acceptodds
Under review as a conference paper at ICLR 2027

PharmPD: Coupling Pharmacophore Semantics and 3D Geometry for Molecular Representation

Abstract

Three-dimensional molecular representations capture spatial relationships, but geometry alone does not specify which of those relationships are chemically relevant to molecular recognition. We introduce PharmPD, an explicit representation that couples interaction semantics and three-dimensional geometry through chemistry-audited pharmacophore centers and their spatial relationships. Across 22 TDC ADMET benchmark tasks, adding PharmPD to a strong 2D baseline improves performance on 18 tasks and on all seven prespecified molecular-recognition endpoints. Controlled ablations show that pharmacophore semantics and geometric information are substantially more predictive when explicitly coupled than when provided separately, while three-dimensional distance provides additional value beyond topological separation. The same explicit structure also makes model attributions chemically inspectable: high-attribution features recover target-specific pharmacophore relationships consistent with established recognition chemistry and can be grounded in molecule-specific structural evidence, improving their verifiability. This complementary value persists on an external OpenADMET PXR task, including against established 3D descriptors and a strong baseline incorporating Uni-Mol v2 representations. Together, these results identify semantic–geometric conjunction as a useful design principle for molecular representations that improves prediction while preserving a direct, verifiable connection to molecular-recognition chemistry.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
molecular representation learning, 3D molecular learning, molecular property prediction, geometric representation learning, model interpretability

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.60 over 5 reviews, above 10% of the others.
Reviewer_mBJtOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes PharmPD, a 200-dimensional descriptor coupling pharmacophore types with 3D intercenter distances. Adding PharmPD improves performance on many TDC ADMET tasks

Critical strengths

Explicitly encoding pharmacophore types and their distances incorporates knowledge of molecular recognition into prediction.

Critical weaknesses
  1. The additional gain of 3D geometry over topology is unclear in table 7. And distances are sensitive to conformers, whereas topology remains unchanged across conformers, raising concerns about the robustness of PharmPD.
  2. The additional benefits of PharmPD are unclear. Full and Rank-1 show small differences when added to MapLight. CATS3D-style achieves higher recognition-role recovery on 2/3 targets in Table 2. Table 14 shows diminishing gains with richer baseline features.
Questions for authors
  1. How sensitive is PharmPD to conformers? Does the Topology provide comparable gains when added to MapLight and the PXR baselines?

  2. Under what conditions does PharmPD provide additional predictive or interpretive value?

4 comments
AuthorsOpenReview ↗

W1 / Q1. Additional Value of 3D Geometry and Conformer Robustness

Response.

We thank the reviewer for raising these important questions. We expanded our evaluation with matched topology controls, conformer sensitivity analyses, and an external PXR comparison to clarify the contribution and robustness of PharmPD's geometric representation.

(1) Contribution of 3D information beyond topology.

We reconstructed the topology control to ensure identical pharmacophore centers, chemical assignments, weights, and auxiliary features, isolating the distance representation as the experimental variable.

In the expanded 22-task analysis, appending aligned 3D features to a fixed graph-distance representation outperformed appending row-permuted features on 21/22 tasks, with nine improvements surviving multiple-testing correction (Table 11).

To further control for the potential penalty introduced by shuffled features, we conducted an additional width-matched experiment comparing graph + 3D against graph + a duplicated graph block. The 3D-augmented representation improved performance on 19/22 tasks, with a corrected improvement on Caco-2 (q = 0.033). These experiments provide complementary evidence for useful 3D relational features beyond the tested graph-distance encoding.

(2) Additional benefit above MapLight and external PXR baselines.

In the matched seven-task MapLight augmentation experiment, both representations improved the baseline, with aggregate gains of +0.0168 for PharmPD and +0.0103 for Topology (Table 10).

We additionally evaluated the corrected representations on external PXR data using 513 molecules and 96 analog-series clusters. PharmPD improved RAE over MapLight by +0.01775, compared with +0.00396 for Topology. The direct difference was +0.01380 (95% CI [-0.00013, +0.02837]), indicating a favorable but statistically unresolved comparison.

Together, these results characterize the contribution of 3D encoding across different predictive settings rather than assuming its universal superiority over topology.

(3) Conformer robustness.

PharmPD aggregates pair-distance features across generated conformers. We evaluated the stability of this representation under expanded conformer sampling.

The pooled features exhibited high split-half reliability (r = 0.963 for max pooling; r = 0.986 for mean pooling). In the three-CYP MapLight-augmentation experiment, increasing the conformer count from K = 4 to K = 64 changed aggregate performance by only +0.0021 (95% CI [-0.0009, +0.0049]).

These results indicate that the evaluated predictor is relatively stable across the tested sampling depths, despite variations in individual conformer geometries. PharmPD therefore does not require a uniquely selected molecular conformation to produce its pooled relational representation.

AuthorsOpenReview ↗

W2 / Q2. Conditions Under Which PharmPD Provides Additional Value

Response.

We appreciate the reviewer's question about when PharmPD contributes beyond existing representations. Our expanded experiments distinguish three aspects of this contribution: the predictive value of semantic–geometric correspondence, complementarity with existing molecular representations, and chemically grounded interpretation.

(1) Predictive value of semantic–geometric correspondence.

Our Rank-1 control preserves pharmacophore-type marginals, distance marginals, total feature mass, dimensionality, and auxiliary features while removing their pair-specific correspondence.

