acceptodds
Under review as a conference paper at ICLR 2027

Beyond Static Pockets: Benchmarking and Modeling Multi-State Transporter Drug Discovery

Abstract

Transporter drug discovery requires consideration of multiple conformations of the same protein, yet existing computational methods and models still struggle to rapidly and accurately identify ligand-preferred conformations. We introduce SLCState, a manually curated benchmark, and SLCHier, a model for conformational recognition. SLCState contains 762 pocket–ligand records from 622 PDB entries, covering 109 proteins, 14 structural folds, and five conformational states. Seven data splits assess generalization to unseen proteins, new ligand associations, and increasingly complex state sets. We also propose SLCHier, which uses the transporter structural hierarchy as a prior and represents it in hyperbolic space through a two-scale representation that preserves state-level pocket embeddings while organizing them with transporter-level centers. Trained separately on each split, SLCHier achieves higher top-1 accuracy than the compared zero-shot baselines in all seven settings. Controlled ablations further confirm the importance of the structural hierarchy and hyperbolic representations. Together, SLCState and SLCHier provide an evaluation framework and modeling approach for multi-state transporter drug discovery.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
membrane transporters, conformational pharmacology, multistate virtual screening, protein--ligand retrieval, hyperbolic representation learning, drug discovery

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_EsWkOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces SLCState, a manually curated benchmark for ligand-conditioned conformational-state ranking in solute carrier (SLC) transporters, and SLCHier, a hierarchy-aware hyperbolic representation learning model. SLCState contains 762 pocket–ligand records from 622 PDB entries, covering 109 transporters, 14 structural folds, and five canonical conformational states. Seven evaluation splits target held-out proteins, held-out ligand–transporter associations, and different conformational-state configurations. SLCHier combines state-level pocket–ligand similarity with transporter-level structural hierarchy supervision through Lorentz-space representations and a two-scale architecture. It achieves the highest top-1 accuracy among the reported methods in all seven settings.

Critical strengths
  1. The paper addresses a practically relevant and underexplored problem in transporter drug discovery: ligand-conditioned conformational-state ranking within the same protein. The manually curated SLCState benchmark, covering 109 transporters and five conformational states, provides a useful resource for studying this problem.

  2. The proposed SLCHier framework incorporates transporter structural relationships while preserving state-specific pocket representations. The ablation studies provide evidence that structural hierarchy supervision and hyperbolic representations contribute to performance, particularly under limited same-fold training support.

Critical weaknesses
  1. The experimental comparison is not entirely fair. SLCHier is trained specifically on SLCState, whereas all competing methods are evaluated zero-shot. Therefore, the reported improvements may partly reflect task-specific supervision rather than the superiority of the proposed architecture.

  2. The definition of ligand-preferred conformations is not sufficiently justified. Observing a ligand bound to a particular conformation does not necessarily establish its preference over other conformations. Treating unobserved states as competing negatives may introduce misleading supervision and weaken the biological interpretation of the results.

  3. The dataset is relatively small and imbalanced across conformational states, raising concerns about the robustness and generalizability of the approach.

  4. The methodological novelty is somewhat limited. Hyperbolic representation learning, hierarchical structural supervision, and contrastive learning have been extensively explored. The proposed framework mainly combines these established techniques, and its broader methodological contribution beyond transporter-specific applications remains unclear.

Questions for authors
  1. For each of the seven splits, how many unique held-out proteins, ligand–transporter pairs, and test records are evaluated? How many of these queries have exactly one candidate state?
  2. How often is the same ligand experimentally associated with more than one canonical state of the same transporter? Are those associations treated as mutually exclusive during state-ranking training?
  3. What is the distribution of the number of pocket representatives per state, and how sensitive are results to fixing this count?
  4. Why are all six external methods evaluated exclusively zero-shot, despite the availability of training data and the use of task-specific supervision for SLCHier?
6 comments
AuthorsOpenReview ↗

Response to Reviewer EsWk(1/3)

W1/Q1/Q4. Fairness of Baseline Comparisons and Additional Task-Adapted Evaluation

We thank the reviewer for raising this important concern and apologize for not sufficiently clarifying the purpose of our original comparisons.

(1) Clarification of the original evaluation. The original comparisons assess zero-shot transfer from general protein–ligand models to conformational-state recognition: the six baselines use released pretrained checkpoints without SLCState adaptation, whereas SLCHier is trained on the corresponding SLCState-Exp partitions. This is a fair and meaningful comparison under the intended task-transfer setting, as all methods rank identical held-out queries and candidate states without test-label adaptation. This evaluation examines whether SLCState-specific learning can generalize to unseen transporters beyond direct transfer of established models. Notably, the original SLCHier uses a randomly initialized ProFSA-based pocket encoder without pretrained ProFSA weights.

Test-set composition (Q1). The table reports mean counts across three valid assignments per protocol. No test query has only one candidate state. Detailed split statistics are provided in Appendix A.9.

ProtocolTest proteinsFully held-out proteinsTest ligand–transporter pairsTest records
HS884576
LS3.673.671672
LH14.67021.6776
2S9.339.3344.6776
IO7742.6775.33
IOF644175
MS5544.6776

Mean counts across three assignments; fully held-out proteins are absent from training. In LH, complete ligand–transporter pairs rather than proteins are held out; IOF includes both fully held-out and other test transporters.

