acceptodds
Under review as a conference paper at ICLR 2027

DrugPrism: Target-Conditioned Riemannian Metrics on Mixed-Curvature Product Spaces for Protein–Ligand Affinity Ranking

Abstract

Recent spherical and hyperbolic affinity ranking models exhibit complementary performance across protein targets, suggesting that a fixed comparison geometry may be restrictive. We show that target specific geometry alone does not guarantee ranking adaptation, as single geometry rescaling changes intrinsic curvature but preserves rankings for fixed representations. This motivates relative scaling across multiple geometric components. We introduce DrugPrism, which maps protein and ligand features into a shared product manifold with spherical, Euclidean, and hyperbolic components, while a pocket conditioned metric head predicts target specific coefficients that rescale the component metrics. Under the resulting product metric, the squared geodesic distance is exactly a weighted sum of the component squared distances, allowing relative weighting to alter rankings when the components disagree. DrugPrism improves mean Spearman correlation over HypSeek by 13.8% on Merck and 19.7% on Hermite. Controlled ablations show that target conditioned weighting substantially outperforms both uniform product weighting and all single geometry variants. On an unseen OpenBind target, DrugPrism with only five tuning compounds already surpasses LigUnity and HypSeek with 25 compounds. The learned metric weights further reveal clear protein family level structure, with related proteins tending to share similar geometric preferences.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Protein--ligand affinity ranking, Riemannian geometry, mixed-curvature product manifolds, target-conditioned metric learning, drug discovery

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_VuWTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

DrugPrism extends HypSeek with spherical, Euclidean, and hyperbolic projection heads. A pocket-conditioned network predicts weights over the three squared geodesic distances, producing a target-specific ranking score. The model retains HypSeek’s backbone and most training objectives, with geometric supervision adapted to the product metric.

The paper reports improvements over HypSeek on Merck and Hermite, slightly lower performance on JACS, and promising few-shot adaptation on one OpenBind target. It also examines how the learned weights vary across protein families.

Critical strengths

Conditioning the comparison metric on the pocket is a reasonable extension of shared-metric affinity models. It adds flexibility through a small prediction head while retaining reusable ligand representations. The product-metric formulation also provides a consistent way to define scoring and geometric supervision.

The improvements on Merck and Hermite persist after five-model ensembling. Ensemble Spearman correlation increases from 0.5447 to 0.6113 and from 0.5929 to 0.6906, respectively. Using the same backbone and pretrained initialization as HypSeek provides a useful basis for comparison. The few-shot result is also encouraging, although it covers only one target.

Critical weaknesses

The experiments do not isolate the proposed contribution. Mixed-curvature product representations are established prior work (Gu et al., 2019), and much of the architecture and supervision is inherited from HypSeek. The contribution therefore rests on the particular combination of heterogeneous geometries and pocket-conditioned weighting. However, the ablations omit both globally learned weights and a capacity-matched model with three Euclidean branches and the same conditioning network. Uniform weights are a weak control because the component distances can have different scales. The current results cannot separate gains from mixed curvature, additional projection capacity, distance calibration, and target conditioning. The rank-invariance result applies to scalar rescaling of fixed representations; it does not establish a limitation of more flexible single-space models.

Several reporting gaps limit confidence in the empirical conclusions. The full model’s Merck Spearman correlation is 0.6326 in Table 2, compared with 0.5731 for the mean and 0.6113 for the ensemble in Table 1, without an explanation of the different settings. Ablations lack uncertainty estimates, and the paper does not diagnose the decline on JACS. Section 4.3 points to Appendix B for component dimensions and optimization details, but that appendix mainly specifies the metric head. The broad practical claims also need narrower wording: FEP+ remains better on Merck and Hermite, while the Boltz-2 timing comparison includes structure prediction and uses a different input and deployment protocol. Joint structure and affinity prediction is part of Boltz-2’s task, so this timing gap does not isolate scoring efficiency.

Questions for authors

Contribution and controls: Can you provide results for globally learned component weights and a capacity-matched three-Euclidean-branch model with the same pocket-conditioned weighting? Please clarify how component distance scales are controlled. These comparisons would materially affect my assessment of the geometric contribution.

Experimental consistency: What accounts for the full-model discrepancy between Tables 1 and 2? Were all ablations retrained under the same split, initialization protocol, training budget, and model-selection procedure? Please report variation across seeds and clarify whether single-geometry variants preserve the total representation dimension.

Implementation: Please specify the geometry-head dimensions, layer configurations, parameter sharing, mapping safeguards, and optimization settings. Figure 2 appears to feed mapped query embeddings to the metric head, whereas Eq. 16 and Appendix B specify pocket tangent coordinates. Which data flow is implemented, and how are the pocket and sequence scores used at inference?

Scope of the gains: Which JACS targets account for the decline relative to HypSeek, and is the difference reliable across seeds? HypSeek also benefits slightly more from ensembling on all three benchmarks. Do the per-seed predictions suggest that DrugPrism mainly improves training stability, or that it corrects different ranking errors? The smaller ensemble gain is not itself a weakness, but this distinction matters for interpreting the improvement.

5 comments
AuthorsOpenReview ↗

Response to Reviewer VuWT (1/4)

W1/Q1. Contribution and controlled evaluation of target-dependent geometry

We thank the reviewer for these important comments. We acknowledge that mixed-curvature product spaces are established prior work, including Gu et al. (2019). DrugPrism's contribution is not the product space or weighted-sum operation itself, but a principled approach to adapting geometric comparisons to individual protein targets, motivated by an empirical observation and supported by controlled experiments.

Motivation and geometric formulation. Despite sharing the same curated training data and similar pretrained backbones, spherical LigUnity and hyperbolic HypSeek exhibit striking target-wise performance reversals. This motivates a natural question: could different protein targets benefit from different geometric inductive biases, rather than one universally preferred geometry?