Restoring this correspondence improves CYP2C9, CYP2D6, and CYP3A4 inhibition by +0.0367, +0.0678, and +0.0492 AUPRC, respectively (all q = 0.0017; Table 1).

These results establish that retaining which chemical roles occur at which distances contributes predictive information beyond preserving the two marginal distributions separately.

(2) Complementarity depends on the existing representation.

We agree that the Full–Rank-1 difference becomes smaller when both representations are added to MapLight. To investigate this behavior, we conducted a controlled backbone analysis using representations ranging from 0 to 3,092 dimensions.

The Full–Rank-1 advantage showed a consistent negative association with backbone width across all three CYP inhibition endpoints (Spearman ρ = −0.829, q ≤ 0.0007). This indicates that the incremental contribution of explicit correspondence depends on the information already available to the predictor, although backbone composition also changes across comparisons.

Importantly, PharmPD retains measurable value in several tested settings:

  • Against the closest-to-published CATS3D reconstruction (C5L), PharmPD improves standalone aggregate performance by +0.0378 (q = 0.0007; Table 2).
  • In external PXR evaluation, adding PharmPD to Uni-Mol v2 84M + MapLight + multi-fidelity features improves RAE by +0.0067 (95% CI [+0.0029, +0.0107]; Table 22).

These findings help identify the representational settings in which explicit pharmacophore–distance features are most useful.

(3) Chemical interpretation and recognition-role recovery.

We acknowledge the reviewer's observation that CATS3D-style achieves stronger role-level attribution concentration on CYP2D6 and CYP3A4.

Our expanded relation-level analysis provides an additional perspective. Both PharmPD and CATS3D-style recover the prespecified CYP2D6 and CYP3A4 pharmacophore relations within their top three attributed pair types in 100% of model seeds. On hERG, PharmPD achieves 80% top-three recovery versus 30% for CATS3D-style (Table 30).

PharmPD also enables each attributed type–distance feature to be traced to its contributing pharmacophore centers, molecular distances, conformer-specific contributions, and aggregation. Independent reconstruction verified the resulting features for 64/64 held-out molecules (Appendix B.1).

This provides a chemically inspectable connection between model attribution and molecular structure, allowing the identified relations to be examined and computationally verified rather than treated solely as abstract feature indices.

Overall, our findings identify when explicit semantic–geometric correspondence contributes to prediction, how its incremental value changes with the surrounding representation, and how its attributed features can be connected to molecular evidence. We hope these additional analyses clarify the contribution and practical scope of PharmPD, and we appreciate the reviewer's reconsideration of the revised work.

AuthorsOpenReview ↗

We sincerely thank the reviewer for the constructive feedback!

We thank the reviewer for the constructive questions, which have helped us clarify these distinctions and strengthen their experimental evaluation. We would appreciate your assessment of whether the revised manuscript and additional analyses adequately address your concerns, and would welcome any further questions or suggestions.

I thank the authors for their response and additional experiments. They alleviate my concerns, and I have raised my rating to 2.

Reviewer_ag4fOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PharmPD, a multi-dimensional molecular descriptor that explicitly combines pharmacophore semantics with 3D spatial relationships. The representation encodes distances between 6 types of pharmacophore centers using Gaussian distance bins and aggregates information across molecular conformers. The authors evaluate PharmPD on 22 TDC ADMET tasks and an external OpenADMET PXR dataset. They report improvements over a MapLight-style baseline on 18 of 22 tasks and conduct controlled ablations to investigate the importance of semantic-geometric coupling. The paper also explores interpretability through SHAP attribution, chemical knowledge alignment, and molecule-level feature verification.

Critical strengths
  • The evaluation covers 22 ADMET tasks and an additional PXR dataset. The Full vs. Rank-1 comparison provides evidence that preserving pharmacophore-type-specific distance relationships can improve prediction on certain endpoints. In particular, the improvements on 3 CYP inhibition tasks remain significant after multiple-testing correction. This is the strongest evidence supporting the proposed representation.
  • The descriptor provides chemically meaningful feature identities that can be traced back to pharmacophore centers, distances, and conformers. The reconstruction experiments in the appendix demonstrate that these attribution-related computations can be independently verified. This is a useful property for molecular prediction applications where interpretability is important.
Critical weaknesses
  • The idea of combining pharmacophore identities with spatial distances has been explored in earlier descriptors, particularly CATS3D. PharmPD introduces modifications to chemical typing, distance encoding, and conformer aggregation, but the methodological distinction is relatively incremental. Moreover, the CATS3D-style comparator is not an exact reproduction of the original method, making it difficult to establish which design choices are responsible for the reported improvements. This substantially limits the paper's contribution at an A*-conference like ICLR.
  • Although PharmPD outperforms the Rank-1 control on some tasks, none of the Full vs. Topology comparisons remains statistically significant after multiple-testing correction. Furthermore, the authors acknowledge that Full and Rank-1 show small and inconsistent differences when added to MapLight. These findings suggest that the benefits of explicitly preserving 3D semantic-geometric relationships may be less substantial than the paper's central framing implies.
Questions for authors
  • Could the authors clarify the key methodological differences between PharmPD and CATS3D? In particular, can the reported improvements be attributed to the semantic-geometric representation itself rather than differences in pharmacophore definitions, weighting, or distance encoding?
  • Given that none of the Full vs. Topology comparisons is statistically significant after correction, what evidence supports the claim that Euclidean 3D geometry provides meaningful additional predictive information beyond topological distances?
  • The Full vs. Rank-1 differences become small and inconsistent when both representations are combined with MapLight. How should readers interpret the practical importance of semantic-geometric coupling when strong molecular descriptors are already available?
5 comments
AuthorsOpenReview ↗

W1 / Q1. Methodological Contribution and Controlled Comparisons with CATS3D

Response.

