A ligand-conditioned design campaign targeting perfluorooctanoic acid produced 187 candidate binders, with the top-ranked design reaching a docking score of -8.26 kcal/mol and a generator confidence of 0.902.
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.
Perfluorooctanoic acid (PFOA) is a fully fluorinated eight-carbon carboxylic acid with the SMILES string O=C(O)C(F)(F)C(F)(F)C(F)(F)C(F)(F)C(F)(F)C(F)(F)C(F)(F)F and the CCD code 8PF. The molecule presents a deceptively simple chemical surface: a terminal carboxylate head group attached to a rigid, hydrophobic perfluoroalkyl tail. That tail is what makes PFOA a challenging target for protein binder design. The carbon-fluorine bonds create a dense, electron-rich surface with low polarizability, and the molecule resists the kinds of hydrogen-bonding and electrostatic complementarity that typically drive high-affinity protein-ligand recognition.
From a computational design standpoint, PFOA is an attractive test case precisely because it is small, rigid, and chemically distinct from the amino acid side chains that proteins normally use to build binding pockets. The ligand has a molecular weight of 414.06 Da, a cLogP of 4.45, one hydrogen-bond donor, one hydrogen-bond acceptor, and six rotatable bonds. It violates none of the Lipinski rules, and its QED drug-likeness score is 0.653. The synthetic accessibility score is 0.814 on the SAScore scale where 1 is easy and 10 is hard, meaning the ligand itself is straightforward to obtain or synthesize. The ADMET aggregate risk score is 0.395, with a toxicity risk component of 0.698, reflecting the known persistence and bioaccumulation concerns associated with perfluorinated compounds.
The design problem is therefore not about optimizing the ligand, but about building a protein surface that can recognize it. The carboxylate offers a single ionizable anchor point, and the fluorinated tail demands a pocket that can accommodate its shape and electronic character without relying on conventional polar contacts.
The campaign began with RDKit conformer generation and partial charge assignment for PFOA. The ligand's six rotatable bonds are all carbon-carbon bonds along the perfluoroalkyl chain, but the chain's rigidity in practice limits the conformational ensemble. The carboxylate head group can rotate relative to the tail, and the tail itself can adopt slight helical twists, but the molecule does not explore a large conformational space.
This rigidity is an advantage for single-site design. The ligand is the binding interface, and the design task is to build a protein pocket around a relatively fixed molecular shape. The prepared conformer set feeds directly into the pharmacophore fingerprinting stage, where the ligand's chemical features are encoded for conditioning the generative model.
The pharmacophore fingerprinting stage identified four features on PFOA. Two are single-atom acceptors, corresponding to the two oxygen atoms of the carboxylate group. One is a single-atom donor, assigned to the acidic proton of the carboxylic acid. The fourth feature is a negatively ionizable acidic group, which captures the carboxylate's ability to deprotonate at physiological pH and present a formal negative charge.
This feature set is sparse but informative. The entire binding epitope is concentrated at the carboxylate head group. The perfluoroalkyl tail contributes no hydrogen-bond donors or acceptors, no aromatic features, and no additional ionizable groups. A successful binder must therefore anchor recognition at the carboxylate while using shape complementarity and van der Waals contacts to engage the fluorinated tail. The pharmacophore fingerprint tells the generative model exactly where the polar contacts can be made, and it leaves the tail as a hydrophobic surface-matching problem.
The campaign used the AffiniBind pipeline version 0.5.0 to generate 187 candidate protein binders. The generation stage split into two source tools. The first, boltzgen, produced 80 candidates that passed the generator prefilter. The second, boltzgen+proofread, contributed 107 candidates that joined the table after the prefilter stage as variant or union rows. The generator prefilter excluded zero candidates from the boltzgen set, meaning all 80 boltzgen candidates carried a generator confidence score and entered the ranking.
The median generator confidence score for the 80 boltzgen candidates was 0.751. The 107 boltzgen+proofread candidates had a higher median confidence of 0.837. This difference reflects the proofreading step's role in refining and validating the generated structures, but it also means the two source tools occupy different regions of the confidence distribution. The boltzgen set spans a wider range of designs, while the proofread set clusters around higher-confidence refinements.
