Benchmarking protein-ligand co-folding across Nav, Cav and Kv-family channels

All-atom protein-ligand co-folding offers a route to modelling biomolecular complexes without prespecifying the receptor conformation or binding site, but its reliability for large, state-dependent membrane proteins remains uncertain. Here, we established a training-cut-off-aware benchmark of experimentally resolved ligand-bound complexes of voltage-gated sodium (Nav), calcium (Cav) and potassium (Kv) channels to evaluate AlphaFold 3, Boltz-2 and Protenix-v2 across multiple ligand representations. The final production set comprised 301 completed method-input jobs and 7,525 predicted structures generated from RCSB/CACTVS SMILES, Protein Data Bank Chemical Component Dictionary identifiers and matched PubChem and ChEMBL representations. Executable coverage differed markedly among methods, particularly for Kv-family systems, necessitating restriction of the balanced pose-accuracy comparison to 22 Nav and Cav complexes completed by all three methods. Median top-ranked ligand heavy-atom root-mean-square deviations were 19.85 Å for AlphaFold 3, 22.12 Å for Boltz-2 and 20.25 Å for Protenix-v2. Best-of-25 selection improved pose accuracy, but recovered a ligand pose within 5 Å of the experimental reference in only two, three and three systems, respectively, indicating that insufficient sampling was the predominant limitation, with ranking errors contributing to a smaller subset. Same-pocket recovery remained below 50% for all methods, despite broad preservation of receptor architecture, and method-native confidence measures showed limited discrimination of ligand-pose accuracy. Alternative ligand representations produced small aggregate changes but substantial system-specific effects on execution, pose category and ranking. Approximate close-contact screening identified additional geometric concerns in a subset of predictions. Collectively, these results show that current co-folding methods remain unreliable as stand-alone predictors of ion-channel ligand-binding modes and highlight pose sampling, pocket selection, ligand representation and independent structural validation as priorities for methodological development.

Voltage-gated ion channels are fundamental determinants of membrane excitability that underlies electrical signalling, calcium entry, secretion and muscle contraction and many other biological processes (Lopez-Mateos et al., 2025; Catterall et al., 2020). Aberrant activities of Nav, Cav and Kv channels, due to channelopathy mutations, are already established and continue to emerge as valid therapeutic targets across neurological, cardiovascular, muscular, inflammatory and psychiatric disorders (Catterall et al., 2020; Wulff et al., 2009). Pharmacological modulators of these as well as other ion channels typically involve pore blockers, gating modifiers and allosteric ligands with diverse mechanisms and degrees of subtype selectivity (Wulff et al., 2009; Li et al., 2024; Yao et al., 2024). The growing availability of high-resolution cryo-EM structures has provided an increasingly detailed view of these mechanisms, revealing ligand-bound within central cavities, lateral fenestrations, voltage-sensing domains, lipid-facing grooves and other allosteric sites (Lopez-Mateos et al., 2025; Li et al., 2024; Yao et al., 2024). These structures have strengthened the foundations for mechanistic pharmacology and structure-guided ligand discovery (Nam et al., 2025), while illustrating the considerable structural diversity of ion-channel recognition sites. Meanwhile, the continuing expansion of experimentally resolved ion-channel-ligand complexes in the PDB is providing an increasingly rich basis for systematic and progressively stringent evaluation of computational structure-prediction methods (Burley et al., 2024).

Despite these advances, ion channels remain particularly demanding targets for computational prediction of protein-ligand complexes. Nav and Cav pore-forming subunits contain four homologous but non-equivalent domains, whereas many Kv channels form oligomeric assemblies with binding sites located within or between subunit interfaces (Catterall et al., 2020). Their ligand-binding regions are frequently embedded within the membrane, incompletely solvent-accessible or coupled to specific functional states (Lopez-Mateos et al., 2025; Wulff et al., 2009). Moreover, the same channel may contain several chemically plausible cavities associated with distinct pharmacological mechanisms (Li et al., 2024; Yao et al., 2024). Ligand recognition may therefore depend on channel conformation, membrane exposure, auxiliary subunits and local rearrangements of pore-forming or voltage-sensing elements (Lopez-Mateos et al., 2025). Predicting an ion-channel complex requires not only the generation of a locally credible ligand pose, but also the identification of the relevant binding region within a large and conformationally heterogeneous membrane-protein architecture (Wulff et al., 2009).

All-atom co-folding methods offer a conceptually different route to this problem. AlphaFold 3 extended deep-learning-based structure prediction to joint models of proteins and other biomolecular components, including small molecules, ions, nucleic acids and modified residues, using an all-atom generative framework (Abramson et al., 2024). Subsequent methods, including Boltz-2 and Protenix-v2, have expanded the range of available co-folding implementations and inference strategies (Passaro et al., 2025; Zhang et al., 2026). Unlike conventional docking workflows, which generally place a ligand into a pre-existing receptor structure, co-folding methods generate the protein-ligand complex jointly from protein sequence and chemical input. This formulation may be advantageous when ligand placement is coupled to receptor adaptation or when the binding site is not specified in advance (Sledz and Caflisch, 2018). It also introduces a more demanding prediction problem, however, because errors in receptor conformation, pocket selection, ligand geometry and model ranking can arise within the same generative process (Buttenschoen et al., 2024; Masters et al., 2025; Morehead et al., 2026).

The rapid development of these methods has made rigorous evaluation increasingly important (Morehead et al., 2026; Skrinjar et al., 2026). Ligand root-mean-square deviation (RMSD) remains a useful measure of geometric agreement with an experimental pose, but it does not by itself establish physical or pharmacological validity (Robin et al., 2023; Leemann et al., 2023). Predictions with favourable RMSD values may nevertheless contain steric clashes or geometrically implausible ligand conformations (Buttenschoen et al., 2024), whereas a ligand with a larger atom-level deviation may still occupy the experimentally observed binding region (Robin et al., 2023; Errington et al., 2025). Evaluation must also distinguish between sampling and ranking (Kryshtafovych et al., 2023; Robin et al., 2023; Masters et al., 2025). A method may generate an accurate pose among several predictions but fail to assign it the highest confidence, which has different implications from failing to generate the pose at all (Masters et al., 2025; Morehead et al., 2026). Recent post-training-cut-off studies have also raised concerns regarding the generalisation of protein-ligand co-folding methods to structures and binding modes that differ substantially from their training data. These considerations motivate benchmarks that account explicitly for training overlap, prediction failures, multiple samples, method-native ranking and complementary measures of binding-site recovery (Kryshtafovych et al., 2023; Errington et al., 2025).

