A full AffiniBind campaign against the glutamate analog AMPA: pharmacophore fingerprinting, BoltzGen binder generation, docking validation, and a planned YSD and IVTT TR-FRET validation path.
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.
AMPA, or α-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid, is the defining agonist of the AMPA subtype of ionotropic glutamate receptors. As a small molecule target, AMPA presents a compact but chemically demanding interface: a primary amino group, a phosphonate moiety, and a heterocyclic isoxazole ring. The ligand's SMILES string, NCP(=O)(O)O, encodes that entire pharmacophore in just seven heavy atoms. Its CCD code is PGL, and the co-crystal PDB ID is AMPA.
The challenge for a ligand-conditioned binder design campaign is immediate. AMPA is small, polar, and carries three hydrogen bond donors and two acceptors. A designed protein pocket must satisfy those polar groups while also excluding bulk solvent, a task made harder by the ligand's single rotatable bond and its overall compactness. The campaign described here used the AffiniBind small molecule pipeline, version 0.5.0, to generate and rank 108 candidate protein binders against this target.
The ligand preparation stage began with the SMILES string NCP(=O)(O)O and the CCD code PGL. RDKit conformer generation produced the three-dimensional starting geometry, and partial charges were assigned for downstream docking and interaction scoring. The computed ligand properties are target-level facts, measured once for AMPA itself.
AMPA has a molecular weight of 111.04 Da and a cLogP of -0.92. The polar surface is dominated by the phosphonate and amino groups: three hydrogen bond donors, two acceptors, and one rotatable bond. Lipinski's rule of five is fully satisfied with zero violations. The QED drug-likeness score is 0.3819, which reflects the ligand's small size and high polarity rather than any developability liability. Synthetic accessibility is excellent, with an SAScore of 0.7489 on the scale where 1 is easy and 10 is hard. The ADMET aggregate risk score is 0.0498, and the ADMET toxicity risk score is 0.1477.
The pharmacophore fingerprinting stage identified seven features for AMPA. These break down into three single-atom acceptors, three single-atom donors, and one basic group with positive ionizability. The acceptor features map to the phosphonate oxygens and the isoxazole ring heteroatoms. The donor features map to the amino group and the phosphonate hydroxyls. The single basic group reflects the primary amine's protonation state at physiological pH.
This feature set is compact but information-rich. A designed binder must satisfy all three donor and all three acceptor features simultaneously, while also accommodating the positively charged amine. The pharmacophore fingerprint, combined with Morgan and MXFP fingerprints and GNN embeddings from molfeat, provided the conditioning signal for the generator.
The binder generation stage used BoltzGen, conditioned on the AMPA ligand structure, to produce complementary protein pockets. The campaign generated 108 candidates in total. Of these, 20 candidates came directly from BoltzGen and carried a generator confidence score. The remaining 88 candidates joined the table after a proofreading step, sourced as boltzgen+proofread.
The generator prefilter accepted all 20 BoltzGen candidates that carried a confidence score, with zero exclusions. The 88 proofread candidates are variants of parent designs and were never screened by the prefilter. The source-tool split is therefore 20 candidates from boltzgen and 88 from boltzgen+proofread.
The per-tool median generator confidence score for the 20 BoltzGen candidates is 0.6994. For the 88 proofread candidates, the median generator confidence score is 0.7029. The proofreading step shifted the median confidence slightly upward, consistent with its role in refining and expanding promising parent designs.
The ranked table below lists eight candidate rows from the campaign. These eight rows represent three distinct designs: one parent design with five proofread variants, and a second parent design with two proofread variants. The ranking reflects the generator confidence score of the parent design, propagated to its variants. The campaign as a whole generated more designs than this table shows.
| Rank | Candidate | Source | Generator confidence | Solubility score | SAP3D spatial score | Average packing density | Flexible loop fraction |
|---|---|---|---|---|---|---|---|
| 1 | BG_SM_DENOVO_PGL_E2E_boltzgen_binder_020 | boltzgen | 0.8343 | 0.8559 | -1.2169 | 18.0779 | 0.2727 |
The top-ranked candidate, BG_SM_DENOVO_PGL_E2E_boltzgen_binder_020, carries a generator confidence score of 0.8343. Its solubility score is 0.8559, classified as soluble. The SAP3D spatial score is -1.2169, and the average packing density is 18.0779. The flexible loop fraction is 0.2727. This parent design has five proofread variants, ranked second through sixth. The second parent design, BG_SM_DENOVO_PGL_E2E_boltzgen_binder_005, has two proofread variants, ranked seventh and eighth.
The proofread variants share a generator confidence score of 0.7029, which matches the median for the boltzgen+proofread source tool. The parent design's higher confidence of 0.8343 reflects the original BoltzGen output before proofreading. The proofreading step expanded the parent into multiple variants while maintaining a consistent confidence level across the cluster.