DrugPrism addresses this question through a product of spherical, Euclidean, and hyperbolic factors, with pocket-conditioned coefficients controlling their relative contributions. Intuitively, the three geometries provide complementary geometric components, while the learned weights specify each target's preferred combination. For fixed representations, scalar rescaling of a single geometry preserves ligand rankings (Theorem 4.2), whereas reweighting multiple components can change rankings when their pairwise preferences disagree (Theorem 4.4). Importantly, the weighted sum of component squared geodesic distances is exactly the squared distance induced by the target-conditioned Riemannian product metric. Thus, the simple scoring form is not an ad hoc aggregation rule, but a structured family of target-adaptive comparison metrics, with geometric representations and metric weights learned jointly.

Controlled evidence: target conditioning and mixed curvature. To isolate the proposed contributions, we introduce two additional baselines:

  • SEH-G: Retains the same mixed-curvature factors and projection architecture, but replaces pocket-conditioned weights with globally learned weights, allowing the model to learn a fixed optimal mixture rather than relying on uniform coefficients.
  • EEE-T: Replaces the spherical/Euclidean/hyperbolic factors with three independent Euclidean components (43/42/43 dimensions) while preserving the pretrained backbone, three-branch projection architecture, and target-conditioned weighting network.

DrugPrism (SEH-T) and EEE-T both use a 128-dimensional representation and exactly 132,070,639 trainable parameters. Training data, objectives, optimization settings, training budget, and evaluation protocols are matched, with geometry-dependent operations adapted appropriately. The following results report five-seed mean ± SD (Pearson / Spearman):

BenchmarkDrugPrism (SEH-T)SEH-GEEE-T
JACS.7171±.0271 / .7131±.0266.6924±.0217 / .6879±.0313.6801±.0349 / .6763±.0394
Merck.6074±.0257 / .5731±.0441.5758±.0434 / .5294±.0428.5587±.0390 / .5104±.0307
Hermite.7082±.0132 / .6392±.0358.6356±.1209 / .5702±.1112.6204±.0556 / .5986±.0674

Target conditioning improves mean Spearman over SEH-G by +0.0252/+0.0437/+0.0690, while mixed curvature improves over parameter-matched EEE-T by +0.0368/+0.0627/+0.0406 on JACS/Merck/Hermite, respectively. Both comparisons favor DrugPrism in Pearson and Spearman across all three benchmarks, providing evidence for the distinct benefits of target-specific weighting and heterogeneous geometry.

AuthorsOpenReview ↗

Response to Reviewer VuWT (2/4)

W1/Q1 (Continued)

Explicit control for component distance scales. We agree that uniform coefficients are a weak control when component distances have different scales. Although SEH-G allows globally learned scale compensation, we further tested whether explicit distance normalization eliminates the SEH–EEE advantage.

For both models, we estimated fixed component-wise squared-distance scales from 2,048 training assays and 16,887 assay–ligand pairs, using identical sampled pairs and an assay-balanced estimator. No evaluation data were used for calibration. We normalized each component's squared distance by its estimated scale before weighting and consistently applied the corresponding normalization to geometric supervision. Both models were then retrained under matched architectures, parameter counts, objectives, and optimization settings.

The following Merck results use the same fixed-seed evaluation setting as the original Table 2:

SettingSEH-T Pearson / SpearmanEEE-T Pearson / Spearman
Original distances.6379 / .6326.5617 / .5276
Scale-calibrated distances.5651 / .5435.4970 / .4378

The mixed-curvature Spearman advantage remains approximately +0.105 before and after calibration. Under calibrated distances, SEH-T outperforms EEE-T on 7 of 8 Merck targets, with paired target-level bootstrap 95% CIs of [+0.0624, +0.1404] for Spearman and [+0.0253, +0.1063] for Pearson. Although normalization reduces absolute performance in this experiment, it does not eliminate the SEH–EEE difference.

These results provide direct evidence that the mixed-curvature advantage is not explained solely by additional projection capacity, target-conditioned weighting, or uncalibrated component-distance scales.

Theoretical and empirical scope. We agree that rank invariance under scalar rescaling applies to fixed representations; it neither rules out more flexible single-geometry models nor guarantees predictive improvement. Our theory characterizes when relative metric weighting enables ranking changes, while the controlled experiments establish its empirical value in the proposed architecture. Complementing these results, DrugPrism achieves 0.7793 Pearson and 0.7489 Spearman across 389 Uni-FEP[1] targets, providing additional evidence beyond the original benchmarks.

Overall, DrugPrism offers a simple, principled, and empirically validated approach to adapting geometric inductive biases across protein targets, without claiming novelty for mixed-curvature product spaces themselves.

W2/Q2. Experimental consistency, seed variation, and ablation reporting

We thank the reviewer for identifying this reporting ambiguity and apologize for not clearly distinguishing the evaluation settings in Tables 1 and 2.

Table 1 vs. Table 2. Table 1 reports results from five independently trained DrugPrism models. On Merck, 0.5731 is the mean Spearman across individual runs, while 0.6113 is obtained by averaging the five models' predictions before evaluation. In contrast, Table 2 reports ablations using one fixed training seed, where the full model achieves 0.6326. Thus, the difference arises from seed aggregation rather than changes in the benchmark, test data, or evaluation metrics.

Ablation consistency and uncertainty. All original Table 2 variants were retrained under the same data splits, pretrained initialization protocol, 50-epoch training budget, optimization settings, and model-selection procedure. To strengthen the statistical evidence for our central contributions, we additionally evaluated SEH-G and EEE-T across five independent seeds, using the same training and evaluation protocols as DrugPrism. These results, reported in W1/Q1, provide mean ± SD estimates for both target-conditioning and mixed-curvature comparisons.

