A ligand-conditioned binder design campaign targeting 2,4,5-T produced 114 candidates, with the best docking score reaching -7.02 kcal/mol and a top generator confidence of 0.8234.
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.
2,4,5-T (2,4,5-trichlorophenoxyacetic acid) is a chlorinated phenoxy herbicide whose structural signature is immediately recognizable: a trichlorinated aromatic ring linked through an ether oxygen to a carboxylic acid-bearing methylene group. The molecule carries three chlorine substituents at the 2, 4, and 5 positions of the phenyl ring, which gives it a distinctive hydrophobic footprint and a substantial halogenated surface area for a compound of only 255.48 Da. Its CCD code is F52, and the SMILES string O=C(O)COc1cc(Cl)c(Cl)cc1Cl captures the full connectivity: the acid group, the oxyacetate linker, and the heavily chlorinated ring.
The ligand presents a compact but information-rich interface for binder design. With a cLogP of 3.11 and a single hydrogen bond donor against two acceptors, 2,4,5-T sits comfortably within Lipinski space, recording zero violations. Its QED drug-likeness score of 0.8443 places it in favorable territory for a small molecule, and its synthetic accessibility score of 0.9078 on the SAScore scale (where 1 is easy and 10 is hard) indicates a readily accessible compound. The ADMET aggregate risk score of 0.3146 is moderate, though the toxicity risk component of 0.9761 is high, consistent with the known toxicological profile of chlorinated phenoxy acids. These ligand-level properties are target facts, computed once for 2,4,5-T, and they frame the design problem: the binder must recognize a small, partially hydrophobic, partially polar ligand whose chlorines dominate the molecular surface.
The campaign began with RDKit conformer generation and partial charge assignment for 2,4,5-T. The ligand's three rotatable bonds give it limited conformational flexibility, which is an advantage for a single-site design campaign: the oxyacetate linker can rotate, but the aromatic core and its three chlorine atoms remain rigid. This means the pharmacophore features derived from the prepared conformers are likely to be stable across the ensemble, and the binder generation stage receives a relatively constrained ligand geometry to condition on.
The prepared ligand carries one hydrogen bond donor (the carboxylic acid proton) and two hydrogen bond acceptors (the ether oxygen and the carbonyl oxygen). The acid group is ionizable, and the pharmacophore analysis identifies it as an acidic group feature. The three chlorine atoms contribute to the hydrophobic character of the ring, and the aromatic system provides a planar recognition surface. The molecular weight of 255.48 Da and the cLogP of 3.11 place the ligand in a range where both polar and nonpolar contacts can contribute to binding, and the design pipeline must balance these competing interaction types.
The pharmacophore fingerprinting stage identified 12 features for 2,4,5-T, distributed across six feature classes. The most abundant class is the hydrophobic three-way attachment feature, with 5 instances, reflecting the dense chlorination pattern and the resulting nonpolar surface. The aromatic ring contributes a single Arom6 feature, and the ether oxygen and carbonyl oxygen together account for 3 single-atom acceptor features. The carboxylic acid proton appears as a single donor feature, and the acid group itself is captured as one acidic group feature. A lumped hydrophobic feature (RH6_6) rounds out the set.
This feature distribution tells a clear story about the ligand's recognition surface. The five hydrophobic attachment points dominate the pharmacophore, suggesting that a successful binder will need a complementary pocket with substantial nonpolar character to engage the chlorinated ring. The three acceptor features and one donor feature provide the polar anchoring points, and the acidic group offers a potential charge-charge or charge-dipole interaction site. The single aromatic feature indicates that pi-stacking or edge-to-face interactions with the phenyl ring are available but not the primary recognition mode. The pharmacophore fingerprint, combined with the GNN embeddings from molfeat, gives the binder generation stage a rich description of where and how the ligand can be contacted.
The campaign used the AffiniBind pipeline version 0.5.0 in a single-site design mode, where the ligand itself serves as the binding interface and no patch discovery step is required. The generator produced 114 candidates in total, with a source-tool split of 20 candidates from boltzgen alone and 94 candidates from the boltzgen+proofread combination. All 114 candidates carry a generator confidence score, and 20 candidates passed the generator prefilter. The prefilter excluded zero candidates, and the 94 boltzgen+proofread rows joined the table after the prefilter stage as variant or union rows.
The 20 boltzgen candidates represent the primary designs, and the 94 boltzgen+proofread rows are variants of those parents. The per-tool median generator confidence score for the 20 boltzgen candidates is 0.6416, while the 94 boltzgen+proofread variants have a median confidence of 0.6582. The boltzgen candidates also carry median developability metrics: a solubility score of 0.7530, a SAP3D spatial score of -1.4790, an average packing density of 15.1186, and a flexible loop fraction of 0.3155. These medians describe the central tendency of the primary designs before proofreading, and they provide a baseline against which individual candidates can be assessed.
