Designing Protein Binders Against Trifluoroacetic Acid (TFA)

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

Trifluoroacetic acid (TFA) is a small, highly fluorinated carboxylic acid that appears throughout protein chemistry and pharmaceutical manufacturing. It is a persistent counterion in peptide synthesis, a mobile-phase modifier in reversed-phase HPLC, and a frequent contaminant in lyophilized biologics. Its three fluorine atoms create a strongly electron-withdrawing environment that drops the pKa of the carboxylic acid well below that of acetic acid, making TFA fully deprotonated at physiological pH. That negative charge, combined with the compact trifluoromethyl group, gives TFA a distinctive recognition surface: a single acidic group, a single hydrogen-bond donor, and a dense fluorinated cap.

The ligand itself is small, with a molecular weight of 114.02 Da, a cLogP of 0.63, one hydrogen-bond donor, one hydrogen-bond acceptor, and zero rotatable bonds. It violates none of the Lipinski rules. Its QED drug-likeness score is 0.5011, and its synthetic accessibility score is 0.8889 on the standard 1-to-10 scale where 1 is easiest. The ADMET aggregate risk score for TFA is 0.3611, driven in part by a toxicity risk score of 0.7364. These target-level properties frame the challenge: TFA is a small, polar, negatively charged ligand with limited hydrophobic surface, which makes high-affinity recognition by a designed protein pocket a demanding test of shape and charge complementarity.

Ligand Preparation

The campaign began with the ligand SMILES string O=C(O)C(F)(F)F, corresponding to the CCD code TFA. RDKit conformer generation produced the three-dimensional starting geometry used throughout the pipeline, and partial charges were assigned to prepare the ligand for downstream docking and generative modeling. Because TFA has zero rotatable bonds, the conformational search is trivial: the ligand is essentially rigid, and the only meaningful degrees of freedom are the orientation of the carboxylic acid group and the overall pose of the molecule relative to a binding partner. That rigidity simplifies the design problem in one sense, since the pharmacophore features are fixed in space, but it also raises the bar for the protein side: the pocket must match the ligand's exact geometry without the benefit of induced-fit flexibility on the ligand side.

Pharmacophore Analysis

The pharmacophore fingerprinting stage identified four features on TFA. Two are single-atom acceptors, consistent with the two oxygen atoms of the carboxylate group. One is a single-atom donor, corresponding to the protonated or hydrogen-bond-capable oxygen of the carboxylic acid. The fourth feature is an acidic group, reflecting the deprotonated carboxylate that dominates at physiological pH. These four features define the recognition logic for the campaign: a successful binder must satisfy the hydrogen-bonding requirements of the carboxylate while also accommodating the hydrophobic trifluoromethyl group. The pharmacophore is compact and highly directional, with all four features clustered within a few angstroms of one another.

Binder Generation

The campaign used the AffiniBind platform, pipeline version 0.5.0, to generate 177 candidate protein binders conditioned on the TFA ligand structure. The generation stage split across two source tools. The first, boltzgen, produced 80 candidates that passed the generator prefilter. The second, boltzgen+proofread, contributed 97 candidates that joined the table after the prefilter stage as variant or union rows. The prefilter itself excluded zero candidates from the boltzgen set, meaning all 80 boltzgen candidates that reached the prefilter advanced.

The generator confidence scores for the boltzgen candidates had a median of 0.6587 across 80 scored designs. The boltzgen+proofread candidates had a median generator confidence score of 0.7495 across 97 scored designs. The proofread step therefore shifted the confidence distribution upward, which is consistent with its role as a refinement pass over the initial generative output.

The top-ranked candidates in the table represent four distinct designs, with six rows being variants of two parent designs. The ranking reflects the parent design's generator confidence score propagated to its variants. The top parent, BG_SM_DENOVO_TFA_L80B_E2E_boltzgen_binder_001, carries a generator confidence score of 0.8277. Its five proofread variants occupy ranks two through six. The second parent, BG_SM_DENOVO_TFA_L80B_E2E_boltzgen_binder_049, carries a generator confidence score of 0.8257, with one proofread variant at rank eight. The table therefore shows a tight cluster at the top, with the two parent designs separated by only 0.002 in generator confidence.

Top Candidates

The table below compares the two parent designs across the metrics that were measured on both. The variant rows carry only their source tool, parent designation, ligand gate status, and structure source, so the comparison focuses on the parent designs where the full metric surface exists.

The two parent designs are remarkably close in generator confidence, with binder_001 at 0.8277 and binder_049 at 0.8257. Their docking scores are also nearly identical: -2.41 kcal/mol for binder_001 and -2.43 kcal/mol for binder_049. The GNINA CNN pose confidence favors binder_049 slightly at 0.8341 versus 0.7977 for binder_001, and the CNN affinity estimate is also higher for binder_049 at 2.9800 versus 2.7060. Both designs are predicted soluble, with solubility scores of 0.8624 and 0.8739, and both carry zero protease-exposed motifs and zero buried unsatisfied polar atoms.