Representation dimensionality. DrugPrism uses 128 intrinsic dimensions partitioned as 43/42/43 across spherical, Euclidean, and hyperbolic factors. The original one-hot single-geometry ablations select one component for scoring and therefore should not be interpreted as full-capacity 128-dimensional single-geometry baselines. Our new EEE-T control addresses this limitation by preserving the complete 128-dimensional three-branch architecture, target-conditioned weighting, and exactly 132,070,639 trainable parameters, while replacing all components with Euclidean geometry.

We will clarify these distinctions and the corresponding evaluation protocols in the revised manuscript.

[1] Rongfeng Zou, Liguo Wang, Xinyan Wang, Ye Ding, and Hang Zheng. Breaking Barriers in FEP Benchmarking: A Large-Scale Dataset Reflecting Real-World Drug Discovery Challenges. 2025.

AuthorsOpenReview ↗

Response to Reviewer VuWT (3/4)

W2. Scope of accuracy and runtime claims

We thank the reviewer for this important clarification. We agree that predictive accuracy and computational efficiency should be distinguished, and that our practical claims should reflect the different tasks and evaluation protocols.

Predictive accuracy. We acknowledge that FEP+ remains more accurate on Merck and Hermite, achieving Spearman correlations of 0.6798 and 0.8720, compared with 0.6113 and 0.6906 for the DrugPrism ensemble. We will therefore emphasize DrugPrism's improvements over the evaluated learning-based baselines and its computational efficiency, rather than claiming uniformly superior accuracy to physics-based methods.

Computational cost. DrugPrism and Boltz-2 were measured on single A100 GPUs, but their inference tasks differ: DrugPrism ranks ligands using prepared protein pockets, whereas Boltz-2 jointly predicts complex structures and affinity. Our Boltz-2 evaluation also uses per-ligand inference rather than optimized batching. Furthermore, the FEP+ estimates are derived from measured throughput reported in NVIDIA's 2022 DGX A100 white paper, rather than direct JACS measurements. We will explicitly distinguish measured runtimes from throughput-derived estimates and present these results as protocol-specific computational costs, not intrinsic speed comparisons under identical tasks.

Q3. Implementation details and metric-head data flow

We thank the reviewer for identifying these ambiguities. We agree that the implementation details should be more explicit and will expand Appendix B with complete architectural configurations, numerical safeguards, and optimization settings. We also acknowledge that the current Figure 2 is inaccurate in its depiction of the metric-head input, and we will revise it to ensure consistency with the implemented data flow and the mathematical formulation.

Metric-head input and inference. The implemented metric head receives the scaled pocket tangent coordinates before exponential mapping, exactly as specified in Eq. 16 and Appendix B, rather than the mapped manifold embeddings currently suggested by Figure 2. Three factor-specific residual-logit networks produce the metric coefficients through softmax normalization, with 17,414 trainable parameters in the metric head. During training, both pocket–ligand and sequence–ligand branches contribute to the objectives, sharing the geometric weights predicted exclusively from the pocket. At inference, only pocket–ligand scores are used for affinity ranking, following HypSeek's protocol; the sequence branch provides auxiliary training supervision. We will revise Figure 2 and clarify this training–inference distinction in Section 4.4.

Projection architecture. DrugPrism uses nine independent projection MLPs—one for each spherical, Euclidean, and hyperbolic component in the pocket, sequence, and ligand towers—without parameter sharing. The component dimensions are 43/42/43 (128 in total). Pocket and ligand features are 512-dimensional (Uni-Mol), while sequence features are 480-dimensional (ESM-2, esm2_t12_35M_UR50D). Each projection uses two linear layers with an intermediate ReLU: 512→512→d_j for pocket/ligand and 480→480→d_j for sequence, without LayerNorm or dropout. The projected tangent coordinates have learned multiplicative scales initialized to 1/√128 and bounded above by one before manifold mapping. The metric head uses factor-specific LayerNorm(d_j)→Linear(d_j,128)→GELU→Dropout(0.1)→Linear(128,1) networks, followed by global-logit addition and softmax.

Numerical safeguards. Spherical and hyperbolic curvature magnitudes are fixed at one. Exponential maps and geodesic distances are computed in FP32, including under mixed-precision training. We smoothly bound spherical tangent radii below π/2, cap hyperbolic sinh arguments, and apply valid-domain clamping, small-distance approximations, and a 10⁻⁶ numerical tolerance.

Optimization. All three pretrained encoders, nine projection MLPs, and the metric head are jointly fine-tuned, totaling 132,070,639 trainable parameters. Training uses four-GPU DDP (global batch size 96), AdamW (learning rate 10⁻⁴, zero weight decay), polynomial decay with 6% warm-up, and gradient clipping at 1.0. Models are trained for 50 epochs, with final parameters averaged over epochs 41–50, without benchmark-based checkpoint selection.

These revisions will make the architecture, training protocol, and inference procedure explicit and fully consistent across the figure, text, and implementation.

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

DrugPrism proposes a protein-target-conditioned metric learning method for ranking protein–ligand binding affinities. It combines spherical, Euclidean, and hyperbolic geometries and learns target-specific weights to determine how much each geometry contributes to the predicted affinity ranking.

The paper evaluates DrugPrism on JACS, Merck, and Hermite benchmarks, reporting improvements over HypSeek on Merck and Hermite, but slightly worse results on JACS. It also evaluates few-shot adaptation on an unseen EV-A71 2A protease target and analyzes whether learned geometric preferences vary across protein families.

The main contribution is a mathematical framework for combining multiple geometries with target-conditioned weighting. However, the evidence that mixed-curvature geometry causes the performance gains remains incomplete, and the few-shot experiment on a single target does not establish broad generalization to unseen proteins. The biological interpretations of the learned geometric preferences are also not independently validated.