(2) Additional task-adapted comparisons. To address the supervision mismatch, we adapted all six baselines on the same SLCState-Exp training partitions across three assignments for each of the seven protocols. They were initialized from official pretrained checkpoints and optimized with a shared state-ranking objective, head warm-up, and partial encoder fine-tuning. Checkpoints were chosen exclusively by validation Top-1 accuracy (MRR as a tie-breaker), with the original pretrained checkpoint included as a no-update option. The no-update option is determined entirely by validation, never test performance, guarding against detrimental updates. We also evaluated a separate SLCHier variant initialized from pretrained ProFSA weights.

MetricHSLSLH2SIOIOFMS
Top-1 (%)69.07 ± 3.2555.86 ± 0.4752.17 ± 2.9978.87 ± 7.1179.70 ± 2.9673.23 ± 4.9855.04 ± 3.75
MRR.8379 ± .0209.7448 ± .0086.7310 ± .0172.8944 ± .0356.8985 ± .0148.8582 ± .0316.7319 ± .0198

ProFSA-pretrained SLCHier; mean ± SD across three train/test assignments.

SLCHier retains the highest mean Top-1 accuracy and MRR across all seven protocols, outperforming the strongest task-adapted baseline by 3.62–16.18 percentage points in mean Top-1 accuracy, with particularly substantial gains on IOF (+16.18 points), LS (+8.99), and LH (+7.77). The full comparisons and adaptation procedures are reported in Appendix E.

(3) Fine-tuning behavior and task-specific modeling. General pretrained protein–ligand models do not consistently benefit from additional adaptation: validation selects epoch 0 in 15/21 assignments for DrugCLIP and TANKBind and 21/21 for RTMScore, compared with only 3/21 for BindNet. This likely reflects both a mismatch between general binding-compatibility objectives and within-transporter conformational discrimination and the limited number of independent transporters with correlated training records. These heterogeneous outcomes indicate that some models benefit from adaptation while others lose transferable performance.

In contrast, SLCHier integrates state-specific representations, ranking supervision, state-aware negatives, and transporter-level hyperbolic hierarchy modeling. State-level ranking separates competing conformations within a transporter, while hyperbolic centers encode structural hierarchy for knowledge sharing across proteins. Its continued advantage under task-adapted comparisons, together with component-wise ablations, supports the combined effectiveness of these task-aligned components.

Together, the original zero-shot and task-adapted evaluations provide complementary evidence for SLCHier while directly addressing the supervision difference raised by the reviewer.

AuthorsOpenReview ↗

Response to Reviewer EsWk(2/3)

W2. Definition of Ligand–Conformation Associations and Negative Supervision

Regarding the definition of ligand-preferred conformations, we agree that an experimentally observed ligand-bound conformation does not necessarily establish thermodynamic preference or exclude binding to other states. The supervision target of SLCState is an experimentally supported ligand–conformation association, verified through original publications, structural inspection, and chain-level annotations. The evaluation measures whether models rank the experimentally observed state highly among candidate conformations. Thermodynamic preferences, equilibrium state occupancies, and functional selectivity require corresponding experimental evidence, and unobserved states should not be interpreted as experimentally verified nonbinding states.

Given that the same substrate may remain bound during multiple stages of the transport cycle, our supplementary computational candidate poses also help explore ligand–conformation combinations not yet captured experimentally, partially addressing gaps in experimental sampling without treating these candidates as experimental evidence. Furthermore, as described in Appendix C, we construct presumed background negatives by first excluding all known positive ligands in the current training partition from the FDA-approved drug library, including those with and without resolved bound conformations, and then sampling from the remaining compounds. This ensures that these known binders are not included in the background drug-negative pool merely because they lack conformation-resolved experimental structures.

Q2. Multistate Ligand Associations and Their Treatment in Training

We analyzed the 762 experimental pocket–ligand records in SLCState. After grouping records by transporter and standardized ligand structure and merging repeated observations within the same canonical state, we identified 412 distinct transporter–ligand pairs. Of these, 59 pairs (14.3%) were experimentally observed in at least two canonical states, involving 29 transporters. A substrate can remain bound as its transporter progresses through multiple substrate-bound conformations, while available experimental structures may capture only a subset of these states.

After adding the 2,043 computational records described above, the expanded database contains 434 transporter–ligand pairs, of which 90 pairs (20.7%) cover at least two canonical states, involving 41 transporters. Specifically, the computational expansion provides candidate binding poses in additional conformations for 31 pairs previously represented by only one experimentally observed binding state. Using the original 412 pairs as a fixed denominator, multistate coverage increases from 14.3% to 21.8% (90/412).

Regarding mutual exclusivity during training, multi-positive pocket–ligand alignment retains all known positive pockets for each ligand. When a ligand has experimental binding records in both inward-facing and outward-facing states, both states are treated as positives in the alignment objective. State ranking encourages the experimentally observed state in each record to receive a higher score, so both known binding states receive positive supervision while also competing with each other. Together, these supervision signals allow the same ligand to receive high compatibility scores with multiple states. If two experimentally supported states both receive high and similar scores, this indicates strong model support for compatibility with both binding states, consistent with a substrate remaining bound across multiple transporter conformations.