Chemical input representation provides an additional and comparatively underexplored source of variability. Co-folding methods may accept ligands through PDB Chemical Component Dictionary identifiers or through linear molecular representations such as SMILES (Westbrook et al., 2015). Chemical Component Dictionary entries contain curated descriptions associated with deposited PDB components, including atom definitions, stereochemical assignments and idealised coordinates. SMILES provides a compact representation of molecular connectivity and stereochemistry, but canonical strings depend on the software and standardisation procedure used to generate them (O'Boyle, 2012; Hahnke et al., 2018). Multiple valid strings may therefore encode the same molecular graph (David et al., 2020; O'Boyle, 2012).

Representations obtained from the RCSB PDB, PubChem and ChEMBL may differ in atom ordering, stereochemical annotation or chemical standardisation even when they nominally describe the same ligand (Burley et al., 2024; Hahnke et al., 2018; Zdrazil et al., 2024). Such differences do not necessarily alter the underlying molecular identity, but they may affect how a generative model tokenises, initialises or interprets the ligand (David et al., 2020). Sensitivity to chemical representation is therefore relevant to both the reproducibility and prospective application of protein-ligand co-folding methods (David et al., 2020; Zdrazil et al., 2024).

General protein-ligand benchmarks provide essential measures of overall method performance, but they may not resolve failure modes associated with specific target classes (Xu et al., 2025; Studer et al., 2026). Voltage-gated ion channels combine large system size, pseudo-symmetry or oligomeric assembly, multiple pharmacologically distinct pockets and pronounced conformational state dependence. These characteristics make them a stringent and therapeutically relevant setting in which to examine whether co-folding methods recover experimentally supported binding modes rather than merely plausible placements within the receptor. The increasing number of recently resolved complexes spanning different channel subtypes, conformational states and binding-site classes also makes it possible to construct benchmarks that are both more representative and less susceptible to training-data overlap (Skrinjar et al., 2026; Xu et al., 2025; Durairaj et al., 2024). Such datasets provide an opportunity to determine whether apparently successful predictions remain stable across chemical representations and whether accurate poses, when sampled, are reliably prioritised.

The specific gap addressed here is that no prior benchmark has jointly evaluated executable coverage, pose sampling versus ranking, binding-site recovery, ligand-representation sensitivity and geometric plausibility across a curated set of voltage-gated ion-channel complexes. CASP-style assessments and broad post-cut-off benchmarks provide essential general baselines, while a recent comparison of 74 small-molecule GPCR complexes identified accurate global receptor structures but variable and often inaccurate ligand placement (Shen et al., 2025). Supplementary Table S16 compares the target classes, dataset sizes, temporal designs and endpoints of these studies with the present benchmark.

Here, we present a benchmark of experimentally resolved ligand-bound structures of Nav, Cav and Kv channels for evaluating AlphaFold 3, Boltz-2 and Protenix-v2. Structures were assessed retrospectively against the relevant method-specific training cut-offs (Abramson et al., 2024; Passaro et al., 2025; Zhang et al., 2026), and direct comparisons were conducted on matched systems to limit bias arising from differential completion. We evaluated ligand placement at both the pose and binding-pocket levels, distinguished method-selected predictions from the most accurate sampled structures, examined approximate steric plausibility and assessed sensitivity to Chemical Component Dictionary and database-derived SMILES inputs (O'Boyle, 2012; Heller et al., 2015; Hahnke et al., 2018). By focusing on a structurally distinctive and pharmacologically important membrane-protein class, this study defines the current capabilities and limitations of protein-ligand co-folding for voltage-gated ion channels and provides a framework for interpreting such predictions in structure-guided channel pharmacology.

Benchmark design and system curation

We established a domain-specific benchmark to evaluate all-atom protein-ligand co-folding methods for voltage-gated sodium (Nav), calcium (Cav) and potassium (Kv) channels. The potassium-channel set also included small-conductance Ca 2+ -activated potassium channels, denoted K Ca /SK in the subfamily annotations (Kaczmarek et al., 2017). These systems were retained within the broader Kv-family category used in the principal tables (Brown et al., 2020). The benchmark comprised protein-ligand complexes for which an experimentally determined structure was available in the RCSB Protein Data Bank (PDB) (Burley et al., 2024).

The benchmark was designed to assess complementary aspects of co-folding performance, including executable coverage for large membrane-protein systems, recovery of experimentally observed ligand-binding modes, selection of accurate poses by method-native ranking, sensitivity to ligand representation, recovery of the correct binding region and the structural plausibility of the predicted complexes. AlphaFold 3, Boltz-2 and Protenix-v2 were included as the principal methods (Abramson et al., 2024; Passaro et al., 2025; Zhang et al., 2026). ESMFold2 was examined only in a limited feasibility analysis because the accessible application programming interface imposed a practical sequence-length restriction that excluded most full-length ion-channel systems. An at-a-glance overview of dataset assembly, model execution and scoring outputs is provided in Supplementary Figure S7.

Benchmark systems were selected manually by searching the RCSB PDB for recently released ligand-bound Nav, Cav and Kv structures. The initial selection was pragmatic and was not conducted as a formal systematic review. Method-specific training cut-offs were not applied during the initial collection of structures. Instead, training-overlap filtering was performed retrospectively after the candidate set had been assembled.

For each retained system, we recorded the PDB identifier, channel family and subtype, ligand name and Chemical Component Dictionary (CCD) identifier (Westbrook et al., 2015), resolution where available, initial PDB release date, protein-chain composition, ligand-chain information, protein token count and available ligand structural descriptors. Structures that could not be prepared consistently for the benchmark or did not satisfy the subsequent method-specific training-cut-off criteria were excluded from the relevant blind-comparison subsets. System-level composition and size metadata are provided in Supplementary Table S1 and run status, failure reason and analysis inclusion are documented for every attempted job in Supplementary Table S2.