My biggest critique is the size of the datasets. They evaluate affinity-ranking performance on three benchmarks: JACS, MERCK, and Hermite. JACS contains 8 protein targets and 199 ligands, MERCK contains 8 protein targets and 262 ligands, and Hermite contains 14 protein targets and 107 ligands. Thus, a total of 8+8+14 targets. This is too small to claim any generalization. Also, the OpenBind experiment uses one target with few-shot adaptation and therefore does not establish robust generalization to a diverse set of unseen proteins.

Post rebuttal: The authors have addressed the central concerns by considering one more dataset and also regarding the contribution of mixed-curvature geometry by introducing a parameter-matched, three-branch Euclidean baseline (EEE-T) with identical trainable parameters and target-conditioned weighting. Overall, the rebuttal substantially strengthens the experimental evidence and addresses most of my major concerns, however, the biological interpretation and reproducibility of learned geometric preferences remain open questions. I have increased my score.

Critical strengths

Strengths are listed below:

  1. The paper addresses an important problem in protein–ligand affinity ranking: different protein targets may benefit from different ways of measuring protein–ligand compatibility.

  2. The proposed method combines spherical, Euclidean, and hyperbolic representations and learns target-conditioned weights to combine their distances.

  3. DrugPrism improves mean Spearman correlation over HypSeek from 0.5034 to 0.5731 on Merck and from 0.5338 to 0.6392 on Hermite.

  4. The ablation studies show that target-conditioned weighting performs better than uniform weighting and individual geometries on the reported Merck evaluation.

  5. The paper evaluates both mean predictions and ensembles across multiple random seeds, providing some evidence about the consistency of the results.

  6. The OpenBind experiment considers few-shot adaptation to an additional target, which is relevant to practical affinity-ranking scenarios where only a small number of compounds are available.

  7. The analysis of learned geometric weights across protein families is an interesting attempt to understand how the method behaves beyond aggregate prediction metrics.

  8. The paper compares against several relevant methods, including HypSeek, LigUnity, and Boltz-2.

Critical weaknesses

Weaknesses in my opinion are:

  1. The benchmarks are small in terms of independent protein targets: JACS and Merck each contain 8 targets, while Hermite contains 14. This is limited evidence for a method whose central claim is that target-conditioned weights improve generalization across proteins.

  2. DrugPrism does not outperform HypSeek on JACS. Its mean Spearman correlation is 0.7131 versus 0.7239 for HypSeek, and its ensemble score is 0.7654 versus 0.7819. The results therefore do not support a consistent improvement across all three benchmarks.

  3. The experiments do not isolate the contribution of mixed-curvature geometry from the effects of the shared pretrained backbone, projection heads, training objectives, or additional model capacity.

  4. The ablations do not establish that mixed curvature is necessary. A simpler target-conditioned model that combines ordinary learned distances or uses a standard gating network could potentially achieve similar results.

  5. The evaluation provides limited evidence for cold-target generalization. The OpenBind experiment uses one target with few-shot adaptation and therefore does not establish robust generalization to a diverse set of unseen proteins.

  6. The paper does not provide enough evidence about target and ligand overlap between training and evaluation data, including protein sequence or structure similarity and chemical-series similarity. Such overlap could affect the difficulty of the evaluation.

  7. The paper does not sufficiently establish the statistical reliability of the improvements across targets. Aggregate scores can conceal cases where the method helps only a few targets, particularly when each benchmark contains so few proteins.

  8. The comparison with prior methods is not fully controlled because some baseline results are taken from published evaluations rather than obtained under identical splits, preprocessing, and training protocols.

  9. The protein-family analysis is descriptive. The reported associations between geometric preferences and biological recognition mechanisms are not independently validated, and their stability across seeds, data subsets, etc. is unclear.

  10. The runtime comparison uses different computational protocols, and the FEP+ cost is estimated from the literature rather than measured directly under the same setup. The efficiency claims should therefore be interpreted cautiously.

Questions for authors

Questions for the authors are:

  1. Can the authors provide per-target results for JACS, Merck, and Hermite, including paired confidence intervals or other appropriate uncertainty estimates? On how many targets does DrugPrism outperform HypSeek and the other main baselines?

  2. Why does DrugPrism perform worse than HypSeek on JACS? Can the authors characterize the targets or ligand series for which the proposed method helps or hurts?

  3. Can the authors compare DrugPrism with a parameter-matched baseline that uses the same pretrained encoders but combines ordinary Euclidean distances through target-conditioned weights? This would help determine whether mixed curvature is necessary.

  4. Can the authors isolate the contributions of the geometric projection heads, target-conditioned weighting, geometric supervision, and regularization through controlled ablations?

  5. Can the authors provide a larger evaluation with more independent protein targets and explicit target-held-out splits? If not, can they narrow their claims about generalization across targets?

  6. How are the training and test targets, ligands, and assay records separated? Can the authors report protein sequence and structure similarity, chemical scaffold or series overlap, and any relevant data overlap between training and evaluation?

  7. In the OpenBind experiment, how are the 25 adaptation compounds selected? Are the results stable across different adaptation subsets, and can the authors repeat the experiment on multiple unrelated targets?

  8. Can the authors distinguish zero-shot cold-target performance from few-shot adaptation performance and report these as separate settings?

  9. Can the authors explain why the theoretical results establish more than the expressivity of the weighted scoring function? What evidence shows that the learned weights improve prediction accuracy rather than merely allowing different rankings?

  10. How stable are the learned geometric weights across seeds and data subsets? Can the authors validate the claimed protein-family associations against independent biological or structural properties?

  11. Can the authors clarify which Table 1 comparisons use results from published evaluations rather than controlled reruns?