AuthorsOpenReview ↗

Response to Reviewer EsWk(3/3)

W3/Q3. Dataset Coverage and Pocket Representation

We thank the reviewer for raising concerns about dataset coverage, conformational-state imbalance, and variation in the number of pocket representatives. Ligand-conditioned conformational recognition requires experimentally supported associations among ligands, specific protein chains, binding pockets, and conformational states. Through systematic structural retrieval and literature-guided curation, SLCState-Exp provides 762 experimental pocket–ligand records spanning 109 transporters and 14 structural folds, including 47 proteins with at least two experimentally resolved states and 22 with at least three. Among the public resources we examined, SLCState-Exp provides the largest collection of experimental records and the broadest fold coverage for this specific ranking task.

We nevertheless acknowledge that the experimental collection provides limited coverage across transporters and conformational states. In response to the reviewer’s suggestion, we introduce SLCState-Aug, which supplements the experimental collection with computational candidates generated using Glide docking and BioEmu2, providing additional binding configurations for 55 of the original 109 proteins.

Regarding conformational-state imbalance, the expansion adds 55 inward-occluded candidates to the original 23 experimental records. This state remains underrepresented, partly reflecting differences in how readily transporter states can be captured experimentally. Despite this imbalance, SLCHier outperforms all zero-shot baselines in mean accuracy and MRR on the two-state ranking task (2S; Table 2).

Table X. Distribution of conformational states in SLCState-Exp and SLCState-Aug.

Conformational stateSLCState-Exp, n (%)SLCState-Aug, n (%)
Inward-facing297 (39.0%)1,288 (46.5%)
Inward-occluded23 (3.0%)78 (2.8%)
Occluded115 (15.1%)290 (10.5%)
Outward-occluded58 (7.6%)167 (6.0%)
Outward-facing269 (35.3%)946 (34.2%)

Pocket-representative counts also vary across transporter–state groups. Among 188 groups, the mean is 4.04, the median is 2, the interquartile range is 1–5, and the range is 1–25; 26.1% have only one representative. Requiring K>1K>1 would therefore exclude part of the benchmark, whereas using K=1K=1 throughout would discard substantial conformational information.

We use max aggregation under a multiple-instance interpretation: a ligand is considered compatible with a state if at least one representative supports the interaction. To assess sensitivity to representative count, we constructed balanced-KK cohorts requiring every candidate state to have at least two or three distinct representatives. Across 100 label-independent repetitions, representatives within each state were randomly ordered, and the same queries were evaluated using the first KK representatives with max aggregation.

Table Y. Sensitivity of MRR to balanced pocket-representative count.

Minimum representatives per candidate stateQuery instancesSplitsK=1K=1K=2K=2K=3K=3
2997210.7217 ± 0.01920.7335 ± 0.0169—
3638180.7201 ± 0.02600.7282 ± 0.02350.7383 ± 0.0233

The reported deviations reflect variability across the 100 selections. MRR increased by 0.0118 from K=1K=1 to K=2K=2 in the first cohort and by 0.0182 from K=1K=1 to K=3K=3 in the second; Top-1 accuracy followed the same qualitative trend. Ranking performance is therefore relatively stable within the tested balanced-count range, with modest benefits from additional representatives. We will clarify these limitations and provide detailed curation, expansion, and sampling procedures in the revised manuscript and appendix.

Reviewer_JjzKOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work addresses the challenge of protein conformations in transporter drug discovery. To this end, the authors introduce SLCState, a benchmark for conformational state recognition, and propose SLCHier, a model that incorporates the structural hierarchy of transporters as a prior. When trained separately on each data split, SLCHier achieves higher top-1 accuracy than the evaluated zero-shot baselines across all seven settings.

Critical strengths
  1. The research problem is interesting: transporter drug discovery requires accounting for multiple conformations of the same protein, yet existing computational methods still struggle to efficiently and accurately identify ligand-preferred conformations.
  2. The paper writing is clear and easy to follow.
Critical weaknesses
  1. Is this primarily a benchmark paper or a methodological paper? The two contributions seem somewhat orthogonal to me. While the paper introduces both SLCState and SLCHier, combining these two directions may limit the depth of each contribution. I am concerned that neither contribution is explored in sufficient depth for an ICLR paper.

  2. Are there any established or widely used benchmarks for this task prior to SLCState? If so, why is SLCHier not evaluated on these existing benchmarks? Such comparisons would help establish the generalizability of the proposed method beyond the newly introduced SLCState benchmark.

Questions for authors

Please see Weakness

2 comments
AuthorsOpenReview ↗

Response to Reviewer JjzK(1/2)

We sincerely thank the reviewer for the thoughtful evaluation and for recognizing the importance of our scientific problem and the clarity of our writing. We appreciate the constructive concerns about our paper's focus and benchmark coverage, which help us better articulate our contributions.