Ligand representations and training-cut-off policy

Three ligand-input conditions were prepared where suitable representations were available. The primary analysis used the CACTVS canonical SMILES (Westbrook et al., 2015) string provided for the corresponding RCSB PDB chemical component. A second condition used the PDB CCD identifier directly (Westbrook et al., 2015), allowing each method to interpret the ligand using its internal chemical-component definitions. A third condition used matched PubChem and ChEMBL SMILES (Hahnke et al., 2018; Zdrazil et al., 2024) strings to examine whether the source and standardisation of the ligand representation influenced prediction outcomes.

Alternative SMILES strings were retained only when curated identifiers and chemical-matching checks indicated that they represented the same compound as the deposited PDB ligand (Heller et al., 2015). Connectivity-only representations that removed stereochemical information (O'Boyle, 2012) were not included in the principal sensitivity analysis, because these inputs would address a distinct question concerning the contribution of stereochemistry.

To reduce the risk of training-set overlap, benchmark eligibility was assessed retrospectively and independently for each method using the initial release date of the corresponding experimental structure. A system was considered a blind test case only when its PDB release date was strictly later than the reported training cut-off of the method. The cut-off dates applied were 30 September 2021 for AlphaFold 3 and Protenix-v2 and 1 June 2023 for Boltz-2 (Abramson et al., 2024; Passaro et al., 2025; Zhang et al., 2026). Structures released on or before the relevant cut-off were not interpreted as blind generalisation cases, although they could be retained as pre-cut-off controls or for analyses not requiring a blind test set.

Separate analysis sets were defined to avoid combining comparisons with different denominators. These comprised a broad-coverage set containing all attempted and completed jobs, a common primary-SMILES subset completed by all three principal methods, a paired CCD-SMILES subset and a matched SMILES-source-sensitivity subset containing available RCSB PDB, PubChem and ChEMBL representations.

Input preparation and structure prediction

Method-specific inputs were generated from a shared curated metadata table to preserve protein sequences, chain composition, ligand identity and representation labels across methods. To reduce differences arising from independent sequence-search procedures, multiple-sequence alignment (MSA) outputs were generated using the AlphaFold 3 data pipeline and reused by Boltz-2 and Protenix-v2 where supported.

AlphaFold 3 template searches were restricted using a maximum template date of 30 September 2021 (Abramson et al., 2024). AlphaFold 3 inference was subsequently performed using the precomputed data-pipeline outputs. Protenix-v2 was run with MSA information enabled and structural-template use disabled. This strategy was intended to reduce avoidable variation in evolutionary information while preserving the method-specific structure-generation and ranking procedures.

AlphaFold 3, Boltz-2 and Protenix-v2 predictions were generated on institutional high-performance-computing infrastructure using NVIDIA A100 GPUs. Each method-system-representation job was evaluated using five independent random seeds, with five samples generated per seed, giving up to 25 predicted structures for a completed production job. Where supported by the software, calculations used 10 recycling or model-cycle iterations and 200 sampling or diffusion steps. Method-native settings were retained when directly equivalent parameters were not exposed.

Data preparation and structural inference for AlphaFold 3 were conducted as separate stages. AlphaFold 3-derived MSA outputs were reused by Boltz-2 and Protenix-v2 where supported. Protenix-v2 inputs retained paired and unpaired MSA paths, whereas Boltz-2 inputs referenced supplied A3M MSA files.

Job completion was assessed using the expected coordinate-file count, execution logs and the final run inventory. Failed jobs were retained as benchmark outcomes rather than excluded from the analysis. Two Protenix-v2 CCD-control jobs, corresponding to PDB entries 21 TP and 21TQ, failed because the internal chemical-component handling did not successfully parse the A1E26 component.

A total of 7,525 coordinate files were generated from the 301 successfully completed AlphaFold 3, Boltz-2, and Protenix-v2 jobs. Optional ESMFold2 API outputs were retained separately and were not included in the primary comparative analyses. Protenix-v1 (Team et al., 2026) webserver predictions were excluded because the available model had a training cut-off of 30 June 2025 and the successfully retrieved compatible systems pre-dated that cut-off.

Structural alignment and ligand-pose evaluation

Experimental structures were obtained in PDBx/mmCIF format (Westbrook et al., 2022). Reference ligands were identified using the curated CCD identifier and deposited ligand-chain information. Water molecules and hydrogen atoms were excluded from ligand-pose, pocket and close-contact calculations. Predictions with missing structures, unidentifiable ligands or incompatible atom mappings were retained with explicit status labels rather than removed without documentation.

Before ligand evaluation, each predicted receptor was superposed onto the corresponding experimental structure. Cα atoms were matched using available chain and residue identifiers, and the rigid-body transformation was determined by least-squares superposition. Where direct residue matching was insufficient because of construct, numbering or chain-order differences, a documented resolved-region fallback was used.

Ligand placement was evaluated using the distance between the predicted and experimental ligand centroids and the ligand heavy-atom root-mean-square deviation (RMSD) following receptor alignment. Heavy-atom RMSD was reported only when the predicted and reference ligands had compatible atom counts and an interpretable atom correspondence (Bell and Zhang, 2019; Meli and Biggin, 2020). The principal RMSD measure was not symmetry corrected. Centroid- and pocket-based measures were therefore retained as complementary endpoints for ligands affected by equivalent-atom permutations or incompatible atom ordering.

Predictions with ligand RMSD below 2 Å were classified as near-native. Predictions with RMSD from 2 to below 5 Å were considered approximately experimental or locally site-correct, whereas predictions with RMSD of 5 Å or greater were considered incorrect at the pose level. The 5 Å threshold was interpreted as a permissive measure of local binding-site recovery rather than evidence of precise atom-level pose reproduction.

Reference-pocket residues were defined as residues containing at least one heavy atom within 5 Å of the experimental ligand. Predicted-pocket residues were defined analogously after receptor superposition. Pocket recall, precision and Jaccard overlap were calculated from ligand-contacted residue sets. A same-pocket classification was additionally assigned using the predefined ligand-centroid and residue-overlap criteria applied in the analysis workflow. These measures were used to distinguish complete binding-site failures from predictions in which the ligand occupied the appropriate pore, fenestration, voltage-sensor-associated or allosteric region but adopted an imperfect orientation. Structural superpositions, predicted binding modes and representative protein-ligand interactions were inspected and visualised using UCSF ChimeraX (Pettersen et al., 2021).