6 comments
AuthorsOpenReview ↗

Response to Reviewer TwPs (1/5)

W1/Q5. Large-scale evaluation across protein targets

We thank the reviewer for highlighting the limited number of protein targets in our original evaluation. We agree that the 30 target–ligand series across JACS, Merck, and Hermite provide limited evidence for broad generalization, particularly given the importance of target-dependent geometry in DrugPrism.

To directly address this concern, we conducted a substantially larger evaluation on the publicly available Uni-FEP benchmark[1], covering 389 protein targets. Since high-quality experimental affinity measurements are costly to obtain, established affinity-ranking benchmarks often contain relatively few target–ligand series. The Uni-FEP evaluation therefore provides a valuable opportunity to assess DrugPrism across a much broader collection of protein–ligand systems.

We evaluated DrugPrism alongside Boltz-2, LigUnity, and HypSeek on the same Uni-FEP evaluation set, using consistent target-wise macro-averaged Pearson and Spearman correlations. All three baselines were evaluated using their publicly released checkpoints. Importantly, DrugPrism, LigUnity, and HypSeek share exactly the same curated training dataset, enabling a particularly controlled comparison in terms of training-data coverage. Boltz-2, which follows a separate training pipeline, serves as an additional external baseline.

MethodPearson rSpearman ρ
Boltz-20.64210.5952
HypSeek0.65770.6289
LigUnity0.70440.6829
DrugPrism0.77930.7489

DrugPrism achieves the best performance on both metrics, improving Spearman by 19.1% over HypSeek, 9.7% over LigUnity, and 25.8% over Boltz-2. Notably, the improvement over the strongest competing baseline, LigUnity, reaches +0.0660 Spearman and +0.0749 Pearson, despite sharing the same training data.

These results substantially strengthen the empirical evidence beyond the original 30-series evaluation, demonstrating that DrugPrism's advantages persist across a much larger and more varied collection of protein targets, rather than being confined to a handful of established benchmarks. We will provide further details on benchmark composition and evaluation protocols in the revision.

W2/Q2. JACS performance and target-dependent geometric preferences

We thank the reviewer for this insightful question. We agree that DrugPrism does not uniformly outperform HypSeek: its mean Spearman is only slightly lower on JACS (0.7131 vs. 0.7239; Δρ = −0.0108), while improving by +0.0697 on Merck and +0.1054 on Hermite. We investigated whether these differences could be explained by the characteristics of the underlying ligand-ranking tasks.

Why are the gains limited on JACS? Our additional analysis of all 30 target–ligand series reveals that JACS contains substantially more ligand pairs that are both structurally similar and close in experimental affinity: 20.1% of pairs have ECFP4 Tanimoto similarity ≥ 0.70 and affinity difference ≤ 1 kcal/mol, compared with only 6.4% on Merck and 10.0% on Hermite. Such fine-grained ranking tasks are already particularly well handled by HypSeek's hyperbolic representations. Indeed, HypSeek achieves 0.7819 ensemble Spearman on JACS, substantially outperforming spherical LigUnity (0.6460) and even exceeding the reported FEP+ reference (0.7595). This suggests that hyperbolic geometry already captures much of the relevant ranking information on JACS, leaving limited room for additional gains from geometric adaptation.

When does DrugPrism offer greater advantages? Merck and Hermite contain more heterogeneous ligand-ranking tasks, where a single fixed geometry may be less sufficient. DrugPrism can instead exploit complementary ranking information from spherical, Euclidean, and hyperbolic components through target-conditioned metric adaptation. Consistent with this interpretation, its improvements over HypSeek are substantially larger on these benchmarks.

Overall, DrugPrism is particularly beneficial when heterogeneous affinity relationships favor flexible, target-dependent geometric comparisons, while its additional gains are smaller when a single geometry is already highly effective. We will clarify this benchmark-dependent behavior and provide further target-wise results in the revision.

[1] Rongfeng Zou, Liguo Wang, Xinyan Wang, Ye Ding, and Hang Zheng. Breaking Barriers in FEP Benchmarking: A Large-Scale Dataset Reflecting Real-World Drug Discovery Challenges. 2025.

AuthorsOpenReview ↗

Response to Reviewer TwPs (2/5)

W4/Q3. Contribution of mixed-curvature geometry

We thank the reviewer for this important suggestion. We agree that the original ablations demonstrate the benefit of target-conditioned weighting but do not fully isolate the contribution of mixed-curvature geometry. To address this, we introduced EEE-T, a parameter-matched three-branch Euclidean baseline that retains DrugPrism's pretrained backbone, independent projection MLPs, and identical target-conditioned weighting network. Both models have exactly 132,070,639 trainable parameters and use the same training data, objectives, optimization settings, and evaluation protocol.

Across five independent seeds, DrugPrism consistently outperforms EEE-T in both Pearson and Spearman correlations on all three benchmarks:

BenchmarkDrugPrism Spearman (mean ± SD)EEE-T Spearman (mean ± SD)Gain
JACS0.7131 ± 0.02660.6763 ± 0.0394+0.0368
Merck0.5731 ± 0.04410.5104 ± 0.0307+0.0627
Hermite0.6392 ± 0.03580.5986 ± 0.0674+0.0406

Since EEE-T preserves the same model capacity and target-conditioned weighting mechanism, these results provide direct empirical evidence that mixed-curvature geometry contributes predictive value beyond adaptive distance weighting or additional parameters alone.

Further architectural and experimental details are provided in our response to Reviewer caQH (W2/Q1).

W3/Q4. Isolating target conditioning, geometric supervision, and projection capacity

We thank the reviewer for this suggestion. We conducted additional controlled experiments to disentangle the contributions of target-conditioned weighting, mixed-curvature geometry, and geometric supervision.