The proofreading step, which produces the boltzgen+proofread variants, is designed to refine the initial designs by correcting local structural issues. The variant rows in the top candidates table carry only their source tool, parent designation, ligand gate status, and structure source, reflecting the fact that the proofread variants inherit their structural context from the parent design while receiving their own identity as distinct rows in the campaign.
The ranked table below lists the eight candidates that appear in the top candidates block, ordered by generator confidence score. The eight rows represent four distinct designs, with six rows being variants of two parent designs. The parent designs are BG_SM_DENOVO_F52_E2E_boltzgen_binder_003 and BG_SM_DENOVO_F52_E2E_boltzgen_binder_006, and their variants carry the __pf1_XX suffix. The ranking reflects the parent's confidence score propagated to its variants, and the variant cluster should be read as a single design family rather than as independent top candidates.
| Candidate | Source Tool | Generator Confidence | Docking Score (kcal/mol) | Solubility Score | SAP3D Spatial Score | Average Packing Density |
|---|---|---|---|---|---|---|
| BG_SM_DENOVO_F52_E2E_boltzgen_binder_003 | boltzgen | 0.8234 | -5.79 | 0.6930 | -1.8919 | 14.5000 |
| BG_SM_DENOVO_F52_E2E_boltzgen_binder_006 | boltzgen | 0.8119 | -7.02 | 0.7693 | -1.8508 |
The table presents the two parent designs that carry the full set of measured metrics. The top-ranked parent, BG_SM_DENOVO_F52_E2E_boltzgen_binder_003, has a generator confidence score of 0.8234, a docking score of -5.79 kcal/mol, a solubility score of 0.6930 with a soluble classification, a SAP3D spatial score of -1.8919, and an average packing density of 14.5000. Its flexible loop fraction is 0.2206, and it carries one protease-exposed motif with zero buried unsatisfied polar atoms. The second parent, BG_SM_DENOVO_F52_E2E_boltzgen_binder_006, has a generator confidence score of 0.8119, the campaign's best docking score of -7.02 kcal/mol, a solubility score of 0.7693 with a soluble classification, a SAP3D spatial score of -1.8508, and an average packing density of 14.5570. Its flexible loop fraction is 0.2532, and it carries zero protease-exposed motifs and zero buried unsatisfied polar atoms.
The six variant rows (__pf1_00 through __pf1_04 for each parent) are proofread versions of these two parents. They carry their own source tool designation (boltzgen+proofread), their parent linkage, a ligand gate status of not_run, and a structure source of refold. The variant cluster for parent 003 includes five variants, and the variant cluster for parent 006 includes one variant in the listed rows. The proofreading step is designed to refine the parent structures, and the variants inherit the parent's design intent while receiving their own structural refinement.
The generator confidence scores for the two parents are close: 0.8234 versus 0.8119. The docking scores, however, diverge more substantially, with parent 006 achieving -7.02 kcal/mol against parent 003's -5.79 kcal/mol. This suggests that the confidence score and the docking score are capturing different aspects of design quality, and the final ranking by confidence does not perfectly align with the docking-based ranking. The solubility scores for both parents are in the soluble range, and the SAP3D spatial scores are both negative, indicating favorable spatial packing. The average packing densities are similar, at 14.5000 and 14.5570 respectively.
The docking stage used AutoDock Vina with the GNINA CNN scoring function, and the results are available for the two parent designs. Parent 006 achieves the campaign's best docking score of -7.02 kcal/mol, with a GNINA CNN pose score of 0.8741 and a GNINA CNN affinity value of 5.3630. The docking run produced 10 modes, with the best mode being mode 1. Parent 003 docks at -5.79 kcal/mol, with a GNINA CNN pose score of 0.9130 and a GNINA CNN affinity value of 5.4710. Its docking run also produced 10 modes, with the best mode being mode 1.
The GNINA CNN pose scores provide a complementary view to the raw docking scores. Parent 003 has a higher pose score (0.9130) than parent 006 (0.8741), suggesting that the CNN model assigns higher confidence to the pose geometry of parent 003 even though its docking score is less favorable. The CNN affinity values are close, at 5.4710 for parent 003 and 5.3630 for parent 006. These values reflect the CNN's internal affinity prediction, which operates on a different scale than the Vina docking score.
The developability metrics for the two parents are strong. Both are classified as soluble, with solubility scores of 0.6930 and 0.7693. The SAP3D spatial scores of -1.8919 and -1.8508 indicate favorable spatial packing, and the average packing densities of 14.5000 and 14.5570 are in a reasonable range. Parent 003 carries one protease-exposed motif, while parent 006 carries zero. Both parents have zero buried unsatisfied polar atoms, which is a positive indicator for structural stability. The flexible loop fractions are 0.2206 for parent 003 and 0.2532 for parent 006, suggesting that both designs have a moderate proportion of flexible regions.