W1. Clarification of the Paper's Focus and the Relationship Between SLCState and SLCHier

We thank the reviewer for raising this important point and apologize for not sufficiently clarifying the relationship between our contributions. We would like to emphasize that SLCState and SLCHier are not two orthogonal research directions, but two closely connected components developed to address the same scientific problem. Our work is primarily centered on formulating and systematically investigating ligand-conditioned conformational-state recognition, with SLCState providing the experimental foundation and SLCHier serving as a task-specific modeling approach that demonstrates the learnability of this problem.

(1) A scientifically important problem. We formulate ligand-conditioned conformational-state ranking, which asks a model to distinguish among different conformational states of the same transporter for a given ligand. This problem is fundamental to transporter drug discovery because conformational transitions underlie transporter function, and ligands recognizing or stabilizing distinct states can exhibit different inhibitory mechanisms and pharmacological effects. Identifying ligand-associated conformations is therefore important for understanding drug mechanisms and developing conformation-selective therapeutics. More broadly, this challenge is relevant to other conformationally dynamic drug targets and allosteric regulatory mechanisms.

(2) SLCState provides the experimental foundation for systematic investigation. Addressing this problem requires experimentally supported associations among ligands, binding pockets, specific protein chains, and conformational states, which are largely absent from existing resources. Through extensive manual curation and verification against original structural studies, we constructed SLCState-Exp with 762 experimental pocket–ligand records from 622 PDB entries, covering 109 proteins, 14 structural folds, and five conformational states. Importantly, 47 proteins have at least two experimentally resolved states, and 22 have at least three, enabling meaningful within-protein comparisons.

Beyond data curation, we establish seven complementary evaluation protocols addressing two fundamental generalization dimensions: protein and ligand generalization (HS, LS, LH), which examines transfer across proteins and ligand–transporter associations under different structural-support conditions; and conformational state-space generalization (2S, IO, IOF, MS), which evaluates recognition across different state combinations, including multi-state discrimination on unseen transporters. We further introduce SLCState-Aug, extending the experimental collection with 2,043 computational candidate records to 2,805 records in total, with experimental and computational provenance explicitly distinguished. Together, these efforts provide a systematic experimental and computational resource for investigating ligand-conditioned conformational recognition (see Appendix A.6 for details).

(3) A task-specific model demonstrating benchmark learnability. Rather than pursuing an independent modeling objective, SLCHier specifically employs two-scale hyperbolic representations to address the hierarchical nature of conformational recognition. State-level pocket representations, ranking supervision, and state-aware negatives distinguish competing conformations within the same transporter, while transporter-level hyperbolic centers and structural hierarchy supervision facilitate knowledge sharing across proteins.

Importantly, SLCHier also demonstrates the learnability of SLCState's experimental supervision. We clarify that its original ProFSA-based pocket encoder is randomly initialized without pretrained ProFSA weights, whereas its ligand encoder is initialized from pretrained Uni-Mol. Thus, meaningful state-specific pocket representations can be learned directly from SLCState supervision.

Our experiments further support this task-specific design. Across three train/test assignments per protocol, SLCHier achieves the highest mean Top-1 accuracy and MRR on all seven protocols, while controlled ablations validate its key components. Furthermore, a ProFSA-pretrained SLCHier variant retains the highest mean Top-1 accuracy and MRR even when all six baselines receive the same task-specific training data (see Appendix E for more details).

In summary, SLCState establishes the scientific task and its evaluation framework, while SLCHier demonstrates benchmark learnability through task-specific geometric modeling.

AuthorsOpenReview ↗

Response to Reviewer JjzK(2/2)

W2. Existing Benchmarks and External Generalization

We thank the reviewer for raising this important question. To our knowledge, there is currently no established independent benchmark specifically designed for ligand-conditioned conformational-state ranking across SLC transporters. This task requires experimentally supported associations among ligand identity, binding pockets, specific protein chains, and conformational states. Ligand-free structures can provide candidate conformations, but cannot establish ligand-associated states, while protein-level activity measurements do not directly identify the bound conformation.

Existing resources primarily address different scientific problems and have limited coverage for our task:

  • IOMemP contains 16 inward-/outward-facing structural pairs (32 experimental structures), primarily supporting alternative-conformation prediction and conformational sampling rather than ligand-conditioned recognition.
  • They All Rock collects 298 structural chains from 207 PDB entries, covering 28 LeuT-fold proteins, with only nine proteins used for multi-state comparisons. It focuses on conformational motions and is restricted to a single structural fold.
  • Pocketome v17.12 is a general-purpose binding-pocket resource containing only three binding sites from three human SLC transporters, without sufficient multi-state coverage.

While valuable for their respective purposes, these resources do not provide the ligand–pocket–chain–state associations required for our ranking task. This gap directly motivates the construction of SLCState.

To establish a dedicated benchmark, we systematically retrieved and manually examined experimental SLC structures, initially collecting 926 PDB entries and 962 PDB entry–chain combinations. Through literature-guided curation, conformational annotation, ligand verification, and pocket extraction, we constructed SLCState-Exp with 762 experimental pocket–ligand records from 622 PDB entries, spanning 109 transporters, 14 structural folds, and five conformational states. Importantly, 47 transporters have at least two experimentally resolved states, and 22 have at least three, enabling meaningful within-protein comparisons. To our knowledge, SLCState-Exp is the largest and most comprehensive experimental benchmark specifically supporting ligand-conditioned conformational-state ranking.