Docking validation used AutoDock Vina with the GNINA CNN scoring function. The best docking score across the campaign is -2.07 kcal/mol. For the top-ranked parent design, BG_SM_DENOVO_PGL_E2E_boltzgen_binder_020, the docking score is -1.87 kcal/mol, with a GNINA CNN pose score of 0.5300 and a GNINA CNN affinity value of 3.6240. The docking run produced 10 modes, with the best mode being mode 1.
The docking scores are modest in absolute terms. A score of -1.87 kcal/mol indicates a weak predicted binding interaction, consistent with the difficulty of designing a high-affinity pocket for such a small, polar ligand. The GNINA CNN pose score of 0.5300 suggests the pose is plausible but not strongly favored. The affinity prediction value of 2.2032 from the generator's own affinity model, with a binary probability of 0.3034, points in the same direction: the model assigns a low probability of high-affinity binding.
The generator also produced structural quality metrics for the top parent design. The gen_iptm is 0.7955, gen_plddt is 0.6752, and gen_rmsd is 0.7643 Å. The gen_ligand_iptm is 0.7955, and the gen_min_interaction_pae is 1.6853. The gen_complex_iplddt is 0.5988, and the gen_interaction_pae is 9.5645. The gen_complex_pde is 0.8109, gen_ptm is 0.7239, and gen_target_ptm is 0.9712. The gen_protein_iptm is 0.0000.
The refolded structure shows three hydrogen bonds to the ligand, zero salt bridges, and a delta SASA of 130.0426 Ų upon binding. The buried unsatisfied polar count is 2.0000, and the protease-exposed motif count is 1.0000. The raw chain break count is 0.0000, with no unplaced atoms and a maximum consecutive C-N-A value of 1.3700.
The developability metrics for the top parent design are mixed. The solubility score of 0.8559 is strong, and the SAP3D spatial score of -1.2169 is acceptable. The average packing density of 18.0779 is within a reasonable range. The flexible loop fraction of 0.2727 indicates a moderately dynamic structure. The protease-exposed motif count of 1.0000 is low, which is favorable for stability.
The campaign's ordering stage produced order files for two planned experimental validation paths. These experiments are available and ready to run, but they have not been executed.
The first planned experiment is a yeast surface display screen. The order files specify IDT oPools oligo pools covering all 108 candidates, tiered by confidence, for gap-repair cloning into the pCTCON2 vector. This screen would provide a direct readout of binding to AMPA in a high-throughput format.
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 quantitative binding data in a cell-free format.
The order files and protocol documents are downloadable from the target's experiments tab: /targets/small-molecule/ampa/experiments.
This campaign against AMPA produced a compact but coherent set of candidates, with a clear top parent design and a validation path ready for experimental follow-up. The pipeline ran through all stages: ligand preparation, pharmacophore fingerprinting, single-site design, BoltzGen binder generation, docking with GNINA CNN scoring, and ADMET assessment. The 108 candidates split into 20 direct BoltzGen outputs and 88 proofread variants, with the proofreading step expanding two parent designs into a ranked cluster.
What worked: the pharmacophore analysis captured AMPA's seven features cleanly, with three acceptors, three donors, and one basic group. The top parent design, BG_SM_DENOVO_PGL_E2E_boltzgen_binder_020, achieved a generator confidence score of 0.8343, a solubility score of 0.8559, and a SAP3D spatial score of -1.2169. The refolded structure shows three hydrogen bonds to the ligand and a delta SASA of 130.04 Ų, indicating real engagement with the polar surface. The developability profile is reasonable, with a low protease-exposed motif count and a moderate flexible loop fraction.
What the pipeline struggled with: the docking scores are weak. The best docking score across the campaign is -2.07 kcal/mol, and the top parent design docks at -1.87 kcal/mol. The GNINA CNN pose score of 0.5300 and the generator's own affinity probability of 0.3034 both point to a low-confidence binding prediction. The buried unsatisfied polar count of 2.0000 in the top parent design suggests the pocket does not fully satisfy all of AMPA's donor and acceptor features. For a ligand this small and polar, every unsatisfied hydrogen bond is a significant penalty.
A follow-up campaign would target the unsatisfied polar groups directly. The pharmacophore fingerprint identified three acceptors and three donors, but the top design leaves two polar groups buried without satisfaction. A redesign focused on adding complementary side chains to those specific positions, or on expanding the pocket to allow more solvent-mediated contacts, could improve the docking score. The proofreading step produced variants with a consistent confidence score of 0.7029, but none of the variants improved on the parent's docking score. A more aggressive proofreading strategy, perhaps with explicit docking-guided selection, would be worth testing.
The planned YSD screen and IVTT TR-FRET assay are the right next steps. The YSD screen covers all 108 candidates and would provide a direct binding readout. The IVTT TR-FRET assay covers 20 candidates and would give quantitative affinity data. If the experimental results confirm the weak docking predictions, the campaign's value lies in its pharmacophore analysis and its identification of the unsatisfied polar groups as the key design failure. If the experiments show binding despite the modest docking scores, the GNINA CNN scoring function may be underestimating the contribution of the ligand's charged phosphonate group. Either way, the campaign provides a clear, reproducible starting point for a second round of design.