Sampling, ranking and robustness analyses

For each completed job, the top-ranked prediction was identified using the method-native scalar confidence or ranking score. The ligand-oracle prediction was defined as the sampled structure with the lowest valid ligand RMSD, and the centroid distance was used as a secondary criterion when RMSD could not be calculated.

The ranking gap was calculated as the difference between the top-ranked and ligand-oracle RMSD values. A positive value therefore indicated that the method had sampled a pose closer to the experimental structure than the pose selected by its native ranking procedure. Jobs were classified as top-ranked near-experimental, sampled but misranked or not sampled. A near-experimental binding mode was considered sampled when at least one predicted structure had a ligand RMSD below 5 Å. This distinction was used to separate limitations in conformational sampling from limitations in model ranking.

Method-native confidence quantities, including pTM-, ipTM- and pLDDT-like scores where available (Jumper et al., 2021; Evans et al., 2022), were treated as method-specific diagnostic measures rather than directly comparable probabilities. Their associations with ligand RMSD, centroid displacement and pocket recovery were evaluated separately for each method and input condition.

Receptor-level structural agreement was assessed independently of ligand placement using US-align in multimer mode (Zhang et al., 2022). Model-normalised and reference-normalised TM-scores (Zhang and Skolnick, 2004) and the aligned fraction of the experimental receptor were retained as scaffold-level quality controls and were not interpreted as measures of ligand-pose accuracy.

CCD and SMILES inputs were compared only for matched method-system pairs in which both calculations completed. SMILES-source sensitivity was evaluated across matched RCSB PDB, PubChem and ChEMBL inputs. Systems were considered failure-sensitive when one or more expected representations did not complete, pose-sensitive when the pose class or RMSD category changed across sources, ranking-sensitive when sampling was comparatively stable but the selected model differed and approximately invariant when both sampling and ranking outcomes remained consistent.

Physical plausibility was assessed independently of RMSD. Top-ranked and ligand-oracle structures were screened for unusually short heavy-atom contacts. Poses were flagged for manual review when any protein-ligand heavy-atom pair was separated by < 2.0 Å or when any non-identical intraligand heavy-atom pair was separated by < 1.0 Å. This screen was used as an approximate close-contact check rather than a full PoseBusters or chemical-validity assessment, and did not evaluate bond geometry, valence, protonation, chirality, aromaticity, ring planarity or ligand strain.

Quantitative analyses, including structural alignment, ligand-pose evaluation, pocket comparison, confidence-score extraction, close-contact screening and aggregation of per-model and per-system results, were performed using in-house Python scripts.

Statistical analysis and reproducibility

Paired statistical procedures were used where the same experimental systems were compared across prediction conditions. Top-ranked and ligand-oracle RMSD values were compared within each method using paired Wilcoxon signed-rank tests, whereas matched CCD and SMILES conditions were assessed using two-sided paired Wilcoxon signed-rank tests. Comparisons among the three principal methods on the common subset were performed using Friedman tests with systems treated as paired blocks, followed by paired post hoc comparisons where appropriate. Benjamini-Hochberg correction was applied across related tests, and exact sample sizes, unadjusted P values and adjusted q values were reported. Missing or unscorable observations were not imputed.

Runtime was included as operational context rather than as a formal comparison of algorithmic efficiency. Available timings represented post-MSA prediction workflows executed on NVIDIA A100 GPUs and generally excluded the initial AlphaFold 3 sequence-search stage. Runtime comparisons were interpreted cautiously because preprocessing, input/output operations, scheduler behaviour and software wrappers differed among methods.

Co-folding predictions were performed utilising AlphaFold 3 (v3.0.1), Boltz-2 (v2.1.1), and Protenix-v2 (v2.0.0). All input generation, run auditing, structural evaluation, statistical analysis and figure preparation were performed using reproducible scripted workflows. A per-model manifest recorded the method, benchmark panel, ligand representation, PDB identifier, random seed, sample index, coordinate-file path and available confidence-file path.

A training-cut-off-aware ion-channel benchmark reveals uneven executable coverage

We assembled a benchmark of experimentally resolved Nav, Kv and Cav protein-ligand complexes to evaluate AlphaFold 3, Boltz-2 and Protenix-v2. The benchmark comprised three principal ligand-input panels: RCSB/CACTVS SMILES, PDB Chemical Component Dictionary (CCD) controls and matched PubChem and ChEMBL SMILES variants. Method-specific training cut-offs were applied retrospectively to define the appropriate post-cut-off comparisons. Protenix-v1 webserver outputs were excluded because the successfully completed compatible systems overlapped its reported training period.

Optional ESMFold2 API predictions were retained only as an ancillary feasibility note. These runs comprised five repeats of the 7W77 RCSB/CACTVS SMILES input and single-repeat CCD and PubChem variants. Because these outputs were not part of the harmonised post-cut-off comparative design and were not scored in the final ligand-RMSD tables, they were not interpreted as a quantitative comparison with AlphaFold 3, Boltz-2 or Protenix-v2.

Executable coverage differed substantially across methods and ligand representations (Figures 1A–C; Table 1). In the primary RCSB/CACTVS SMILES panel, AlphaFold 3 completed all 44 systems, whereas Boltz-2 and Protenix-v2 completed 23 and 25 systems, respectively. The difference was most pronounced among Kv-family systems, including K Ca /SK entries: AlphaFold 3 completed all 20 primary-SMILES jobs, compared with none for Boltz-2 and three for Protenix-v2. Detailed denominators, run-level outcomes, failure categories and training-cut-off classifications are provided in Supplementary Figure S1 and Supplementary Tables S2-S4.

CCD inputs were less consistently executable. AlphaFold 3 and Boltz-2 each completed 21 of 44 CCD-control jobs. Protenix-v2 completed 23 of the 25 jobs in its compatible paired-CCD set while the remaining two failed during chemical-component parsing. Within the selected method-specific SMILES-source-sensitivity sets, all submitted jobs completed for each method: 65 of 65 for AlphaFold 3, 38 of 38 for Boltz-2 and 41 of 41 for Protenix-v2.

The final production inventory comprised 301 completed AlphaFold 3, Boltz-2 and Protenix-v2 method-input jobs, corresponding to 7,525 model coordinate files. Prediction failures were preserved as benchmark outcomes because their exclusion would have altered the distribution of channel families and ligand representations available to each method. Direct accuracy comparisons were therefore restricted to a common subset completed by all three principal methods.