(1) Target-conditioned vs. globally learned weights. We introduce SEH-G, which replaces DrugPrism's pocket-conditioned metric weights with a single globally learned set of weights, while retaining the same mixed-curvature components and projection architecture. Results are reported as mean ± SD across five runs:

BenchmarkDrugPrism (SEH-T) SpearmanSEH-G SpearmanGain
JACS0.7131 ± 0.02660.6879 ± 0.0313+0.0252
Merck0.5731 ± 0.04410.5294 ± 0.0428+0.0437
Hermite0.6392 ± 0.03580.5702 ± 0.1112+0.0690

DrugPrism consistently achieves higher mean Spearman across all three benchmarks, showing that target-specific geometric weighting is more effective than a single globally learned mixture.

(2) Geometric supervision and regularization. Our original Table 2 already reports ablations of the geometry-aware loss (LgeomL_{\mathrm{geom}}) and regularizers (Rang,Rhet,RspecR_{\mathrm{ang}}, R_{\mathrm{het}}, R_{\mathrm{spec}}). We further removed cone-related supervision (radial and angular losses, together with angular-margin regularization), while retaining the geometric representations, projection heads, target-conditioned weighting, and ranking objectives.

The following results are from matched runs on Merck using the same fixed seed:

ModelPearson rSpearman ρ
SEH-T w/o cone supervision0.57260.5817
EEE-T w/o cone supervision0.45560.4448

Removing cone supervision reduces SEH-T's Spearman from 0.6326 to 0.5817. Importantly, SEH-T still outperforms EEE-T without cone supervision (0.5817 vs. 0.4448), indicating that the benefits of mixed curvature are not solely attributable to auxiliary geometric losses.

(3) Projection capacity. DrugPrism partitions its 128-dimensional representation into 43/42/43 dimensions, rather than using three separate 128-dimensional representations. Moreover, the parameter-matched EEE-T control retains the same backbone, three-branch projection architecture, and target-conditioned weighting network, with exactly 132,070,639 trainable parameters in both models. This controls for additional projection capacity as an explanation for the mixed-curvature gains (see W4/Q3).

Together, these experiments provide complementary evidence for the respective contributions of target conditioning, mixed-curvature representations, and geometric supervision.

AuthorsOpenReview ↗

Response to Reviewer TwPs (3/5)

W5/Q7. Few-shot generalization and robustness to adaptation-set selection

We thank the reviewer for this valuable suggestion. We conducted additional experiments to evaluate generalization across unseen targets and robustness to adaptation-set selection.

Generalization across unseen targets. Beyond the original OpenBind EV-A71 2A protease, we evaluated DrugPrism on five additional held-out ChEMBL targets. All five show substantial few-shot improvements, with macro-averaged Spearman increasing from −0.0528 (zero-shot) to 0.3740 (5-shot) and 0.6267 (10-shot). Every target achieves Spearman above 0.52 with only 10 compounds. Detailed per-target results and interquartile ranges are provided in our response to Reviewer caQH (Q2).

Robustness to adaptation-set selection. In the original OpenBind experiment, the 25-compound adaptation pool was randomly selected, with 100 sampled few-shot subsets. To further assess whether the results depend on this initial selection, we repeated the evaluation using five independent random 25/469 adaptation/test splits from the 494 compounds.

The following results report mean ± SD across the five splits:

ShotSpearman ρ (mean ± SD)
0−0.1975 ± 0.0107
50.5147 ± 0.0302
100.5662 ± 0.0368
250.6140 ± 0.0124

The improvements are consistent across all five splits. At 5 shots, Spearman ranges from 0.4627 to 0.5396, while at 25 shots it ranges from 0.5974 to 0.6265, with a small standard deviation of 0.0124. These results indicate that DrugPrism's few-shot gains are robust to the initial selection of adaptation compounds, rather than relying on a particularly favorable tuning pool.

Together, these additional experiments strengthen the evidence for few-shot adaptability across unseen targets and stability under different adaptation-set selections. We will include the complete split-level results in the revision.

W6/Q6. Training–evaluation separation and data leakage prevention

We thank the reviewer for raising this important concern and apologize for not making our data separation protocol sufficiently clear.

Importantly, as already documented in Appendix D.1, we performed strict protein-sequence-level exclusion against all evaluation targets in JACS, Merck, and Hermite before training. Specifically, we collected the sequences of all benchmark targets and removed every training record associated with an identical protein sequence. This ensures that no exact evaluation-target sequence or its associated affinity records were included in training.

DrugPrism also follows the same curated training dataset and data-cleaning procedures as HypSeek and LigUnity, including ligand- and assay-level deduplication, ensuring consistent training-data coverage across the three methods.

We acknowledge that exact-sequence exclusion does not guarantee complete sequence-, structure-, or scaffold-level disjointness, as homologous proteins and chemically related ligands may remain. We will make this distinction explicit and supplement the revision with protein sequence/structure similarity and ligand scaffold-overlap analyses to more fully characterize train–test similarity.

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

The paper proposes DrugPrism, a method for ranking ligands by binding affinity to a protein target. Proteins and ligands are placed in a combined space with spherical, Euclidean, and hyperbolic parts. A target-conditioned head predicts a weight for each part, so their importance can change from target to target. The paper gives a theoretical analysis of how metric scaling affects rankings, and tests the method on JACS, Merck, and Hermite. DrugPrism beats HypSeek on Merck and Hermite, and ablations show that target-conditioned weights help. The paper also reports few-shot results on an unseen target.