We appreciate the reviewer's request to clarify PharmPD's methodological distinction from established pharmacophore representations and identify the contributions of its design choices. We have expanded the descriptor comparisons, conducted matched ablations, and evaluated the predictive and explanatory value of the resulting representation.

  • (1) CATS3D reproduction and direct comparisons: Exact reproduction of the original MOE-based CATS3D presents practical limitations, as the open implementation jCompoundMapper itself differs from the original specification. We therefore explicitly distinguish a five-type, 150-dimensional reconstruction (C5L) from our local six-type, 210-dimensional implementation (C6), documenting their configurations. Against C5L, PharmPD improves the standalone aggregate by +0.0378 (q = 0.0007), while the MapLight-augmented contrast is +0.0081 (q = 0.063). In the expanded 22-task C6 comparison, PharmPD achieves higher point estimates on 16 tasks, although the individual direct contrasts remain unresolved after correction.

  • (2) Predictive contribution of semantic–geometric correspondence: The Full–Rank-1 control preserves pharmacophore-type and distance marginals, feature dimensionality, weights, and auxiliary information while removing their correspondence. Restoring correspondence improves CYP2C9, CYP2D6, and CYP3A4 inhibition by +0.0367, +0.0678, and +0.0492 AUPRC, respectively (all q = 0.0017). We additionally applied the same intervention to C6, obtaining +0.0404, +0.0530, and +0.0372 AUPRC on these endpoints and +0.0359 AUROC on hERG, all significant after correction within the four-comparison family. This demonstrates that the predictive value of correspondence is not specific to PharmPD's pharmacophore definitions or distance encoding.

  • (3) Contributions of individual design choices: Sequential ablations isolate changes in pharmacophore weighting, distance encoding, and chemical assignments while holding conformer ensembles, pair categories, and pooling fixed. Among twelve component contrasts, soft-pKa weighting on CYP2C9 shows a corrected improvement (+0.00566, q = 0.006). These experiments distinguish the contribution of particular implementation choices from that of preserving correspondence.

  • (4) Chemically verifiable representation and explanation: Beyond predictive accuracy, PharmPD provides explicit relational coordinates whose contributions can be traced to molecular centers, distances, and conformers. Independent reconstruction reproduced features for all 64 molecules in the verification cohort. In fixed eight-leaf surrogate experiments, the PharmPD-augmented explanation space achieved significantly higher fidelity in 6/12 comparisons against MapLight-only and 2/12 against CATS3D-style. Thus, we evaluate not only predictive utility but also whether model behavior can be expressed through chemically defined, independently verifiable relations.

Where to find these revisions.

  • Section 3 and Figure 1: Chemical definitions and representation construction.
  • Section 4.3, Tables 1–2: Matched Rank-1 control and direct CATS3D comparisons.
  • Appendices A.5.1–A.5.4: CATS3D configurations, expanded comparisons, and component ablations.
  • Section 4.4 and Appendices B.1 and B.6: Attribution reconstruction and explanation fidelity.
  • Additional results reported here: Rank-1 intervention applied to C6.

Together, these experiments establish PharmPD as an explicit, chemically inspectable representation and provide controlled evidence for the predictive and explanatory value of preserving pharmacophore identity–distance correspondence.

AuthorsOpenReview ↗

W2 / Q2. Additional Predictive Value of 3D Geometry Beyond Topology

Response.

We thank the reviewer for raising the distinction between the direct predictive advantage of Euclidean distances and their complementary value alongside topological features. We have corrected the original topology control and expanded the analysis to 22 tasks, including additional comparisons designed to isolate the contribution of 3D-derived features.

  • (1) Corrected topology controls: Our audit identified differences in molecular preparation and pharmacophore assignments between the original Topology arm and Full. We replaced that comparison with a matched graph-distance control preserving pharmacophore centers, weights, and auxiliary features, ensuring that distance encoding is the principal difference between the representations.

  • (2) Direct Euclidean–graph comparisons: In the expanded 22-task analysis, both the direct Euclidean-versus-graph comparison and the distance-resolution-matched comparison favor Euclidean distances on 11/22 tasks, with no individual contrast surviving correction (Table 11). We distinguish these direct comparisons from the question of whether 3D features contribute when graph-distance information is already available.

  • (3) Complementary predictive contribution of 3D features: Holding the graph-distance block fixed, adding aligned rather than row-permuted 3D features yields 21/22 positive estimates, with nine corrected improvements. To address possible penalties introduced by shuffled features, we additionally compared graph+3D against graph+duplicate graph, matching input dimensionality without introducing randomized features. The 3D-augmented representation achieves 19/22 positive estimates, including a corrected improvement on Caco-2 (q = 0.033). Against graph-distance features alone, graph+3D also produces 18/22 positive estimates.