Near-experimental ligand poses are infrequently sampled in the common subset

The primary comparative analysis comprised 22 RCSB/CACTVS SMILES systems completed by AlphaFold 3, Boltz-2 and Protenix-v2. Each method generated 25 models per system. Ligand-pose accuracy was evaluated after alignment of the predicted receptor to the corresponding experimental structure. RMSD values reported in this section refer to ligand heavy-atom RMSD rather than receptor backbone RMSD.

Top-ranked ligand poses were generally distant from their experimental counterparts (Figures 2A–D; Table 2). Median top-ranked RMSDs were 19.85 Å for AlphaFold 3, 22.12 Å for Boltz-2 and 20.25 Å for Protenix-v2. None of the methods produced a top-ranked pose below the near-native threshold of 2 Å. At the more permissive threshold of 5 Å, top-ranked recovery was observed for two AlphaFold 3 systems, one Boltz-2 system and no Protenix-v2 systems. Per-model metrics and per-job top-ranked and oracle summaries are provided in Supplementary Tables S5, S6.

Selecting the lowest-RMSD structure from the 25 sampled models reduced the median RMSD to 12.40 Å for AlphaFold 3, 18.61 Å for Boltz-2 and 12.90 Å for Protenix-v2. Best-of-25, or oracle, recovery within 5 Å nevertheless remained uncommon, occurring in two, three and three systems for AlphaFold 3, Boltz-2 and Protenix-v2, respectively. Under the sampling regime used here, most method-system combinations therefore did not contain a near-experimental ligand pose.

Best-of-25 selection significantly improved ligand RMSD relative to the corresponding top-ranked model within each method (Figure 2B; AlphaFold 3, P value = 1.5 × 10 -4; Boltz-2, P value = 1.5 × 10 -4; Protenix-v2, P value <1 × 10 -4; Supplementary Table S15). By contrast, paired comparisons among methods did not provide strong evidence for differences in top-ranked RMSD after correction for multiple testing. The common-subset analysis therefore did not identify a consistently superior method, but instead revealed a shared difficulty in sampling experimentally consistent ion-channel ligand poses.

Classification based strictly on ligand heavy-atom RMSD further distinguished sampling from ranking behaviour (Figure 2D). AlphaFold 3 sampled a pose within 5 Å in two of 22 systems, both of which were selected as top-ranked. Boltz-2 sampled such a pose in three systems but ranked it first in only one. Protenix-v2 sampled a pose within 5 Å in three systems by strict ligand heavy-atom RMSD, none of which was selected as top-ranked. Near-experimental poses were therefore not sampled in 20 of 22, 19 of 22 and 18 of 22 systems for AlphaFold 3, Boltz-2 and Protenix-v2, respectively. Improved ranking could recover a small number of additional Boltz-2 and Protenix-v2 predictions, but would not address the more frequent absence of an accurate pose from the sampled ensemble. System-level RMSD values and sampling classifications are shown in Supplementary Figure S2 and Supplementary Table S7.

Consistent with this interpretation, the median difference between top-ranked and oracle RMSD was 3.88 Å for AlphaFold 3, 2.48 Å for Boltz-2 and 2.24 Å for Protenix-v2 (Figure 2C). Method-native ranking often failed to select the most accurate available model, although the best sampled pose was itself generally far from the experimental ligand position. All-model RMSD distributions, near-experimental sampling fractions and confidence ranks of the best sampled poses are shown in Supplementary Figure S3.

Ligand representation introduces system-specific instability

We next examined whether ligand representation altered co-folding outcomes independently of protein sequence and target identity. Matched CCD and RCSB/CACTVS SMILES calculations showed only modest differences in ligand-pose accuracy (Figure 3A; Supplementary Table S8). Median CCD-minus-SMILES differences in top-ranked ligand RMSD were −0.75 Å for AlphaFold 3, −1.25 Å for Boltz-2 and −0.24 Å for Protenix-v2. The corresponding differences for the best-of-25 models were −0.76 Å, −1.33 Å and −0.83 Å, respectively. These shifts were small relative to the absolute pose errors and did not identify CCD input as consistently superior.

A broader pattern emerged when chemically matched SMILES representations from RCSB/CACTVS, PubChem and ChEMBL were compared (Figure 3B; Supplementary Figure S4; Supplementary Table S9). AlphaFold 3 produced seven invariant, ten ranking-sensitive, eight pose-sensitive and fifteen failure-sensitive method-system groups. Boltz-2 produced three, seven, seven and four groups in the corresponding categories, whereas Protenix-v2 produced nine, five, four and five. Pose-sensitive groups were defined by a change in pose class or RMSD category across ligand sources. Ranking-sensitive groups retained a broadly similar sampled outcome but differed in the model prioritised by the method-native ranking procedure.

Despite this system-level variability, the aggregate source effects were small. Relative to RCSB/CACTVS, median PubChem-minus-CACTVS differences in top-ranked RMSD were −0.11 Å for AlphaFold 3, −0.06 Å for Boltz-2 and −0.42 Å for Protenix-v2. Median ChEMBL-minus-CACTVS differences were −0.02 Å, 0.24 Å and −0.05 Å, respectively. No SMILES source conferred a consistent advantage across methods.

The principal representation effect was therefore not a uniform shift in benchmark-wide accuracy, but instability in individual target-ligand systems. Alternative encodings of the same nominal ligand could alter completion, the pose category sampled or the identity of the top-ranked model. This variability indicates that ligand representation is an experimental factor in co-folding rather than a neutral input convention.

Pocket recovery and confidence reveal limitations beyond ligand RMSD

Ligand RMSD does not distinguish between an incorrectly oriented ligand within the experimental site and placement in a different region of the receptor. We therefore evaluated recovery of the experimentally defined local pocket using ligand-contacted residue sets (Figure 4A; Table 3).

Among top-ranked predictions, same-pocket recovery was observed in three of 22 AlphaFold 3 systems, five of 22 Boltz-2 systems and six of 22 Protenix-v2 systems. Selection of the best-of-25 model increased recovery to seven, eight and nine systems, respectively. Thus, pocket recovery was more frequent than strict pose recovery, but remained below half of the common subset for all three methods. Representative-level pocket classifications are provided in Supplementary Table S10.