Critical strengths
  1. The motivation is clear. The theory explains a simple point: with fixed representations, scaling a single geometry cannot change the ranking, but changing the relative weights of several geometries can.
  2. The method is tested on three benchmarks, with useful ablations. Table 2 shows that target-conditioned weights clearly beat uniform weights and each single geometry. The few-shot results suggest the method may help when a target has little affinity data.
Critical weaknesses
  1. Limited novelty. DrugPrism reuses the HypSeek backbone and training losses. The main new part is a target-dependent weighted sum of distances. The Riemannian framing makes it look more complex, but the final score is basically a weighted sum of three distance scores. The theory describes properties of scaling and weighted sums, but does not show why mixed-curvature geometry is needed for affinity ranking.
  2. The benefit of mixed curvature is not isolated. Table 2 shows that target-conditioned weights beat uniform weights and single geometries. But it does not show whether spherical and hyperbolic spaces help more than a strong Euclidean model with the same kind of target-dependent weights. Useful controls would be a parameter-matched multi-branch Euclidean model, or a learned Mahalanobis metric. Without them, the gain may come from adaptive score mixing, not from the geometry.
Questions for authors
  1. Can you compare with a parameter-matched multi-branch Euclidean model that uses the same target-conditioned weighting? This would show whether mixed curvature adds anything beyond adaptive weighting.
  2. Can you run the few-shot experiment on several unseen targets and report the variation? What happens if you fine-tune only the metric head and freeze the projection layers?
  3. Why is DrugPrism worse than HypSeek on JACS but better on Merck and Hermite? Are there target properties that predict when the method helps?
  4. On Merck, Table 2 reports a full-model Spearman of 0.6326, but Table 1 reports 0.5731 (mean) and 0.6113 (ensemble). What explains the difference? Are the evaluation settings or averaging different?
5 comments
AuthorsOpenReview ↗

Response to Reviewer caQH (1/4)

Q1/W2. Does mixed-curvature geometry provide benefits beyond target-conditioned weighting?

We thank the reviewer for this important suggestion. We agree that the original ablations establish the benefit of target-conditioned weighting, but do not fully isolate the contribution of mixed-curvature geometry. To directly address this concern, we introduce EEE-T, a parameter-matched, three-branch Euclidean control that retains the same weighting mechanism as DrugPrism (SEH-T).

Controlled comparison. EEE-T replaces the spherical, Euclidean, and hyperbolic components (43/42/43 dimensions) with three independent Euclidean components of identical dimensions. Importantly, EEE-T retains the same pretrained backbone, independent projection MLPs across all three towers (ligand/pocket/sequence), and identical target-conditioned weighting network. Both models have exactly 132,070,639 trainable parameters. We further keep the training data, supervision objectives, loss weights, optimization settings, 50-epoch training budget, and evaluation protocol unchanged, with geometry-dependent operations adapted to their Euclidean counterparts. This design controls for both additional representation capacity and adaptive score weighting, allowing us to assess the contribution of mixed-curvature geometry.

Results. We train EEE-T with five independent random seeds and evaluate it under the same protocol as DrugPrism. The following results report five-seed mean ± standard deviation:

BenchmarkDrugPrism (SEH-T) Pearson / SpearmanEEE-T Pearson / SpearmanSpearman gain
JACS.7171±.0271 / .7131±.0266.6801±.0349 / .6763±.0394+0.0368
Merck.6074±.0257 / .5731±.0441.5587±.0390 / .5104±.0307+0.0627
Hermite.7082±.0132 / .6392±.0358.6204±.0556 / .5986±.0674+0.0406

Key finding. DrugPrism consistently outperforms EEE-T in both Pearson and Spearman correlations across all three benchmarks, with Spearman gains of +0.0368, +0.0627, and +0.0406 on JACS, Merck, and Hermite, respectively. Notably, these improvements persist even though EEE-T has the same number of trainable parameters, three independent geometric branches, and identical target-conditioned weighting mechanism.

These results directly address the reviewer's concern: the improvements cannot be explained solely by additional projection capacity or target-adaptive distance weighting. Rather, the controlled comparison provides empirical evidence that combining geometries of different curvatures contributes additional predictive value beyond an equally expressive Euclidean multi-branch counterpart.

AuthorsOpenReview ↗

Response to Reviewer caQH (2/4)

W1. Novelty and target-dependent geometry

We thank the reviewer for raising this important point. We would like to clarify that the central contribution of DrugPrism is not the weighted-sum operation itself, but identifying and addressing target-dependent geometric preferences in protein–ligand affinity ranking.

Observation and motivation. Our work began with an intriguing observation: despite using the same training data and similar backbone architectures, spherical LigUnity and hyperbolic HypSeek exhibit markedly different, sometimes reversed, performance across protein targets. This raises an important yet underexplored question: rather than assuming one geometry is universally optimal, could different protein targets favor different geometries for affinity ranking? This observation motivated DrugPrism.

A simple, principled solution. DrugPrism addresses this question by learning in a mixed-curvature product manifold with three complementary geometric components—spherical, Euclidean, and hyperbolic—and their target-specific weights. Intuitively, the three geometries act like basis directions in a coordinate system, while the learned weights serve as coordinates, allowing each target to express its own geometric preference. With a single component, metric rescaling merely changes distance magnitudes without altering ligand rankings. With three complementary components, however, adjusting their relative weights introduces additional degrees of freedom, enabling different targets to induce different rankings. This motivates our target-conditioned Riemannian product metric, whose squared geodesic distance is exactly the weighted sum of component squared distances. Thus, the simple weighted-sum form is not an ad hoc aggregation rule, but the exact realization of a structured family of target-adaptive comparison metrics, with geometric representations and metric weights learned jointly.

Evidence supporting the hypothesis. Our analyses suggest that the learned geometric preferences are not arbitrary. On Merck, the learned weights broadly agree with the target-wise strengths of spherical LigUnity and hyperbolic HypSeek. Across 91 InterPro families, 67% exhibit strong geometric preferences and 84% show low within-family variation, suggesting systematic preferences among related proteins. Crucially, our new parameter-matched EEE-T control isolates the contribution of mixed curvature: despite identical model capacity and target-conditioned weighting, DrugPrism improves Spearman by 0.0368/0.0627/0.0406 on JACS/Merck/Hermite. Furthermore, on the publicly available Uni-FEP benchmark[1] covering 389 protein targets, DrugPrism achieves 0.7793 Pearson and 0.7489 Spearman, outperforming HypSeek, Boltz-2, and LigUnity by 19.1%, 25.8%, and 9.7% in relative Spearman improvement, respectively, under a unified evaluation protocol.