The boltzgen candidates also carried a richer set of structural and developability metrics. Across the 80 scored boltzgen candidates, the median solubility score was 0.756, the median SAP3D spatial score was -1.455, the median average packing density was 15.116, and the median flexible loop fraction was 0.295. These medians describe the central tendency of the generated binders before proofreading, and they indicate that the boltzgen output tends toward soluble, reasonably packed proteins with a moderate fraction of flexible loop residues.
The ranked table lists eight candidate rows representing three distinct designs. Seven of the eight rows are variants of two parent designs, and the ranking reflects the parent's generator confidence score propagated to its variants. The top-ranked candidate is the parent design BG_SM_DENOVO_8PF_L80_E2E_boltzgen_binder_029, with a generator confidence score of 0.902. The next four rows are proofread variants of that parent, and the final three rows are proofread variants of a second parent, BG_SM_DENOVO_8PF_L80_E2E_boltzgen_binder_031.
The parent design carries the full set of per-candidate metrics. Its generator confidence of 0.902 is the highest in the listed set. The docking score is -8.26 kcal/mol, with a GNINA CNN pose confidence of 0.592 and a GNINA CNN affinity prediction of 5.216. The structure was generated by refolding, and the docking run explored 10 modes, with the best mode ranked first.
The structural quality metrics for the parent are strong. The solubility score is 0.634, placing it in the soluble class. The SAP3D spatial score is -1.890, indicating good spatial packing quality. The average packing density is 14.954, and the flexible loop fraction is 0.185. The parent has zero protease-exposed motif counts and zero buried unsatisfied polar atoms, both favorable for developability.
The generative model's own confidence metrics reinforce the ranking. The interface predicted template modeling score (gen_iptm) is 0.849, and the ligand iptm is also 0.849. The predicted local distance difference test (gen_plddt) is 0.701, and the complex iplddt is 0.641. The interaction predicted aligned error is 6.201, and the minimum interaction PAE is 1.049. The complex predicted distance error is 0.598. The affinity prediction value is 1.838, with a binary affinity probability of 0.266. The refolded structure has zero raw chain breaks, a maximum consecutive C-N distance of 1.40 Å, and zero unplaced atoms.
The seven variant rows carry their own source tool designation (boltzgen+proofread), their own ligand gate status (not_run), and their own structure source (refold). The variants are refinements of the parent designs, and their ranking in the table reflects the parent's confidence score, but their own measured values are limited to the source tool, ligand gate status, and structure source fields.
| Candidate | Source Tool | Generator Confidence | Docking Score (kcal/mol) | Solubility Score | SAP3D Spatial Score |
|---|---|---|---|---|---|
| BG_SM_DENOVO_8PF_L80_E2E_boltzgen_binder_029 | boltzgen | 0.902 | -8.26 | 0.634 | -1.890 |
The table above includes only the parent design, because it is the only listed row that carries a full set of measured values across the confidence, docking, solubility, and SAP3D columns. The seven variant rows are described in prose, where their own measured values (source tool, ligand gate status, structure source) can be stated without borrowing metrics from the parent.
The parent design's docking score of -8.26 kcal/mol is the best docking score in the campaign, and it is close to the campaign's overall best of -8.91 kcal/mol. The GNINA CNN pose confidence of 0.592 indicates that the docking pose is moderately well supported by the convolutional neural network scoring. The affinity prediction of 5.216 is a GNINA internal value, not a direct binding affinity in molar units, and it should be interpreted as a relative ranking signal rather than an absolute affinity estimate.
The docking stage used AutoDock Vina with GNINA CNN rescoring. The parent design's docking score of -8.26 kcal/mol places it in a favorable range for a designed binder, though it is not the strongest docking score in the full campaign. The campaign's best docking score was -8.91 kcal/mol, achieved by a candidate outside the top-ranked confidence list. This divergence between confidence ranking and docking ranking is expected: the generator confidence score reflects the model's belief in the design's structural plausibility, while the docking score reflects the predicted binding energy in a rigid or semi-rigid docking protocol.