These results separate two distinct classes of error. In some systems, the ligand was placed in the correct local environment but adopted an inaccurate orientation or conformation. In others, it occupied a different region of the receptor altogether. This distinction is particularly consequential for ion channels, where central-cavity, fenestration, voltage-sensor-associated and extracellular pore sites may be spatially proximal yet correspond to different mechanisms of modulation.

Method-native confidence features provided only limited discrimination of these outcomes (Figures 4B,C; Supplementary Figure S5; Supplementary Table S11). In the primary RCSB/CACTVS SMILES panel, correlations between the principal scalar confidence measure and ligand RMSD were −0.407 for AlphaFold 3, −0.095 for Boltz-2 and −0.285 for Protenix-v2. Correlations involving ipTM-like quantities were similarly modest, whereas harmonised pLDDT-like values were available only for Boltz-2 and Protenix-v2.

The confidence measures therefore contained some method-specific information but did not reliably distinguish accurate from inaccurate ligand poses. Their differing definitions and calibration further preclude direct comparison as equivalent observables across methods. In the present benchmark, confidence was most informative when interpreted together with sampling behaviour and pocket identity rather than as an independent measure of ligand-pose correctness.

Binding-site context, receptor architecture and geometric plausibility delineate distinct failure modes

We next examined whether performance varied with the biological context of the binding site (Figure 5A; Supplementary Table S12). The small number of poses recovered within 5 Å was concentrated in Nav pore-associated or selectivity-filter settings. Fenestration or lipid-facing sites, voltage-sensor-associated pockets and several Cav systems with ambiguous, multiple or uncurated site assignments remained challenging.

These patterns should be interpreted cautiously because some classes contained few systems and several Cav annotations lacked sufficient specificity for detailed comparison. Nevertheless, the site-stratified analysis indicates that ligand-pose error was not uniformly distributed across the benchmark. Binding-site architecture and pharmacological context contributed to the observed heterogeneity, particularly for sites formed by membrane-facing, state-dependent or allosteric channel elements.

To determine whether ligand-pose failure primarily reflected an incorrect receptor structure, we assessed receptor-level similarity using US-align in multimeric protein mode (Figure 5B; Supplementary Table S14). Median reference-normalised TM-scores for the top-ranked structures were 0.807 for AlphaFold 3, 0.780 for Boltz-2 and 0.759 for Protenix-v2. The corresponding median aligned fractions of the experimental receptors were approximately 0.80-0.81. Similar values were obtained for the best-of-25 representatives.

The receptor scaffolds were therefore often broadly consistent with the experimental structures even when ligand placement was inaccurate. Construct-length differences contributed to reduced aligned fractions in some Cav systems, and the TM-scores do not establish the accuracy of local binding-site conformations. Nonetheless, the analysis argues against a general model in which ligand failure arose predominantly from collapse of the global channel fold. Rather, many predictions preserved the overall receptor architecture while failing at the level of local pocket selection or ligand placement.

Approximate geometric plausibility provided an additional and independent quality axis (Figure 5C; Supplementary Figure S6B; Supplementary Table S13). Among 602 top-ranked and best-of-25 representatives, 154 were flagged for unusually short heavy-atom contacts and 448 had no close-contact flag. The flagged structures were evenly divided between top-ranked and best-of-25 representatives, with 77 in each group. By method, 89 AlphaFold 3 and 65 Protenix-v2 representatives were flagged.

None of the evaluated Boltz-2 representatives contained a protein-ligand heavy-atom separation below 2.0 Å or a non-identical intraligand heavy-atom separation below 1.0 Å. This result is specific to the predefined close-contact criteria and does not establish ligand-pose accuracy or broader chemical validity. The method subsets also differed in composition, and the absence of a flag should not be interpreted as evidence of overall superiority in ligand geometry.

More generally, this screen did not evaluate bond geometry, valence, protonation, chirality, aromaticity, ring planarity or conformational strain. Its purpose was to identify grossly short contacts and to demonstrate that local steric plausibility is separable from ligand RMSD, pocket correctness and ranking. Representative examples of outer-pore recovery, a large ranking gap with wrong-site top placement, and central-cavity recovery are shown in Figure 6. The remaining ligand-representation, Kv-context and close-contact examples are retained as Supplementary Figures S6C-E.

Runtime measurements were analysed separately as operational context (Supplementary Figure S6A). The available values were derived from completed one-seed calculations on NVIDIA A100 GPUs and represent post-MSA prediction workflows. AlphaFold 3 timings excluded the preceding data-pipeline stage, whereas Boltz-2 and Protenix-v2 reused AlphaFold 3-derived MSA outputs where supported. Because preprocessing, job organisation, input/output overhead and execution wrappers differed among methods, these measurements were not interpreted as an end-to-end comparison of computational efficiency.

In summary, the results resolve several distinct limitations of current ion-channel co-folding. In the balanced Nav and Cav comparison, the predominant failure was the absence of a near-experimental ligand pose from the sampled ensemble. Ligand representation introduced additional system-specific instability, while pocket analysis showed that many predictions failed to recover the experimental local environment. Method-native confidence measures only weakly discriminated pose accuracy, and broad preservation of the receptor scaffold did not ensure correct ligand placement. Approximate close-contact screening further demonstrated that geometric plausibility constitutes an additional criterion independent of sampling, ranking and pocket recovery. Under the conditions evaluated here, these failure modes collectively limited reliable prediction of experimentally resolved Nav, Cav and Kv-family ligand-binding modes. A consolidated comparison of executable coverage and the principal accuracy, pocket, plausibility and confidence outcomes is provided in Table 4.

Coverage is calculated on the 44-system primary RCSB/CACTVS SMILES panel. RMSD, pocket recovery, confidence correlation and dominant sampling failure are calculated on the common 22-system Nav/Cav subset. Close-contact rates use all top-ranked and oracle representatives across completed panels (method-specific denominators; 602 representatives in total) and are an approximate review flag, not a cross-method measure of chemical quality.