  • (4) Predictive and interpretive value of spatial relations: These experiments support a complementary role for 3D-derived features alongside the tested graph-distance representation, rather than a general claim that Euclidean distances outperform graph distances. Importantly, PharmPD's Euclidean-distance coordinates also retain direct molecular interpretations: attributed relations can be reconstructed from specific pharmacophore centers, spatial separations, and conformer contributions. This connects the predictive representation to independently checkable structural evidence.

Where to find these revisions.

  • Section 4.3: Summary of corrected geometry controls.
  • Appendices A.3.1 and D.5: Matched topology construction and correction of the original control.
  • Appendix A.4, Tables 11–12: Expanded 22-task graph-distance and 3D comparisons.
  • Section 4.4 and Appendix B.1: Molecular reconstruction and attribution verification.
  • Additional results reported here: Duplicate-graph control and graph-only comparisons.

These results clarify the role of 3D geometry: its contribution is not necessarily superior standalone accuracy, but complementary predictive value and chemically inspectable spatial relationships.

AuthorsOpenReview ↗

W2 / Q3. Practical Value of Semantic–Geometric Coupling Above Strong Molecular Descriptors

Response.

We appreciate the reviewer's question about the practical importance of semantic–geometric coupling when strong molecular descriptors are available. We address this by distinguishing the incremental value of correspondence, the contribution of the complete PharmPD block, and the usefulness of its chemically explicit coordinates for explaining model predictions.

  • (1) Baseline-dependent contribution of correspondence: We confirmed that the Full–Rank-1 difference diminishes when richer molecular descriptors are available. In a ten-seed backbone analysis, standalone differences of +0.0392, +0.0737, and +0.0499 on CYP2C9, CYP2D6, and CYP3A4 inhibition decrease to +0.0028, −0.0041, and −0.0002 above MapLight. Intermediate backbone comparisons show a general attenuation of this difference. An additional ErG-removal experiment did not substantially restore the standalone gap, indicating that the attenuation cannot be explained solely by the presence of ErG.

  • (2) Incremental value of the complete descriptor: A small Full–Rank-1 difference does not imply that adding PharmPD provides no benefit. In a separate exploratory analysis of stored predictions, adding the full PharmPD block produces positive estimates in all 20 combinations of five backbones and four endpoints, with 15/20 surviving BH correction within that grid. Above full MapLight, two of four endpoint-level increments survive correction. These results concern the complete descriptor rather than correspondence alone and complement the primary 22-task benchmark.

  • (3) Chemically verifiable attribution: PharmPD provides a structured explanation vocabulary independent of the magnitude of its incremental predictive gain. Each attributed relational coordinate can be traced to pharmacophore centers, pairwise distances, Gaussian contributions, and the conformers involved in aggregation. Independent reconstruction verified the stored features for all 64 molecules in the evaluation cohort, establishing that feature-level explanations can be checked against their underlying molecular computations.

  • (4) Explanation fidelity beyond PharmPD-trained predictors: We also evaluated whether these relational coordinates can represent model behavior at fixed explanation complexity. In eight-leaf surrogate comparisons, the PharmPD-augmented explanation space achieved significantly higher fidelity in 6/12 comparisons against MapLight-only and 2/12 against CATS3D-style. Notably, three of the six improvements over MapLight-only occurred for teacher models that did not use PharmPD as input. This demonstrates that the relational representation can also help explain predictions learned from other molecular feature spaces, rather than merely reproducing the inputs of PharmPD-based models.

Where to find these revisions.

  • Section 4.2 and Appendix A.1: Predictive improvements over MapLight across 22 TDC tasks.
  • Section 4.3 and Appendix A.5.5: Correspondence comparisons and ErG group-removal diagnostics.
  • Section 4.4 and Appendix B.1: Molecule-specific attribution provenance and independent verification.
  • Appendix B.6, Tables 34–35: Fixed-complexity explanation fidelity comparisons.
  • Additional results reported here: Backbone ladder, ErG-specific control, and exploratory baseline-increment analysis.

These findings support two complementary contributions of PharmPD: predictive value that depends on the surrounding representation, and chemically interpretable, independently verifiable relations that can improve the fidelity of model explanations. Even when incremental accuracy gains are modest, retaining these explicit relations provides a meaningful way to inspect and evaluate how molecular models arrive at their predictions.

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

This paper introduces PharmPD (Pharmacophore Pair Distance), a molecular descriptor that explicitly couples pharmacophore semantics with three-dimensional molecular geometry to improve molecular property prediction and chemical interpretability. The authors concatenate PharmPD with the MapLight representation and evaluate its performance across 22 TDC ADMET prediction tasks and an external OpenADMET PXR dataset. Experimental results show that incorporating PharmPD improves the mean predictive performance on 18 of the 22 tasks. Furthermore, ablation studies demonstrate that preserving the complete correspondence between pharmacophore types and spatial distances outperforms representations that retain only marginal information on several CYP inhibition prediction tasks.