The developability metrics for the parent design are encouraging. The solubility score of 0.634 places it in the soluble class, and the SAP3D spatial score of -1.890 indicates that the protein's spatial packing is well within the range typically associated with stable, well-folded proteins. The average packing density of 14.954 is consistent with a compact core, and the flexible loop fraction of 0.185 suggests that only a small portion of the protein is in disordered or flexible regions. The zero buried unsatisfied polar count means that all polar atoms in the core are satisfied by hydrogen bonds or salt bridges, which is a strong indicator of folding stability.
The ADMET assessment for this campaign is a target-level property, not a per-candidate property. PFOA itself has an ADMET aggregate risk score of 0.395 and a toxicity risk score of 0.698. These values describe the ligand, not the protein binders. The protein binders are not assessed for ADMET properties in this pipeline stage, because the ADMET-AI tool is applied to the small molecule target. The ligand's QED of 0.653 and its synthetic accessibility of 0.814 are also target-level facts.
The ligand's drug-likeness profile is relevant to the campaign's framing. PFOA is not a drug candidate, and its QED of 0.653 reflects a molecule that is moderately drug-like by the standard filters but carries significant toxicity risk. The binder design campaign is therefore not aimed at improving PFOA's drug-likeness, but at creating a protein that can recognize and bind PFOA for detection, sequestration, or catalytic applications.
The campaign's ordering stage produced order files for two planned experimental validation routes. These experiments are available for reproduction and have not been run.
The first planned experiment is a yeast surface display screen using IDT oPools oligo pools covering all 187 candidates. The candidates are tiered by confidence, and the screen is designed for gap-repair cloning into the pCTCON2 vector. This approach allows high-throughput display of the designed binders on the yeast surface, where binding to PFOA can be assessed by flow cytometry or related methods.
The second planned experiment is an IVTT TR-FRET assay using 96-well IDT gBlock constructs covering 80 candidates. The constructs use the T7 promoter with a GG + SD + Kozak + ATG + FLAG context, and the codon table is the E. coli MFC table. This cell-free expression and assay format is designed to test binding in a controlled in vitro setting, with the TR-FRET readout providing a quantitative signal for ligand engagement.
The order files and protocol documents are downloadable from the target's experiments tab. These resources allow another lab to reproduce the planned experiments exactly as designed.
This campaign produced a credible set of ligand-conditioned binder designs for PFOA, with a clear top candidate and a well-defined validation path. The parent design BG_SM_DENOVO_8PF_L80_E2E_boltzgen_binder_029 stands out with a generator confidence of 0.902, a docking score of -8.26 kcal/mol, and a strong developability profile including a solubility score of 0.634, a SAP3D spatial score of -1.890, and zero buried unsatisfied polar atoms. The structural quality metrics from the generative model are consistent: the interface iptm of 0.849 and the complex iplddt of 0.641 indicate that the model is confident in the designed interface, and the refolded structure has no chain breaks or unplaced atoms.
The campaign's main limitation is the sparse metric coverage on the proofread variants. The seven variant rows carry only their source tool, ligand gate status, and structure source. This means the ranking of the top eight rows is driven entirely by the parent designs' confidence scores, and the variants themselves cannot be compared on docking, solubility, or structural quality within the listed set. A follow-up campaign would benefit from running the full docking and developability pipeline on the proofread variants, so that the refinement step's impact can be quantified rather than inferred.
The divergence between the confidence ranking and the docking ranking also deserves attention. The campaign's best docking score of -8.91 kcal/mol belongs to a candidate outside the top confidence list, which suggests that the generator confidence score and the docking score are capturing different aspects of design quality. A follow-up campaign could re-rank candidates by a composite score that weights docking, confidence, and developability metrics together, rather than relying on confidence alone for the primary ranking.
The pharmacophore analysis correctly identified the carboxylate head group as the sole polar anchor, and the generative model appears to have responded by building pockets that complement the fluorinated tail through shape and van der Waals contacts. The planned yeast surface display and IVTT TR-FRET experiments will test whether these computational predictions translate into measurable binding. The order files are ready, the protocols are documented, and the campaign's output is structured for direct experimental follow-up.