Beyond dataset construction, we establish seven complementary evaluation protocols covering unseen proteins, unseen ligand–transporter associations, different levels of structural-fold support, and increasingly complex conformational-state configurations. We further introduce SLCState-Aug, adding 2,043 computational candidate records generated through Glide docking (1,688) and BioEmu2 (355), expanding the collection to 2,805 records. Experimental observations and computational candidates remain explicitly distinguished, and additional evaluations on SLCState-Aug have been incorporated into the revised manuscript.

We systematically assess cross-transporter generalization through strict protein-holdout protocols, including MS with at least three candidate states, and three independently constructed train/test assignments per protocol. SLCHier retains the highest mean Top-1 accuracy and MRR across all seven protocols, supporting generalization to unseen transporters and ligand associations within SLCState.

Reviewer_9KFXOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper presents SLCState, a benchmark for multi-state transporter drug discovery, and SLCHier, a model for conformational recognition. The benchmark provides experimental structures, ligand data, annotations, and seven evaluation protocols. SLCHier encodes the transporter hierarchy in hyperbolic space, preserving state-level pocket embeddings around transporter-level centers

Critical strengths
  1. A scientifically meaningful problem formulation. The paper emphasizes that ligand binding depends not only on protein identity but also on the protein's conformational state. This is an important perspective for transporter drug discovery, where conformational changes can alter the geometry and chemical environment of binding pockets. More broadly, the same scientific question arises in allosteric pockets and other conformationally dynamic drug targets. Explicitly formulating ligand-conditioned conformational recognition as a distinct evaluation problem is a meaningful contribution.

  2. Considerable effort in benchmark construction. The authors curate experimentally resolved structures, ligand information, and conformational annotations to construct SLCState. Collecting and organizing these heterogeneous data requires substantial effort and provides a potentially valuable resource for evaluating conformation-aware ligand recognition.

  3. A dedicated model for multi-conformational ligand recognition. The authors propose SLCHier to evaluate pocket-ligand compatibility across multiple conformational states. Its hierarchical representation is designed to capture both state-specific pocket–ligand relationships and broader structural relationships among transporters. This design is aligned with the paper's central problem and offers a concrete modeling approach beyond simply evaluating ligand binding against a single static pocket.

  4. Broad experimental evaluation and informative ablations. The seven splitting protocols cover complementary generalization settings. The ablation studies further validate the contributions of key components.

Critical weaknesses
  1. Limited methodological novelty. SLCHier combines pretrained ProFSA and Uni-Mol representations with hyperbolic geometry and hierarchical supervision. However, hyperbolic protein–ligand representations have been explored in Learning Protein–Ligand Binding in Hyperbolic Space. The paper insufficiently clarifies its methodological advances beyond combining existing techniques, warranting more direct comparisons and clearer contributions.

  2. Limited dataset scale, imbalanced state distribution, and potential label ambiguity. SLCState contains 762 pocket–ligand records from 622 PDB entries, covering 109 proteins and five conformational states. The dataset is relatively small, and its state distribution is highly imbalanced, with inward-occluded structures accounting for only 3.0% of the records. Moreover, assigning experimentally observed ligand-associated conformations as ground-truth labels may not reflect true thermodynamic preference or functional selectivity, as structural observations depend on experimental conditions and sampling. Thus, label validity requires further justification.

  3. Insufficient analysis of split construction and robustness. The seven splitting protocols evaluate complementary aspects of generalization, but each protocol can potentially yield multiple valid train/test assignments. The paper does not sufficiently explain the rationale for selecting the specific assignments used in the experiments, nor does it report enough information to assess how sensitive the results are to alternative valid splits. Without repeated splits or an analysis of variability across different assignments, it is difficult to determine whether the reported improvements are robust or depend on particular choices of held-out proteins and ligand–transporter pairs.

  4. Insufficient details on validation subset construction. Section 7.1 states that hyperparameters and checkpoints are selected using a validation subset drawn exclusively from the corresponding training data, after which the selected checkpoint is frozen for test evaluation. However, the paper does not clearly specify how this validation subset is constructed.

  5. Potentially unfair baseline comparisons and limited external validation. SLCHier is trained on SLCState training splits, whereas baselines appear to be evaluated zero-shot, complicating fair comparisons. External validation is also needed to establish generalizability beyond the proposed benchmark and evaluation protocol.

Questions for authors
  1. Methodological novelty. How does SLCHier advance beyond prior hyperbolic protein-ligand learning? Can controlled experiments isolate the contributions of its hierarchical design and hyperbolic geometry from pretrained encoders and existing methods?

  2. Data splitting and validation protocol. Under each splitting protocol, how sensitive are the results to different valid split assignments, particularly given the highly imbalanced conformational state distribution? How was the validation set constructed, and were the corresponding splitting constraints consistently maintained?

  3. External generalization. Can SLCHier be evaluated on independent transporter datasets to assess generalization beyond SLCState?