Critical strengths
  1. The paper introduces a molecular descriptor that explicitly integrates pharmacophore semantics with three-dimensional spatial distances, allowing the representation to preserve the relationships between different pharmacophore types and their spatial arrangements. This design provides chemically meaningful features and a certain degree of interpretability.
  2. Experimental results demonstrate that incorporating PharmPD into existing 2D molecular representations improves predictive performance across multiple ADMET tasks. Specifically, improvements in mean predictive performance are observed on 18 of the 22 TDC benchmark tasks, suggesting that PharmPD provides complementary information to existing molecular representations.
  3. The authors conduct controlled experiments, including the Rank-1 ablation, to investigate the importance of preserving the correspondence between pharmacophore types and spatial distances. Furthermore, chemical-rule audits and SHAP-based feature attribution analyses are performed to examine the descriptor and its interpretability.
Critical weaknesses
  1. Limited methodological novelty. Existing methods, such as ErG and CATS3D, have already incorporated pharmacophore features with molecular topology or 3D geometry, limiting the novelty of the proposed approach. Although PharmPD introduces several modifications, including refined pharmacophore definitions, pH-dependent weighting, and Gaussian distance encoding, sufficient ablation studies are lacking to demonstrate their individual contributions.

  2. Lack of systematic comparisons with strong molecular prediction models. The main experiments primarily use MapLight as the predictive baseline, together with internal ablation controls such as Rank-1. However, systematic comparisons with competitive molecular prediction models, such as ADMET-AI (Chemprop-RDKit), MapLight+GNN, and MolE, are missing across the 22 TDC benchmark tasks. Although Uni-Mol v2 and CheMeleon are considered in the PXR experiments, broader comparisons under consistent evaluation protocols are needed to better establish the predictive performance and practical advantages of PharmPD.

  3. Limitations in generalizability and interpretability evaluation. The downstream evaluations are primarily focused on ADMET-related tasks, with limited independent validation on other molecular property prediction or molecular recognition benchmarks. Furthermore, the interpretability experiments mainly demonstrate the computational traceability of pharmacophore feature attributions. However, it remains insufficiently validated whether these highly attributed features correspond to actual protein–ligand interactions or the underlying mechanisms of molecular activity.

Questions for authors
  1. Could the authors provide additional comparisons with representative molecular prediction models across the 22 TDC ADMET benchmark tasks and report the incremental performance gains obtained by incorporating PharmPD into different strong baselines? Furthermore, could the authors evaluate PharmPD on additional molecular property prediction or molecular recognition benchmarks to provide stronger evidence of its generalizability?
  2. Could the authors further validate the chemical interpretability of PharmPD, for example, by leveraging experimentally determined protein–ligand complex structures or established structure–activity relationships? Such analyses could help determine whether highly attributed pharmacophore features correspond to actual molecular recognition patterns.
  3. Existing methods, such as ErG and CATS3D, have already encoded topological or three-dimensional spatial relationships between pharmacophore features. Could the authors further clarify the main methodological contributions of PharmPD compared with these approaches? In particular, could the authors conduct controlled comparisons under consistent pharmacophore definitions, conformer generation procedures, and distance encoding schemes, together with component-level ablation studies, to demonstrate the contributions of the proposed design choices?
5 comments
AuthorsOpenReview ↗

W2 / Q1. Strong Molecular Predictors and Generalizability

Response.

We thank the reviewer for suggesting stronger molecular prediction baselines and evaluation beyond ADMET. We substantially expanded the experiments in four directions.

  • (1) Comparisons with strong predictors: We evaluated GAT, AttentiveFP, Chemprop-RDKit, Uni-Mol v1, and Uni-Mol2 84M across all 22 TDC tasks under matched conditions. MapLight+PharmPD achieved 46 significant wins and one loss across these five comparator families, using scaffold-cluster bootstrap and multiple-testing correction.

  • (2) Incremental gains across strong representations: We evaluated PharmPD on seven PXR base configurations incorporating MapLight, CheMeleon, Uni-Mol v2, and multi-fidelity information. All seven incremental estimates were positive, with five nominal confidence intervals excluding zero. Notably, adding PharmPD to Uni-Mol v2 + MapLight + multi-fidelity information improved RAE by +0.0067 [0.0029, 0.0107]. The complete analysis also reports the full-stack configuration.

  • (3) Generalization beyond ADMET: We added BACE and KIBA bioactivity evaluations. On the 229-target molecule-disjoint KIBA benchmark, MapLight+PharmPD achieved MSE 0.3755 ± 0.0338, within the numerical range of competitive published cold-drug DTA results. Standalone PharmPD also demonstrated predictive signal across the KIBA target panel. Published studies use different partitions, so their scores provide external performance context rather than matched rankings.

  • (4) Additional external cohorts: We evaluated BindingDB CYP2D6/CYP3A4 and Karim hERG datasets, extending predictive evaluation across distinct molecular cohorts.

Where to find these revisions.

  • Section 4.2; Appendix A.2, Table 4: Matched learned-model comparisons.
  • Appendix A.7.1, Table 22: Seven-base PXR augmentation.
  • Appendix A.8, Tables 23–24: BACE, KIBA, and external-cohort evaluations.
  • Appendix D.7: Statistical procedures and evaluation protocols.

Together, these results substantially broaden the empirical evaluation across strong baselines, representation backbones, and molecular prediction tasks.

AuthorsOpenReview ↗

W3 / Q2. Chemical Interpretability and Experimental Structural Validation

Response.

We appreciate the reviewer's suggestion to validate whether attributed pharmacophore relationships correspond to experimentally observed molecular recognition patterns. We extended the evaluation using experimental protein–ligand complexes and independent chemical-signature analyses.