This study provides a target-class-specific assessment of all-atom protein-ligand co-folding for Nav, Cav and Kv-family channels. Under the post-training-cut-off conditions and sampling regime evaluated here, AlphaFold 3, Boltz-2 and Protenix-v2 did not consistently reproduce experimentally resolved ligand-binding modes. In the strict common subset, the principal limitation was failure to sample a near-experimental pose rather than ranking alone. Uneven executable coverage, incorrect pocket selection, ligand-representation sensitivity, limited discrimination by method-native confidence scores and unusually short heavy-atom contacts introduced further uncertainty.

Executable coverage should be regarded as part of benchmark performance (Leemann et al., 2023; Skrinjar et al., 2026; Studer et al., 2026) rather than a peripheral implementation detail. AlphaFold 3 completed all systems in the primary RCSB/CACTVS SMILES panel, whereas Boltz-2 and Protenix-v2 completed smaller and compositionally different subsets. The disparity was particularly marked for Kv-family systems. In the primary SMILES and CCD-control panels, Boltz-2 did not complete any attempted Kv-family jobs. Although the run ledger records these cases under a general not-completed category, the raw logs indicated GPU execution failures under the one-A100 settings, including out-of-memory warnings and CUDA illegal-memory-access errors during inference. We therefore treated these cases as executable-coverage failures rather than ligand-pose failures.

This result does not imply that Boltz-2 is intrinsically unable to model Kv channels. Completion may depend on the input, software configuration, memory demand and hardware. The conclusion is restricted to the attempted primary-SMILES and CCD-control Kv-family systems under the settings tested and no Kv-family Boltz-2 jobs were submitted to the final SMILES-source-sensitivity panel. Nevertheless, excluding failed jobs would alter the family, system-size and ligand-site composition of the benchmark and could overstate practical performance. Executable coverage and pose accuracy are therefore complementary properties (Leemann et al., 2023; Skrinjar et al., 2026; Studer et al., 2026).

The Kv-family category included K Ca /SK channels (Nam et al., 2025; Kaczmarek et al., 2017; Brown et al., 2020), which were identified separately in the subfamily annotations. These small-conductance Ca 2+ -activated potassium channels extend the benchmark beyond canonical pore and fenestration ligands through calmodulin-, gating- or cofactor-associated ligand environments (Brown et al., 2020). They should therefore be interpreted as Kv-family/K Ca -SK cases rather than conventional voltage-gated Kv systems. Because Boltz-2 completed no Kv-family systems in the primary-SMILES panel, the strict 22-system common subset contained Nav and Cav complexes only. Direct pose-accuracy comparisons consequently apply to the balanced Nav and Cav subset and should not be extrapolated to Kv or K Ca /SK ligand placement.

Within the common subset, selecting the lowest-RMSD structure from 25 predictions improved ligand-pose accuracy relative to method-native top ranking for all three methods. Strict ligand heavy-atom RMSD recovery within 5 Å was nevertheless observed in only two AlphaFold 3 systems and three systems each for Boltz-2 and Protenix-v2. Near-experimental poses were absent from most sampled ensembles. Positive ranking gaps, together with the failure of Boltz-2 and Protenix-v2 to prioritise several of their best sampled poses, establish ranking as a secondary limitation (Masters et al., 2025; Morehead et al., 2026). Improved confidence estimation or post hoc rescoring could recover individual cases in which a useful pose had already been generated, but would not address the more frequent failure to enter the correct pose basin.

This interpretation remains conditional on the sampling budget. Twenty-five models per method-input job provide a meaningful assessment of ensemble behaviour but do not exhaust ligand and receptor conformational space. Greater sampling or pocket-guided constraints might improve recovery (Ghersi and Sanchez, 2009). The data therefore show that experimental poses were rarely obtained under a standardised, computationally tractable regime, not that they are inaccessible to the methods. Even so, the low oracle success rate indicates that reranking alone is unlikely to yield reliable blind ion-channel pose prediction.

Receptor-level analysis localised these failures more precisely. Median reference-normalised TM-scores of approximately 0.76–0.81 indicate that the broad channel architecture was often recognisable even when ligand placement was inaccurate. Global scaffold recovery is thus insufficient evidence that a predicted complex is correct. Relatively small changes in pore, fenestration, voltage-sensor or interfacial regions can alter pocket geometry without substantially affecting the overall fold (Lopez-Mateos et al., 2025; Li et al., 2024; Yao et al., 2024).

Pocket-level evaluation further separated distinct modes of error. Same-pocket placement was more frequent than strict pose recovery, particularly after oracle selection, but remained below half of the common subset for every method. Some predictions placed the ligand in an incorrect region of an otherwise recognisable channel, whereas others identified the experimental pocket but failed to reproduce its orientation or local geometry. These outcomes have different pharmacological implications. Displacement between a central cavity, fenestration, voltage-sensor-associated site and extracellular pore region may imply a different mechanism of modulation (Lopez-Mateos et al., 2025; Li et al., 2024; Yao et al., 2024), while an inaccurate pose within the correct pocket may retain limited site information but remain unsuitable for interaction-level interpretation or structure-guided chemical design.

Several fundamental biological and physical features of voltage-gated ion channels make the joint receptor-ligand search exceptionally demanding. Foremost among these is the pronounced conformational heterogeneity and state dependence of ligand binding. Ion channels are highly dynamic assemblies that transition between resting, open, and inactivated states, with each conformation exposing distinct pharmacological pockets. Because current co-folding models tend to default to a single dominant conformation, typically the state most frequently represented in their training data, they often fail to generate the specific local geometry required for a given ligand. Consequently, local rearrangements of an activation gate, fenestration, selectivity filter, or voltage-sensing domain can materially reshape a binding site without substantially altering the global fold.

The structural ambiguities introduced by state dependence are compounded by the absence of an explicit membrane environment during inference. Many ion-channel modulators, particularly those targeting lateral fenestrations or voltage-sensor domains, partition into the lipid bilayer and bind directly at protein-lipid interfaces. Without lipid representations or spatial constraints to define these boundary grooves, generative models struggle to reconstruct membrane-facing binding sites accurately. The immense chemical diversity of ion-channel ligands, spanning charged pore blockers to flexible lipophilic allosteric modulators, exacerbates this sampling problem. Incomplete specification of protonation, tautomeric state, or stereochemistry introduces further uncertainty, especially given the limited availability of high-resolution structural training data for diverse, state-dependent channel complexes compared to soluble proteins.