  4. Fairness of baseline comparisons. Why is SLCHier trained on SLCState while baselines are evaluated zero-shot? Can comparisons use consistent supervision, and how do the authors ensure the benchmark does not favor SLCHier?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 9KFX(1/3)

W1/Q1. Clarification of the Core Contributions and Methodological Advances

We thank the reviewer for recognizing the scientific importance of our problem and for highlighting the need to better clarify our methodological contributions. We agree that hyperbolic representations and contrastive learning have been explored in prior work. However, our central contribution lies in formulating a previously underexplored scientific problem and establishing its dedicated benchmark, while SLCHier provides a task-specific geometric modeling approach and an initial demonstration of the benchmark's learnability.

(1) A new scientific problem. To our knowledge, this work is the first to systematically formulate ligand-conditioned conformational-state ranking for SLC transporters as a distinct learning and evaluation task. Unlike conventional protein–ligand retrieval or affinity prediction, our task fixes the transporter and ligand while distinguishing among different conformations of the same protein. This problem is fundamental to transporter drug discovery because conformational transitions underlie transporter function, and ligands recognizing or stabilizing different states can exhibit distinct inhibitory mechanisms and pharmacological effects. More broadly, conformationally dependent ligand recognition is relevant to other dynamic drug targets and allosteric regulatory mechanisms.

(2) A dedicated benchmark enabling systematic investigation. This task requires experimentally supported associations among ligands, binding pockets, specific protein chains, and conformational states, which are largely absent from existing resources. Previous transporter datasets are relatively limited and primarily address different problems. To our knowledge, SLCState-Exp is the largest and most comprehensive experimental benchmark specifically supporting ligand-conditioned conformational-state ranking, comprising 762 curated records from 622 PDB entries, covering 109 proteins, 14 structural folds, and five conformational states. Importantly, 47 proteins have at least two experimentally resolved states, and 22 have at least three. We further establish seven complementary generalization protocols and introduce SLCState-Aug, extending the collection to 2,805 records through explicitly distinguished computational candidates(see Appendix A.6 for details).

(3) Task-specific modeling and evidence of benchmark learnability. To address this problem, SLCHier adopts a two-scale hyperbolic representation that explicitly integrates within-transporter conformational discrimination with cross-transporter structural hierarchy. State-level representations and ranking supervision distinguish competing conformations of the same transporter, while transporter-level hyperbolic centers and structural hierarchy supervision facilitate knowledge sharing across proteins.

Across all four state-space evaluation protocols (2S, IO, IOF, and MS), SLCHier consistently outperforms the strongest zero-shot baselines in Top-1 accuracy, including the challenging MS setting involving unseen proteins with at least three candidate conformations. Together with our ablation studies, these results support the effectiveness of our task-specific modeling design and, importantly, demonstrate that SLCState provides meaningful supervision for learning conformation-aware pocket representations, establishing the learnability of this newly formulated task.

We also clarify that the original SLCHier uses a randomly initialized ProFSA-based pocket encoder without pretrained ProFSA weights, whereas its ligand encoder is initialized from pretrained Uni-Mol. Thus, state-specific pocket representations are learned directly from SLCState supervision rather than inherited ProFSA representations. We have clarified this initialization in Sections 6.2 and 7.1 of the revised manuscript.

Our controlled ablations directly examine the contributions of the key modeling components. On the original LS assignment, removing state-ranking supervision reduces Top-1 accuracy from 52.11% to 46.48%, replacing Lorentz geometry with a matched Euclidean counterpart reduces it to 29.58%, removing structural distance/order supervision reduces it to 33.80%, and shuffling the structural hierarchy reduces it to 21.13%(see Section 7.4 and Table 5). Together, these experiments support the distinct contributions of state-ranking supervision, hyperbolic geometry, and meaningful structural hierarchy, while demonstrating that SLCState provides meaningful supervision for learning conformation-aware pocket representations.

In summary, our primary contribution is the establishment of a new scientific task and its dedicated benchmark, complemented by SLCHier as a task-specific geometric framework demonstrating the benchmark's learnability. We will clarify this contribution structure in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer 9KFX(2/3)

W2. Dataset Scale, State Imbalance, and Label Validity

We thank the reviewer for highlighting these concerns. We emphasize that ligand-conditioned conformational recognition requires experimentally supported associations among ligands, specific protein chains, binding pockets, and conformational states. Through systematic structural retrieval and literature-guided curation, SLCState-Exp provides 762 experimental pocket–ligand records spanning 109 transporters and 14 structural folds, including 47 proteins with at least two experimentally resolved states and 22 with at least three. Among the public resources we examined, SLCState-Exp provides the largest collection of experimental records and the broadest fold coverage for this specific ranking task.

In response to the reviewer’s suggestion, and to further broaden substrate and binding-pose coverage, we introduce SLCState-Aug, which supplements the experimental collection with 2,043 computational candidate records generated using Glide docking (1,688 records) and BioEmu2 (355 records). The resulting collection contains 2,805 records, covering additional binding configurations for 55 of the original 109 proteins. Experimental observations and computational candidates are explicitly distinguished (see Appendix A.6 for details).

Regarding the distribution of conformational-state data, the expansion also adds 55 inward-occluded candidates to the original 23 experimental records. We acknowledge that this rare state remains underrepresented, reflecting limitations in available experimental structural coverage. One contributing factor is that transporters dynamically transition among multiple states, which differ in how readily they can be enriched and captured under experimental conditions. Despite this limitation, SLCHier achieves higher mean accuracy and mean reciprocal rank (MRR) than all zero-shot baselines on the two-state ranking task (2S) in Table 2.