  • (1) Experimental structural correspondence: We tested whether highly attributed distance channels correspond more closely to intraligand distances observed in experimentally determined bound conformations.
AnalysisEffect (Å) [95% CI]
CYP3A4, 67 co-crystal ligands−0.108 [−0.162, −0.046]
CYP3A4, 64 bound–free comparisons−0.0440 [−0.0783, −0.0105]
KLIFS, 92 kinase targets / 4,110 structures−0.0244 [−0.0387, −0.0102]

These analyses provide evidence of retrospective geometric correspondence between model attribution and experimentally observed bound conformations, extending the evaluation beyond CYP3A4 to 92 kinase targets. We distinguish these pose-referenced associations from prospective bound-pose prediction and document the attribution-input provenance in the Appendix.

  • (2) External chemical consistency: Prespecified pharmacophore pairs showed attribution enrichment relative to abundance on external hERG (PI–AR: 18.2×) and CYP2D6 (PI–HP: 6.4×; PI–AR: 22.2×). In KIBA, AR–HBA exceeded its abundance share in 158/167 target models, with median enrichment of 2.04×. These results extend the chemical-consistency analysis across datasets and target families.

  • (3) Molecular verification: We independently reconstructed PharmPD features for 64/64 molecules and verified all evaluable structured attribution claims, linking model attribution to identifiable molecular centers, distances, and conformer contributions.

  • (4) Structural controls: We additionally evaluated native-pose, in-pocket rotation, and pharmacophore-label controls to assess the specificity of contact-based interpretations.

Where to find these revisions.

  • Section 4.4: Chemical attribution and structural correspondence.
  • Appendices B.1–B.2: Molecular reconstruction and computational verification.
  • Appendix B.5, Tables 32–33: External chemical-signature analyses.
  • Appendix B.9, Table 38: CYP3A4 and 92-target experimental structural analyses.
  • Appendix B.10, Table 39: Structural contact controls.

These analyses extend the interpretability evaluation from computational traceability to recurring chemical patterns and experimental bound-conformation evidence.

AuthorsOpenReview ↗

W1 / Q3. Methodological Contribution and Controlled Comparisons with ErG and CATS3D

Response.

We appreciate the request to clarify PharmPD's methodological distinction from established pharmacophore representations and isolate the contributions of its design choices. We conducted additional matched comparisons and component-level experiments.

  • (1) Representation-level contribution: PharmPD preserves chemically distinguishable interaction roles and their type-specific spatial relationships. The Rank-1 control retains pharmacophore-type and distance marginals while removing their correspondence. Restoring that correspondence improves CYP2C9, CYP2D6, and CYP3A4 inhibition by +0.0367, +0.0678, and +0.0492 AUPRC, respectively (all q = 0.0017).

  • (2) Controlled CATS3D comparison: We distinguished the closest-to-published CATS3D reconstruction (C5L) from our local CATS3D-style implementation (C6), with consistent molecular preparation and conformer generation within the matched experiments. Against C5L, PharmPD improves the standalone aggregate by +0.0378 (q = 0.0007); the MapLight-augmented contrast is +0.0081 (q = 0.063). Both settings are reported together.

  • (3) Component-level ablations: We evaluated sequential changes in pharmacophore weighting, distance encoding, and chemical assignments while holding conformer ensembles, pair categories, and pooling fixed. Among twelve comparisons, soft-pKa weighting on CYP2C9 survives correction (+0.00566, q = 0.006). The remaining component effects are documented to distinguish their individual empirical contributions.

  • (4) ErG and expanded descriptor controls: Group-removal experiments showed larger score decreases after removing PharmPD than ErG in all four tested tasks. We also expanded the C6 comparison to 22 TDC tasks, reporting direct predictive accuracy separately from permutation-controlled differences.

Where to find these revisions.

  • Section 3 and Figure 1: Chemical definitions and representation construction.
  • Section 4.3, Tables 1–2: Rank-1 and direct CATS3D comparisons.
  • Appendices A.5.1–A.5.5: CATS3D comparison families, component analyses, and ErG experiments.
  • Appendix A.4: Corrected topology and geometry controls.

These experiments provide a clearer empirical assessment of PharmPD's design choices and the predictive value of preserving pharmacophore identity–distance correspondence.

Reviewer_oCJJOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

The Paper proposes a molecular representation based on pharmacophore centers and their 3D relationships, which can be used for the molecular property prediction and explainability. PharmPD use the concept of pharmacophore and thus find out centers of each pharmacophore type, get distance between each pair distance in 3D, bin each distance with 21 different pair types and max pool across conformers. Its novelty lies in capturing 3D distance across chemically meaningful pharamcophore centers. They used MapLight-style with PharmPD to show that it improves performance across TDC ADMET tasks, with controlled analysis across information. Most interesting part is showing the explainability analysis, and that PharmPD can show match-decoy enrichment: suggesting it can share chemically meaningful representation.