These difficulties are not exclusive to ion channels and appear to reflect a broader limitation of current co-folding models when applied to complex membrane proteins. In a recent evaluation of 74 AlphaFold 3 GPCR-small-molecule complexes (Shen et al., 2025), global receptor architecture and orthosteric pocket geometry were generally recovered, whereas ligand positioning remained variable and was frequently inaccurate, particularly for allosteric modulators. This parallel observation suggests that the limitations documented here indicate a general deficiency in modelling state-dependent membrane-protein complexes rather than an isolated ion-channel failure. Voltage-gated ion channels nevertheless impose additional and distinctive structural burdens. Features such as four-domain pseudo-symmetry in Nav and Cav proteins, alongside oligomeric interfaces in many Kv-family channels, substantially expand the number of physically plausible but pharmacologically incorrect placements.

Method-native confidence measures did not resolve these ambiguities. Their correlations with ligand RMSD were weak to moderate and differed among methods. Because the scores arise from distinct architectures and training objectives, they should not be treated as interchangeable or as calibrated probabilities of pose correctness (Skrinjar et al., 2026). They are best used as within-method diagnostics alongside ensemble consistency, pocket identity and structural inspection. Confidence measures that distinguish receptor quality, pocket correctness, pose accuracy and chemical plausibility would be more informative for future applications.

The findings are consistent with concerns about generalisation but do not demonstrate training-set memorisation (Masters et al., 2025; Xu et al., 2025). Retrospective filtering by PDB release date reduces the direct risk that a reference structure pre-dated a method-specific cut-off, yet chronological separation is not equivalent to structural novelty (Durairaj et al., 2024; Rogers and Hahn, 2010). Post-cut-off complexes may remain close to earlier channel, ligand or pocket examples. Future versions of the benchmark should therefore complement temporal filtering with quantitative protein-, ligand- and pocket-similarity analyses.

Ligand representation was another source of variability. CCD inputs produced small median improvements relative to RCSB/CACTVS SMILES, but these shifts were modest compared with the overall errors and were not consistent across systems. PubChem and ChEMBL variants likewise produced small aggregate differences while altering completion, ranking or RMSD category for individual complexes. Nominally equivalent encodings were therefore not always prediction-equivalent (Westbrook et al., 2015). Ligand source and stereochemical specification should consequently be reported, and agreement across chemically equivalent encodings may provide a useful robustness check. Because protonation, tautomerism and formal charge were not varied systematically, the chemical basis of this sensitivity remains unresolved.

The close-contact analysis provided a separate, deliberately limited assessment of structural plausibility. A subset of top-ranked and oracle representatives contained protein-ligand or intraligand heavy-atom separations below the predefined thresholds. These flags identify structures requiring inspection but do not constitute comprehensive chemical validation. Their absence likewise does not establish correct bond geometry, valence, protonation, chirality, aromaticity, ring planarity or conformational strain. The absence of flags among the evaluated Boltz-2 representatives does not establish chemical superiority, because the completed subsets differed and coordinate-generation procedures may contribute. Chemically informed validation or force-field-based refinement (Sledz and Caflisch, 2018; Errington et al., 2025; Fischer et al., 2021) would be required for a fuller comparison.

Several limitations temper the conclusions. The benchmark was assembled pragmatically from recently released RCSB PDB structures, and its composition does not represent the full pharmacological diversity of Nav, Cav and Kv-family channels. The common subset was small and contained no Kv-family systems, while several Cav site annotations remained broad or uncertain. Site-stratified trends should therefore be regarded as descriptive. Ligand RMSD was calculated only where atom counts and ordering were compatible and was not symmetry corrected, potentially overestimating error for ligands with equivalent atoms (Bell and Zhang, 2019; Meli and Biggin, 2020). Centroid and pocket measures provide complementary information but do not eliminate this limitation, and same-pocket classification is not a direct measure of native interaction recovery.

AlphaFold 3-derived MSA outputs reduced one source of variation but did not make the inference inputs identical. Boltz-2 referenced supplied A3M files, whereas Protenix-v2 retained paired and unpaired MSA paths, and each method processed these data through its own architecture. The benchmark also did not evaluate conventional or pocket-informed docking, molecular dynamics, free-energy calculations, binding-affinity prediction or prospective compound prioritisation (Warren et al., 2006; Durrant and McCammon, 2011). It therefore supports comparison among the evaluated co-folding workflows but not claims of superiority or inferiority to receptor-based methods.

As channel-ligand structures accumulate, the benchmark can be expanded to include balanced Kv and K Ca /SK representation, additional states and broader regulatory and allosteric site coverage. Future analyses should combine temporal filtering with similarity controls (Skrinjar et al., 2026; Xu et al., 2025), strengthen state and site annotation, evaluate larger or adaptively selected ensembles and incorporate symmetry-aware ligand mapping and chemically typed interaction analyses. Comparison with pocket-informed docking and hybrid refinement (Ghersi and Sanchez, 2009; Fischer et al., 2021) would help establish whether the principal value of co-folding lies in receptor generation, pocket identification, pose sampling or ensemble diversification.

For current structure-guided ion-channel research, the results support a cautious but constructive role for co-folding. These methods may assist hypothesis generation, propose candidate binding regions or provide ensembles for further analysis, but should not be treated as stand-alone evidence of a ligand-binding mechanism (Buttenschoen et al., 2024; Masters et al., 2025; Morehead et al., 2026). More reliable workflows will combine repeated predictions and alternative ligand encodings with pocket-aware inspection, chemical validation, physics-based refinement, comparison with experimental structures and functional or structural testing (Sledz and Caflisch, 2018; Buttenschoen et al., 2024; Durrant and McCammon, 2011).

The principal conclusion is not that one evaluated method is categorically preferable, but that ion-channel protein-ligand co-folding remains limited along several separable axes. Under the conditions tested, sampling failure was the dominant obstacle in the balanced Nav and Cav comparison, while executable coverage, ranking, local pocket selection, ligand representation and structural plausibility introduced additional uncertainty. Kv-family and K Ca /SK systems broadened the benchmark towards regulatory and non-classical channel-modulator sites while exposing the limitations of balanced cross-method assessment for large potassium-channel assemblies. Broad receptor recovery did not guarantee an accurate ligand-binding mode. By distinguishing these failure modes, the benchmark provides a foundation for evaluating future methodological progress.