The generator-level structural metrics provide additional detail. Parent 003 has a gen_iptm of 0.7843, a gen_plddt of 0.6561, and a gen_rmsd of 0.7622. Its gen_ligand_iptm is 0.7843, and its gen_complex_iplddt is 0.5922. The gen_interaction_pae is 6.6566, and the gen_min_interaction_pae is 1.4860. The gen_delta_sasa_refolded is 369.9269, indicating substantial burial of surface area upon binding. Parent 006 has a gen_iptm of 0.7813, a gen_plddt of 0.7037, and a gen_rmsd of 0.8674. Its gen_ligand_iptm is 0.7813, and its gen_complex_iplddt is 0.6828. The gen_interaction_pae is 6.7374, and the gen_min_interaction_pae is 1.5658. The gen_delta_sasa_refolded is 403.4364, indicating even greater surface burial than parent 003.
Both parents have zero raw chain breaks, zero unplaced atoms, and a maximum consecutive C-N-A value of 1.43 for parent 003 and 1.41 for parent 006. The gen_affinity_pred_value for parent 003 is 0.3113 with a binary probability of 0.3932, while parent 006 has a gen_affinity_pred_value of -0.2266 with a binary probability of 0.6464. These affinity predictions suggest that parent 006 may have a more favorable predicted binding affinity according to the generator's internal model.
The ADMET assessment for the campaign is reported at the ligand level, since the target ligand's properties are computed once. The QED drug-likeness of 0.8443, the Lipinski violation count of zero, and the synthetic accessibility score of 0.9078 all indicate that 2,4,5-T is a tractable small molecule target. The ADMET aggregate risk score of 0.3146 is moderate, and the toxicity risk score of 0.9761 is high, which is expected for a chlorinated phenoxy acid. These ligand-level properties are the correct values to cite for drug-likeness and ADMET, and they apply to the target rather than to any individual protein candidate.
The campaign's ordering stage produced order files for two planned experimental validation routes, both available for reproduction from the target's experiments tab. The order files and protocol documents are downloadable from that tab.
The first planned experiment is a yeast surface display screen using IDT oPools oligo pools covering all 114 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 highest-ranked designs to be prioritized.
The second planned experiment 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 using the E. coli MFC codon table. This assay is designed to test a focused subset of candidates in a cell-free format, providing a direct readout of binding.
Both experiments are planned and available for another lab to reproduce. They have not been run, and no experimental results exist for this campaign. The order files and protocol documents provide everything needed to execute the validation.
This campaign produced a focused set of protein binder candidates for 2,4,5-T, with two parent designs emerging as the strongest leads. The best docking score of -7.02 kcal/mol from parent 006 is a solid result for a small, heavily chlorinated ligand, and the generator confidence scores of 0.8234 and 0.8119 for the two parents indicate that the design pipeline was able to produce internally consistent structures. The developability metrics are encouraging: both parents are classified as soluble, both have zero buried unsatisfied polar atoms, and the SAP3D spatial scores are negative, indicating favorable packing. The flexible loop fractions of 0.2206 and 0.2532 suggest that neither design is overly rigid, which may be beneficial for induced-fit binding to a small ligand.
The pipeline's strengths are evident in the quality of the parent designs and the consistency of the structural metrics. The proofreading step produced 94 variant rows from the 20 boltzgen parents, and the variant clusters in the top candidates table reflect a deliberate refinement strategy. The pharmacophore analysis correctly identified the hydrophobic dominance of the ligand, with five three-way attachment features, and the binder generation stage appears to have responded to this by producing designs with substantial surface burial (369.93 and 403.44 Ų of delta SASA upon refolding).
The main challenge in interpreting this campaign is the divergence between the generator confidence ranking and the docking-based ranking. Parent 003 ranks first by confidence (0.8234) but docks at -5.79 kcal/mol, while parent 006 ranks second by confidence (0.8119) but achieves the campaign's best docking score of -7.02 kcal/mol. This suggests that the confidence score and the docking score are capturing different aspects of design quality, and a follow-up campaign would benefit from a more explicit integration of these two signals. The GNINA CNN pose scores add another layer: parent 003 has a higher pose score (0.9130) than parent 006 (0.8741), even though parent 006 docks more favorably. The CNN affinity values are close, at 5.4710 and 5.3630, and the generator's internal affinity prediction favors parent 006 with a binary probability of 0.6464 against parent 003's 0.3932.
A follow-up campaign targeting 2,4,5-T would benefit from several adjustments. First, the ranking could be re-weighted to give docking scores and CNN pose scores more influence relative to the generator confidence score, since the docking data provide a more direct readout of binding potential. Second, the proofread variants could be screened through the docking stage to determine whether the proofreading step improves or degrades the docking scores relative to the parents. Third, the high toxicity risk score of 0.9761 for the ligand itself suggests that any binder developed for 2,4,5-T would be most useful in a detection or sequestration context rather than a therapeutic one, and the validation plan should reflect that application focus. The planned yeast surface display screen and IVTT TR-FRET assay are well suited to this goal, and the availability of the order files and protocols means that the experimental validation can proceed as soon as a lab is ready to run it.
| 14.5570 |