Critical strengths
  • The controlled representation ablations including Rank-1/Topology/Unjoined has been well constructed to support the claim that PharmPD is meaningful.
  • Readable feature is clear strength compared to fingerprints, but would like to see more where this could be possibly used in cases like drug discovery.
Critical weaknesses
  • Topological ablation shows that none of the contrasts are meaningful. This undermines the claim that 3D geometry adds value. Authors defends this by saying this is task dependent but the number of task shown is limited and additional ablations should have been provided.
  • I am not fully convinced about the clear novelty/effectiveness compared with CATS3D, which is also a pharamacophoric descriptor with similar architecture. The comparison only covers a limited number of tasks (7 of the 22 TDC ADMET tasks), and for PXR contrast the confidence interval includes zero.
Questions for authors
  1. A runtime analysis for PharmPD and related main experiments would be helpful to show that it is not substantial to add on top of MapLight.
  2. I would appreciate the authors' insight on Weakness 2. I would like illustration of the difference and effect.
  3. I am curious about topology-related ablations such as using no scale (without 1.4x).
  4. Max Pooling is applied per channel across conformers, which loses information about the actual conformer. Even though max pooling seems like a natural fit, I would love to see ablation across other pooling methodology.
  5. The computational resource has not been shared (CPU/GPU type), and it would be helpful for reproducibility.
  6. Required "AI Use Statement" is not included in the manuscript? Has no AI been used in the manuscript?
10 comments
AuthorsOpenReview ↗

W1 / Q3: Corrected Topology Controls and Extended Geometry Ablations

Response.

Thank you for the detailed suggestions. We expanded the geometry controls to all 22 TDC tasks and revised the abstract to distinguish predictive gains against shuffled features from direct superiority over graph distances.

W1 — Corrected topology controls. We replaced the original Topology comparison, whose molecular preparation and pharmacophore assignments differed from Full, with controls matching centers, weights, and auxiliary features. The expanded experiment evaluates seven arms across 22 tasks and ten seeds, reporting all five prespecified contrasts.

With the graph-distance block held fixed, molecule-aligned 3D features outperformed their row-permuted counterparts on 21/22 tasks, with nine improvements surviving Benjamini–Hochberg correction. All seven prespecified recognition endpoints had positive estimates, and the three CYP inhibition improvements survived correction across all 22 tasks.

By contrast, direct Euclidean-versus-graph and resolution-matched comparisons each yielded 11/22 positive estimates, with no individually significant differences after correction. Our conclusion is therefore that properly aligned 3D features provide predictive value relative to the specified shuffled control, not that Euclidean distances universally outperform graph distances or contain information irrecoverable from topology.

Q3 — Distance scaling and calibration. Removing the 1.4 scaling factor changed the three CYP scores by only 0.0006–0.0023, with inconsistent directions. Train-only isotonic recalibration did not yield a significant aggregate improvement (q = 0.906). An earlier MapLight-augmented matched-resolution comparison was positive (+0.009727, q = 0.0130), but we explicitly distinguish its evaluation setting from the expanded standalone experiment.

Manuscript locations: Section 4.3; Appendices A.3.2 and A.4; Tables 9–12.

AuthorsOpenReview ↗

W2 / Q2: Contribution Relative to CATS3D and Expanded Benchmark Evaluation

Response

We agree that CATS3D is the closest established comparator and that broader evaluation was necessary.

Contribution relative to CATS3D. We do not claim novelty for combining pharmacophore identities with spatial distances. Our contribution is the chemistry-audited construction, weighted continuous distance encoding, molecule-level attribution provenance, and controlled evaluation of the predictive value of semantic–geometric correspondence.

For example, our Rank-1 control preserves pharmacophore-type and distance marginals, total feature mass, dimensionality, and auxiliary features while removing pair-specific distance profiles. Full PharmPD improves over Rank-1 on the three CYP inhibition tasks, with all three improvements surviving correction.

Comparator implementations and performance. We now distinguish the local six-type, 210-dimensional CATS3D-style implementation (C6) from the five-type, 150-dimensional closest-to-published reconstruction (C5L).

Against C5L, PharmPD achieves a significant standalone aggregate advantage (+0.0378, q = 0.0007). However, the MapLight-augmented comparison does not survive correction (+0.0081, q = 0.0630).

We additionally extended the C6 comparison to all 22 TDC tasks, using four arms and ten seeds per task (880 fits).

  • Primary difference-in-differences (DiD): 17/22 positive estimates, with one task surviving correction.
  • Secondary direct performance comparison: 16/22 positive estimates, but none surviving correction; all 22 confidence intervals include zero.
  • Recognition endpoints: Both comparisons had positive estimates on all seven endpoints, although these are not seven independently significant improvements.

C6 itself outperformed its shuffled control on 18/22 tasks. Thus, our results do not imply that CATS3D lacks useful predictive signal.

External PXR evaluation. We agree that the direct PharmPD–CATS3D comparison does not establish superiority (+0.0091, 95% CI [−0.0030, +0.0205], p = 0.149). However, the prespecified DiD was positive (+0.0266, 95% CI [+0.0142, +0.0389], p < 0.0005).

These results answer different questions. The DiD measures the difference between gains relative to descriptor-specific shuffled controls; it does not overturn the unresolved direct accuracy comparison.

We therefore emphasize the explicit chemical construction and controlled evidence of predictive contribution rather than universal superiority over CATS3D.

Manuscript locations: Sections 4.2–4.3; Tables 1–2; Appendices A.5.1, A.5.3–A.5.4, and A.7; Tables 18 and 21.

AuthorsOpenReview ↗

Q4: Conformer Pooling Ablations

Response

Thank you for requesting additional pooling experiments. We now report controlled comparisons in which molecular coordinates, retained conformers, conformer counts, and ordering were held fixed.

In the standalone setting, mean pooling outperformed max pooling on two CYP inhibition tasks, demonstrating that max is not universally optimal.

However, in the MapLight-augmented setting, max achieved the highest aggregate point estimate among the tested single pooling operators. Concatenating max, mean, and minimum statistics also provided no significant additional benefit in this setting (+0.0001, p = 0.813), despite a positive standalone result.