In summary, DrugPrism's novelty lies in recognizing target-dependent geometric preferences as an important modeling problem, developing a principled yet simple mechanism to learn these preferences, and validating its effectiveness through controlled ablations and large-scale evaluation. The simplicity of the resulting scoring rule is intentional: it enables efficient and interpretable geometric adaptation rather than diminishing the underlying contribution.

[1] Rongfeng Zou, Liguo Wang, Xinyan Wang, Ye Ding, and Hang Zheng. Breaking Barriers in FEP Benchmarking: A Large-Scale Dataset Reflecting Real-World Drug Discovery Challenges. 2025.

AuthorsOpenReview ↗

Response to Reviewer caQH (3/4)

Q2. Few-shot generalization and metric-head-only adaptation

We thank the reviewer for this suggestion. We conducted two additional experiments to examine few-shot generalization across unseen targets and metric-head-only adaptation.

(1) Generalization across unseen targets. Beyond the original OpenBind target, we evaluated DrugPrism on five additional held-out ChEMBL targets, using PDB-derived binding pockets. For each target, we reserved 10 compounds for adaptation and the remainder for independent testing. We evaluated 0/2/3/5/8/10-shot settings, reporting median Spearman correlations and interquartile ranges (Q25–Q75) across sampled adaptation subsets.

Target (PDB)Tune/TestZero-shot ρ5-shot ρ (Q25–Q75)10-shot ρ
DP2/CRTH2 (6D26)10/2030.12980.3777 (0.2236–0.4887)0.6781
Dopamine D3 (3PBL)10/1300.02120.2508 (0.1282–0.3528)0.5268
Dopamine D2 (6CM4)10/130−0.04390.4331 (0.2597–0.5121)0.6330
5-HT2A (6A93)10/130−0.35600.3598 (0.2029–0.4540)0.6470
5-HT7 (7XTC)10/72−0.01510.4484 (0.3746–0.5057)0.6488
Macro-average—−0.05280.37400.6267

All five targets show substantial improvements, with macro-averaged Spearman increasing from −0.0528 (zero-shot) to 0.3740 (5-shot) and 0.6267 (10-shot). Every target exceeds 0.52 at 10 shots, supporting consistent few-shot adaptation beyond the original single-target experiment. We will include the full shot-dependent results and variability in the revision.

(2) Metric-head-only adaptation. On the original OpenBind target, we additionally fine-tuned only the target-conditioned metric head while freezing all projection layers. This improved Spearman from −0.1947 (zero-shot) to 0.1786, while projection-layer fine-tuning achieved 0.6057 under the same protocol.

These results illustrate the complementary roles of metric weighting and representation learning. As characterized by Theorem 4.4, with fixed representations, adjusting metric weights can reverse a ligand pair's ordering only when the geometric components disagree. Thus, metric-head-only adaptation can exploit existing geometric signals but cannot correct errors shared by all components. Fine-tuning the projection MLPs additionally refines the component representations and distances, enabling more flexible target-specific adaptation.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Revision Summary

Global Comment: Summary of Revisions

We sincerely thank all reviewers for their constructive feedback. We have substantially strengthened the revised manuscript, focusing particularly on clarifying our central contribution, isolating the sources of empirical improvements, and broadening the evaluation.

Clarifying our contribution. DrugPrism was motivated by a striking observation: despite sharing the same curated training data and similar pretrained backbones, spherical LigUnity and hyperbolic HypSeek exhibit markedly different, sometimes reversed, performance across protein targets. This raises a fundamental question: could different targets favor different geometric inductive biases rather than a single universal geometry? DrugPrism addresses this through a pocket-conditioned Riemannian product metric. Intuitively, the three geometries act like complementary basis directions, with learned metric weights serving as target-specific coordinates. Our contribution lies in this target-adaptive formulation and its theoretical and empirical investigation, rather than in mixed-curvature product spaces or weighted-sum distances themselves.

New controlled experiments addressing the central concerns. We introduce SEH-G (globally learned weights) and EEE-T (parameter-matched three-branch Euclidean control), evaluated across five independent seeds (Table 4). DrugPrism improves mean Spearman over SEH-G by +0.025–0.069 and over EEE-T by +0.037–0.063 across JACS, Merck, and Hermite. SEH-T and EEE-T share the same 128-dimensional representation budget and exactly 132,070,639 trainable parameters. Furthermore, explicit component-distance calibration leaves their Merck Spearman gap essentially unchanged (+0.1050 → +0.1056; Appendix F). Together, these controls provide stronger evidence for the distinct contributions of target conditioning and mixed-curvature geometry, beyond additional capacity or uncalibrated distance scales.

Broader evaluation and improved reproducibility. We extend evaluation to 389 protein targets on Uni-FEP, achieving 0.7793 Pearson / 0.7489 Spearman, with a +0.0660 Spearman improvement over the strongest evaluated baseline, LigUnity (Appendix E). Few-shot experiments on five additional ChEMBL targets improve macro-Spearman from −0.0528 (zero-shot) to 0.6267 (10-shot) (Table 2). We also substantially expand Appendix B with architectural configurations, numerical safeguards, optimization settings, and explicit clarification of dual-branch training versus pocket-only inference. Baseline provenance and evaluation protocols are further clarified.

Together, these revisions provide more rigorous controlled evidence, substantially broader empirical validation, and clearer implementation details, directly addressing the reviewers' principal concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.