Regarding label validity, we agree that an experimentally observed ligand-bound conformation does not necessarily establish thermodynamic preference or exclude binding to other states. The supervision target of SLCState is an experimentally supported ligand–conformation association, verified through original publications, structural inspection, and chain-level annotations. The evaluation measures whether models rank the experimentally observed state highly among candidate conformations. Thermodynamic preferences, equilibrium state occupancies, and functional selectivity require corresponding experimental evidence, and unobserved states should not be interpreted as experimentally verified nonbinding states.

Given that the same substrate may remain bound during multiple stages of the transport cycle, our supplementary computational candidate poses also help explore ligand–conformation combinations not yet captured experimentally, partially addressing gaps in experimental sampling without treating these candidates as experimental evidence.

Furthermore, as described in Appendix C, we construct presumed background negatives by first excluding all known positive ligands in the current training partition from the FDA-approved drug library, including those with and without resolved bound conformations, and then sampling from the remaining compounds. This ensures that these known binders are not included in the background drug-negative pool merely because they lack conformation-resolved experimental structures.

We will clarify these distinctions in the revised manuscript, with detailed curation, computational expansion, and negative-construction procedures provided in the appendix.

W3/Q2. Split Construction and Robustness Across Multiple Assignments

We thank the reviewer for raising this important concern. To assess sensitivity to train/test partitioning, we randomly generated two additional valid assignments for each of the seven protocols, yielding three assignments per protocol. All assignments were constructed under the original biological holdout constraints and audited for protocol compliance.

The additional assignments cover complementary biological characteristics, including different levels of same-fold support (HS/LS), held-out ligand–transporter associations (LH), different two-state combinations (2S), inward/outward state proportions (IO/IOF), and proteins with at least three documented states (MS).

We re-evaluated all methods across the three assignments and report mean ± standard deviation for each protocol. The complete results for SLCHier and all six baselines are presented in Tables 1 and 2 of the revised manuscript, with detailed split statistics and biological coverage analyses provided in Appendix A.11. SLCHier retains the highest mean Top-1 accuracy and MRR across all seven protocols, with mean Top-1 improvements of 3.63–14.43 percentage points over the strongest baseline in each protocol.

AuthorsOpenReview ↗

Response to Reviewer 9KFX(3/3)

W4. Clarification of Validation Subset Construction

We apologize for the insufficient description.

Validation subsets are drawn only from the corresponding outer training partitions, without using test data. Using a fixed seed, we stratify by query counts, structural folds, and state counts, targeting approximately 15% of eligible training queries.

The original biological holdout units are preserved: entire transporters for six protein-holdout protocols and complete ligand–transporter pairs for LH; IOF further limits validation to inward-/outward-facing queries of held-out transporters. All 21 validation splits passed leakage and protocol-compliance audits, with disjoint holdout units and all candidate states of each query kept together. Checkpoints are selected by validation Top-1 accuracy (MRR as a tie-breaker). Full construction and audit details will be provided in Appendix A.11.

W5/Q4. Fairness of Baseline Comparisons and Additional Task-Adapted Evaluation

We thank the reviewer for identifying this concern and apologize for the unclear comparison setting.

(1) Original zero-shot evaluation. Our original experiments assess zero-shot transfer from general protein–ligand models to conformational-state recognition. The six baselines use released pretrained checkpoints without SLCState adaptation, whereas SLCHier is trained on the corresponding SLCState-Exp partitions. All methods rank the same held-out candidate states without test-label adaptation. Notably, the original SLCHier uses a randomly initialized ProFSA-based pocket encoder without pretrained ProFSA weights, while its ligand encoder uses pretrained Uni-Mol. This is a meaningful task-transfer comparison, but it does not isolate architectural advantages under matched task-specific supervision.

(2) Additional task-adapted comparisons. To address this limitation, we adapted all six baselines on the same SLCState-Exp training partitions across three assignments per protocol. Starting from official pretrained checkpoints, each baseline uses a shared state-ranking cross-entropy objective with head warm-up and partial encoder fine-tuning. Checkpoints are selected solely by validation Top-1 accuracy (MRR as a tie-breaker), and the original pretrained checkpoint (epoch 0) is included as a predefined no-update option to avoid detrimental updates. We also evaluate a separate SLCHier variant initialized from pretrained ProFSA weights.

SLCHier retains the highest mean Top-1 accuracy and MRR across all seven protocols even when all baselines are offered the same task-specific training data. Thus, the advantage extends beyond access to task-specific supervision. Complete results and adaptation procedures are provided in Appendix E.

(3) Analysis of fine-tuning behavior. General pretrained protein–ligand models do not consistently benefit from further adaptation: validation often retains the original checkpoint for RTMScore, DrugCLIP, and TANKBind, while BindNet more frequently selects an updated checkpoint. These heterogeneous outcomes suggest that task adaptation is challenging rather than universally ineffective. Two plausible factors are the mismatch between general binding-compatibility pretraining and fine-grained within-transporter state ranking, and the limited number of independent transporters with correlated training records. This does not imply that pretraining itself is ineffective.

