcity of hydrophobic surface means the designed pocket cannot rely on burial of a large aliphatic patch. Instead, the pharmacophore model pushes the generator toward a pocket rich in carboxylate, carbonyl, and hydroxyl groups positioned to satisfy the guanidinium donors and the acceptor lone pairs. The 14-feature fingerprint is compact but information-dense, and it served as the conditioning signal for the ligand-conditioned design stage.

Binder Generation

The campaign used pipeline version 0.5.0 and generated 120 candidates total. The source-tool split is 20 candidates from boltzgen and 100 from boltzgen+proofread. All 120 candidates carry a generator confidence score, and 20 candidates passed the generator prefilter. The prefilter excluded zero candidates, which means the 100 boltzgen+proofread rows joined the table after the prefilter stage as variant or union rows and were never screened by that prefilter. The 20 boltzgen candidates are the prefilter-passing set.

The generator confidence score distribution tells a clear story about the campaign's internal consistency. The median generator confidence score for the 20 boltzgen candidates is 0.6056, and the median for the 100 boltzgen+proofread candidates is also 0.6056. That identical median reflects the variant structure of the campaign: the proofread rows are refinements of the same parent designs that the boltzgen generator produced, so their confidence scores inherit the parent's value. The boltzgen set also carries median solubility score of 0.8272, median SAP3D spatial score of -1.1282, median average packing density of 15.6308, and median flexible loop fraction of 0.4488.

The design inventory for the eight listed top candidates is important to state plainly. Those eight rows represent four distinct designs, and six of the rows are variants of two parent designs. The ranking by generator confidence score therefore reflects the parent's score propagated to its variants, not eight independent top candidates. The campaign as a whole generated more designs than the table shows, and the source-tool split of 20 boltzgen plus 100 boltzgen+proofread is the correct campaign-level count.

Top Candidates

The ranked table below lists the eight candidate rows from the campaign's top set. The columns shown are those that carry a real measured value on every listed row, so the comparison focuses on generator confidence, docking, and the interaction metrics that all eight rows share.

The table's structure reflects the variant clustering directly. Rows 1 through 6 are all variants of the parent design BG_SM_DENOVO_STX_E2E_boltzgen_binder_007, and they carry identical generator confidence, docking, and interaction metrics because the proofread refinements preserved the parent's scored properties. Rows 7 and 8 belong to the parent design BG_SM_DENOVO_STX_E2E_boltzgen_binder_016, with row 7 as a proofread variant and row 8 as the parent. The ranking therefore represents two distinct parent designs, not eight independent solutions.

The top parent design, binder_007, achieves a generator confidence score of 0.7302 and a docking score of -5.02 kcal/mol. Its gen_iptm of 0.6622 and gen_ligand_iptm of 0.6622 indicate that the model assigns moderate confidence to the interface geometry, while the gen_complex_iplddt of 0.5429 is lower, reflecting the challenge of modeling a small polar ligand in a designed pocket. The gen_affinity_pred_value of 1.7221 and gen_affinity_probability_binary of 0.2277 suggest the generator's internal affinity predictor is cautious about this design.

The second parent, binder_016, scores slightly lower on generator confidence at 0.7177 and docks at -3.09 kcal/mol. However, its gen_plddt of 0.7827 and gen_complex_iplddt of 0.7262 are higher than the top parent's, and its gen_affinity_pred_value of 1.9980 with a binary probability of 0.3873 indicates a more favorable internal affinity read. The trade-off between docking score and generator confidence is the central tension in this campaign's ranking.

Docking and Developability

The docking stage used AutoDock Vina with 10 modes per candidate, and the best mode was mode 1 for every listed candidate. The top docking score across the campaign is -5.02 kcal/mol, achieved by the binder_007 parent and all five of its proofread variants. The binder_016 parent and its variant dock at -3.09 kcal/mol. The GNINA CNN pose score for the top parent is 0.4371, with a CNN affinity of 5.0220, while the second parent shows a pose score of 0.4003 and CNN affinity of 4.8180.

The developability metrics for the two parent designs are available and informative. The binder_007 parent has a solubility score of 0.8547 with a soluble classification, a SAP3D spatial score of -0.8079, an average packing density of 15.4828, zero protease-exposed motifs, zero buried unsatisfied polar atoms, and a flexible loop fraction of 0.4828. The binder_016 parent shows a solubility score of 0.8095 with soluble classification, a SAP3D spatial score of -0.9803, an average packing density of 15.2830, one protease-exposed motif, zero buried unsatisfied polar atoms, and a flexible loop fraction of 0.4151.

The SAP3D spatial scores are negative for both parents, which is the favorable direction for this metric. The binder_016 parent's score of -0.9803 is more negative than the binder_007 parent's -0.8079, suggesting better spatial packing quality in the second design despite its weaker docking score. The flexible loop fractions are high for both, at 0.4828 and 0.4151 respectively, which is a developability concern that a follow-up campaign would want to address through loop stabilization or redesign.

The ligand-level ADMET aggregate risk score of 0.1919 and toxicity risk score of 0.4040 are target properties, not candidate properties. They describe saxitoxin itself, and they set the baseline for any downstream application. The QED of 0.2796 and the single Lipinski violation are also target-level facts. These values do not change across the candidate set because the ligand is fixed; the binder design problem is about complementarity, not about improving the ligand's drug-likeness.

Validation Plan

The campaign's ordering stage produced order files for two planned experimental validation routes, 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 120 candidates, tiered by confidence, for gap-repair cloning into the pCTCON2 vector. This screen is designed to test the full candidate set in a high-throughput format, with the confidence tiering allowing the top-ranked designs to be interrogated first.

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 ready to run and provides a cell-free route to measuring binding signal for the top subset of candidates. The order files and protocol documents are downloadable from the experiments tab, making both validation routes available for another lab to reproduce.

Honest Assessment

This campaign produced a coherent set of ligand-conditioned binder designs for saxitoxin, with the top parent design achieving a docking score of -5.02 kcal/mol and a generator confidence score of 0.7302. The pipeline's strength is its ability to condition on a small, highly polar pharmacophore and generate complementary pockets with reasonable developability profiles. The top parent's solubility score of 0.8547, zero protease-exposed motifs, and zero buried unsatisfied polar atoms are all favorable, and the negative SAP3D spatial score of -0.8079 indicates acceptable packing quality.

What the campaign struggled with is the modest docking scores and the internal tension between docking and generator confidence. The best docking score of -5.02 kcal/mol is workable but not strong for a small polar ligand, and the second parent's better gen_plddt and gen_complex_iplddt values did not translate into a better docking score. The high flexible loop fractions, 0.4828 for the top parent and 0.4151 for the second, are a developability liability that the current ranking does not penalize heavily. The generator's internal affinity predictor was also cautious, with binary probabilities of 0.2277 and 0.3873 for the two parents, suggesting the model itself is uncertain about the interface quality.

A follow-up campaign would target the loop regions for rigidification, either through computational redesign or by adding loop-stability filters to the ranking. It would also benefit from a docking score threshold that is applied before the proofread stage, so that variant expansion focuses on designs with stronger predicted binding. The planned YSD screen and IVTT TR-FRET assay are the right next steps, and their results will show whether the generator's confidence scores or the docking scores are the better predictor of experimental binding for this challenging, guanidinium-rich target.