We therefore retain max pooling based on its measured performance in the setting where it is deployed, rather than assuming that it has a special mechanistic interpretation.

We also clarify that different feature channels may select different conformers, so the pooled representation does not necessarily correspond to a single physically realized molecular pose.

Manuscript locations: Section 3.2 and Appendix A.6.

Reviewer_AHwcOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces PharmPD, a deterministic 200-dimensional molecular descriptor that combines chemically defined pharmacophore features with their pairwise 3D distances. The representation is constructed using manually curated pharmacophore definitions, Gaussian distance kernels, and conformer pooling.

The authors evaluate PharmPD using standard gradient-boosting models (CatBoost and XGBoost) on 22 TDC ADMET tasks and an external OpenADMET PXR dataset. They investigate its complementary predictive value relative to existing molecular descriptors, conduct controlled representation ablations, and evaluate whether feature attributions can be related to established molecular-recognition chemistry.

Critical strengths

S1. Generally sound experimental setup. The authors evaluate the proposed descriptor across a broad collection of established ADMET benchmarks and an external PXR dataset. The experiments include multiple training seeds, controlled representation ablations, and comparisons with alternative molecular descriptors.

S2. Chemically inspectable representation. PharmPD uses explicit pharmacophore-distance features whose values and associated model attributions can be traced back to molecular structures. While this facilitates chemical inspection, the practical interpretability advantages over existing pharmacophore descriptors remain unclear

Critical weaknesses

W1. Limited novelty and contribution to machine learning.

The main contribution is a manually engineered molecular descriptor based on established pharmacophore concepts. The representation contains no learned parameters, and the predictive models use standard gradient-boosting algorithms without methodological modifications.

While chemically informed feature engineering can be valuable, the proposed representation is conceptually close to existing pharmacophore-distance descriptors, particularly CATS3D. The paper does not establish a sufficiently substantial methodological or empirical advance over these existing approaches to justify its contribution to a machine learning conference such as ICLR.

W2. Limited evidence for meaningful improvements over existing representations.

Although PharmPD improves mean performance on 18 of the 22 TDC tasks, only 12 improvements are nominally statistically significant, while four tasks show nominally significant deteriorations. These tests are based on repeated training seeds and are not corrected for multiple comparisons.

More importantly, the controlled ablations and descriptor comparisons do not convincingly establish the advantages of the proposed representation over simpler or existing alternatives:

  • Full PharmPD does not significantly outperform the graph-distance Topology control on any of the seven recognition endpoints after multiple-testing correction.
  • Improvements over the established CATS3D-style descriptor are relatively small, and the standalone comparisons in Table 9 lack statistical inference.
  • On the external PXR benchmark, the difference between PharmPD and CATS3D-style is not statistically significant.
  • The descriptor comparisons change several components simultaneously, including pharmacophore definitions, weighting, and distance encoding. It is therefore unclear which components are responsible for the observed improvements.
  • Furthermore, although the TDC benchmark is widely used, the paper does not establish how its results compare with published state-of-the-art performance on these tasks.

Overall, the experiments suggest that PharmPD can provide useful additional predictive information, but do not convincingly demonstrate a substantial advance over existing molecular representations.

Additional comments to improve paper

C1. Mathematical presentation and notation. The paper devotes considerable space to mathematical formalization of relatively straightforward descriptor construction, while several variables and symbols are not properly defined. This would normally make the methodology difficult to follow, but fortunately Figure 1 provides a much clearer and more intuitive explanation of the proposed descriptor. The mathematical formulation adds little to the understanding of the method in its current form. All variables should be properly defined, and the mathematical presentation could be substantially simplified.

C2. Organization of experimental results. Important experimental comparisons are relegated to the appendix, while considerable space in the main paper is devoted to detailed results for individual CYP endpoints. For example, the descriptor comparisons in Tables 9 and 10 are arguably more important for evaluating the contribution than the endpoint-level results in Table 1. Reporting aggregate metrics across endpoints could free space to present these more informative comparisons in the main paper.

C2. Statistical reporting. Uncertainty estimates should be reported systematically alongside the main results. In particular, Table 9 provides only mean performance, making it difficult to assess whether the relatively small differences between descriptors are meaningful. The limitations of the nominal paired-seed tests used for the 22-task benchmark should also be more prominent.

C3. Methodological clarity. Section 3.4 should provide a more complete overview of the experimental comparisons, including the established descriptor baselines introduced later in Section 4.2.3.

Questions for authors

Q1. Given the conceptual similarity to existing pharmacophore-distance descriptors such as CATS3D, what do the authors consider the principal methodological novelty of PharmPD, beyond the manually curated pharmacophore definitions and descriptor construction choices?

Q2. The full representation does not significantly outperform the topological-distance control on any of the seven recognition endpoints. What evidence supports the need for explicit 3D geometry rather than a simpler graph-distance representation?

Q3. Could the authors provide a controlled comparison with CATS3D-style descriptors using identical pharmacophore definitions and conformer-generation procedures? This would help distinguish improvements due to the specific representation from those arising from the chemistry-aware feature definitions.

Q4. How does the proposed method compare with published state-of-the-art results on the TDC ADMET benchmarks? Given the relatively small improvements over existing descriptors, this comparison seems important for establishing the practical significance of the contribution.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.