In contrast, SLCHier combines state-specific ranking supervision and state-aware negatives with transporter-level hyperbolic structural hierarchy learning, explicitly addressing both within-transporter conformational discrimination and cross-transporter knowledge sharing. These results and controlled ablations support the combined effectiveness of these task-specific components. The zero-shot and task-adapted comparisons provide complementary evidence while addressing the supervision mismatch.

Q3. External Generalization Beyond SLCState

We thank the reviewer for emphasizing the importance of external validation.

To our knowledge, no established independent benchmark supports ligand-conditioned conformational-state ranking across SLC transporters. IOMemP addresses conformational sampling, They All Rock examines LeuT-fold dynamics, and Pocketome collects general binding pockets; none provides the required ligand–pocket–chain–state associations.

We acknowledge the lack of independent experimental validation as a limitation. Instead, our current evidence covers cross-transporter generalization within SLCState through strict protein holdouts, multiple train/test assignments, and varied state configurations. SLCState-Aug is a computational expansion, not an independent experimental benchmark. We will clarify this scope in the revision.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Comment: Clarification of Contributions and Summary of Revisions

We sincerely thank all reviewers for their thoughtful evaluations and constructive suggestions. In response to concerns about the paper's focus, dataset coverage, robustness, and comparison fairness, we have clarified our central contribution and substantially expanded the benchmark and evaluation.

(1) Clarifying the unified scientific contribution. We address a scientifically important problem in transporter drug discovery: ligand-conditioned conformational-state recognition. Conformational transitions underlie transporter function, and ligands recognizing different states can have distinct pharmacological effects. The task ranks states for a fixed transporter–ligand pair. SLCState and SLCHier are complementary, not orthogonal, contributions to this scientific problem.

Our primary contribution is formulating this scientific problem and establishing SLCState, its dedicated benchmark. Its experimentally curated collection, SLCState-Exp, is, to our knowledge, the largest and most comprehensive experimental benchmark for this task. It contains 762 manually curated pocket–ligand records from 622 PDB entries, covering 109 transporters, 14 structural folds, and five conformational states; 47 transporters have at least two experimentally resolved states and 22 have at least three. Existing resources lack the required ligand–pocket–chain–state associations. We also establish seven evaluation protocols spanning protein/ligand generalization (HS, LS, LH) and state-space generalization (2S, IO, IOF, MS).

SLCHier is a task-specific model demonstrating the benchmark's learnability. It integrates state-level conformational discrimination with transporter-level structural hierarchy through two-scale hyperbolic representations. Importantly, the original SLCHier's ProFSA-based pocket encoder is randomly initialized without pretrained ProFSA weights, while its ligand encoder is initialized from pretrained Uni-Mol. Its state-ranking, geometry, and structural-hierarchy ablations further support the role of these modeling components.

(2) Expanding the resource with SLCState-Aug. To address limited experimental coverage and explore additional substrate–conformation configurations, we introduce SLCState-Aug, adding 2,043 computational candidate records: 1,688 generated by Glide docking into experimentally resolved conformations and 355 BioEmu2-derived records retained after structural and chemical screening. Together, these additions provide new candidate configurations for 55 of the original 109 transporters, increasing the full collection from 762 to 2,805 records (Section 5.1; Appendix A.6).

Experimental associations and computational candidates remain explicitly distinguished; the new poses broaden structural exploration but are not presented as experimentally established binding preferences. We also evaluate SLCState-Aug across all seven protocols (Section 7.3).

(3) Expanding robustness evaluation and ensuring matched task supervision. We generated two additional valid train/test assignments for each protocol, yielding three assignments per protocol (21 in total), all satisfying the original biological holdout constraints and passing leakage checks. We now report mean ± standard deviation across assignments, with detailed split and validation procedures in Appendix A.11. SLCHier maintains the highest mean Top-1 accuracy and MRR across all seven SLCState-Exp protocols, improving Top-1 accuracy by 3.63–14.43 percentage points over the strongest zero-shot baseline for each protocol (Tables 1–2).

To address fairness, we clarify that the original baselines were evaluated zero-shot to measure transfer from general-purpose protein–ligand models, whereas SLCHier was trained on SLCState. We agree that this does not isolate architectural benefits under matched supervision. Therefore, we additionally adapted all six baselines on the same SLCState-Exp training partitions across all 21 assignments, using a shared state-ranking objective, conservative partial fine-tuning, and validation-only checkpoint selection. We also evaluated a separate SLCHier variant initialized from pretrained ProFSA weights. SLCHier retains the highest mean Top-1 accuracy and MRR across all seven protocols under task-adapted comparisons, exceeding the strongest adapted baseline by 3.62–16.18 percentage points in Top-1 accuracy (Appendix E). These settings provide complementary evidence of effectiveness.

In summary, our contribution is a unified framework for a previously underexplored scientific problem: SLCState establishes the experimental foundation and evaluation protocols, while SLCHier demonstrates task-specific benchmark learnability. SLCState-Aug, multi-assignment robustness analysis, and task-adapted baselines strengthen the breadth and reliability of our evidence.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.