The structural quality metrics diverge in interesting ways. binder_001 has a higher average packing density at 17.1139 versus 16.0000 for binder_049, and a lower flexible loop fraction at 0.3165 versus 0.4308. Its SAP3D spatial score is more negative at -1.2265 versus -0.8862, which typically indicates better spatial packing quality. However, binder_049 shows a higher pLDDT at 0.7715 versus 0.7120, a lower RMSD at 0.5561 Å versus 0.8061 Å, and a higher complex ipLDDT at 0.6874 versus 0.6085. The refolded PLIP analysis found two hydrogen bonds for binder_049 and zero for binder_001, and the delta SASA upon binding is larger for binder_049 at 208.27 Ų versus 192.68 Ų. These differences suggest that while binder_001 packs more tightly, binder_049 may form more specific polar contacts with the TFA ligand.

Docking and Developability

The docking stage used AutoDock Vina with GNINA CNN rescoring, producing ten modes per candidate. The best docking score across the campaign was -4.17 kcal/mol, which is modest for a small, charged ligand. The two top parent designs scored -2.41 and -2.43 kcal/mol, placing them in the weaker half of the docking distribution. The GNINA CNN pose confidence values of 0.7977 and 0.8341 indicate that the docked poses are reasonably well-defined, but the CNN affinity estimates of 2.7060 and 2.9800 are not strongly predictive of tight binding.

The developability metrics for the two parent designs are encouraging. Both are predicted soluble, with solubility scores above 0.86. Both have zero protease-exposed motifs, zero buried unsatisfied polar atoms, and zero raw chain breaks. The unplaced atom fractions are zero for both designs, and the maximum consecutive C-N distances of 1.39 Å and 1.43 Å are within normal backbone geometry. The generator confidence scores of 0.8277 and 0.8257 are the highest in the campaign, and the median generator confidence across all 80 boltzgen candidates was 0.6587, so these two parents sit well above the typical output.

The target-level ADMET aggregate risk score of 0.3611 and toxicity risk score of 0.7364 are properties of TFA itself, not of the designed binders. The QED drug-likeness of 0.5011 and synthetic accessibility of 0.8889 are also ligand-level values. These numbers frame the target as a small, synthetically accessible molecule with moderate drug-likeness and a notable toxicity risk, which is consistent with its role as a manufacturing contaminant rather than a therapeutic target.

Validation Plan

The campaign's ordering stage produced order files for two planned experimental validation routes. The first is a yeast surface display screen using IDT oPools oligo pools covering all 177 candidates, tiered by confidence, for gap-repair cloning into the pCTCON2 vector. The second is an IVTT TR-FRET assay using 96-well IDT gBlock constructs for 80 candidates, with the T7 promoter plus GG plus SD plus Kozak plus ATG plus FLAG construct context using the E. coli MFC codon table. Both experiments are planned and available for reproduction; they have not been run, and no binding affinities or titration curves exist yet. The order files and protocol documents are downloadable from the target's experiments tab.

Honest Assessment

This campaign set out to design protein binders against trifluoroacetic acid, a small, rigid, negatively charged ligand with a compact pharmacophore of four features: two hydrogen-bond acceptors, one donor, and one acidic group. The pipeline produced 177 candidates, with 80 from the boltzgen generator and 97 from the boltzgen+proofread refinement path. The generator confidence scores reached a maximum of 0.8277, and the median for the boltzgen set was 0.6587, indicating that the generative model produced a wide distribution of designs with a meaningful fraction of high-confidence outputs.

What worked well is the structural quality of the top designs. Both parent candidates are predicted soluble, carry zero protease-exposed motifs, zero buried unsatisfied polar atoms, and zero chain breaks. Their packing densities are reasonable, their SAP3D spatial scores are negative, and their pLDDT values of 0.7120 and 0.7715 suggest folded, ordered structures. The proofread step shifted the median generator confidence upward from 0.6587 to 0.7495, which is a tangible improvement in the quality of the refined set.

What the pipeline struggled with is docking affinity. The best docking score across the entire campaign was -4.17 kcal/mol, and the two top parent designs scored only -2.41 and -2.43 kcal/mol. For a small, charged ligand like TFA, these scores are weak. The GNINA CNN affinity estimates of 2.7060 and 2.9800 are also not indicative of tight binding. The refolded PLIP analysis found zero hydrogen bonds for binder_001 and only two for binder_049, which is a low number for a ligand with four pharmacophore features. The delta SASA values of 192.68 Ų and 208.27 Ų suggest that the ligand is partially buried, but the polar contacts may not be fully satisfied.

A follow-up campaign would target the docking bottleneck directly. The generator is producing structurally sound proteins with high confidence, but the ligand engagement is not translating into strong docking scores. One approach would be to increase the weight on the ligand interaction terms during generation, perhaps by conditioning more strongly on the pharmacophore features or by adding a docking-based filtering step earlier in the pipeline. Another would be to expand the proofread step to include explicit ligand pose refinement, since the current proofread variants carry no docking scores and their effect on ligand engagement is unknown. The two parent designs are close enough in confidence that a larger pool of top candidates, rather than a single winner, should be advanced to the planned yeast surface display and IVTT TR-FRET assays. The experimental data from those assays, when they are run, will reveal whether the generator confidence scores or the docking scores are the better predictor of actual TFA binding.