acceptodds
Withdrawn from ICLR 2027

CSIL: A Chirality-Structural Interaction Learning Framework for Chiral Drug-Drug Interaction Prediction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Accurate prediction of drug-drug interactions (DDIs) is essential for drug safety and healthcare research. Many drugs are chiral, and their molecules can exist in non-superimposable mirror-image forms. These forms share the same molecular formula and atomic connectivity but differ in chiral configuration. When two chiral drugs interact, different configuration combinations may alter biological recognition and lead to different interaction outcomes. This motivates the study of Chiral Drug-Drug Interaction (CDDI) prediction. However, many existing DDI methods focus on molecular connectivity and substructures without explicitly considering chirality, which may limit their ability to capture differences caused by chiral configuration and accurately predict CDDIs. Moreover, their substructure-based explanations do not explicitly account for chirality-related information, making it difficult to explain how chiral configuration affects CDDIs. To overcome these limitations, we propose a Chirality-Structural Interaction Learning (CSIL) framework for CDDI prediction. CSIL integrates hierarchical motif representations with configuration-sensitive chirality representations and employs chirality-motif collaborative cross-attention to model their cross-drug dependencies. It further highlights interaction-relevant motifs and configuration-sensitive cues associated with CDDI predictions, providing fine-grained attributions from both structural and chirality-related perspectives. Extensive experiments on a CDDI dataset curated from DrugBank demonstrate that CSIL achieves the best overall rank against representative baselines across four settings. The source code and datasets are submitted as Supplementary Materials for Reproducibility.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Drug-drug interaction prediction, Molecular representation learning, Molecular chirality

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_YK8NOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CSIL for drug–drug interaction (DDI) event prediction between chiral drugs. The model combines a shared atom encoder, motif representations, and a chirality branch using explicit stereochemical descriptors. Representations are aggregated across conformers and candidate configurations, followed by cross-drug attention and classification. CSIL achieves the best reported mean in 22 of 24 task–metric combinations.

Critical strengths

Separating unseen forms of known parent structures from unseen-parent generalization is a useful direction for evaluating stereochemistry-aware DDI prediction. This could expose limitations that conventional drug-pair splits do not directly address. The chirality-branch ablation also shows improvements in several generalization metrics, supporting an incremental benefit within CSIL.

Critical weaknesses
  1. The evaluation provides limited direct evidence for configuration-dependent DDI prediction. The dataset is derived by selecting DrugBank interactions between chiral drugs, but this does not necessarily demonstrate that the benchmark tests differences caused by configuration. Neither the main text nor Appendix A.2 quantifies matched examples in which different configurations of the same parent interact with the same partner, particularly examples with different event labels. Performance on this subset is also not reported, leaving some uncertainty about how directly the evaluation supports the paper's central motivation.

  2. The chirality-specific contribution is difficult to assess without clearer feature definitions and ablation details. Section 4.2 identifies the categories of chirality descriptors, but I could not locate a complete specification of the initial atom features, bond/distance encodings, or the construction of the signed-volume and determinant features in the main text or appendix. Moreover, Eq. (4) introduces chirality-related role embeddings before the shared atom encoder. It is unclear whether these embeddings or other stereochemical inputs remain in the "w/o chirality branch" variant.

    Comparing Tables 1 and 2, the variant without the chirality branch achieves competitive performance to the best external baseline, based on reported means. The remaining architecture is therefore already competitive without the dedicated branch. It would be helpful to clarify whether the chirality branch can also improve other DDI models.

Questions for authors
  1. How many matched examples contain different fully specified configurations of the same parent paired with the same partner, and how often do their event labels differ? How does CSIL perform on these cases, particularly those with different event labels?

  2. What are the exact attributes, extraction tools, and encodings used for the initial atom and edge features? How are CIP assignments, signed chiral volumes, and determinant features constructed, including neighbor ordering and sign conventions? In the chirality-branch ablation, are the role embeddings in Eq. (4), any stereochemical atom features, and candidate-configuration processing retained?

  3. What makes motif-level modeling particularly useful for the proposed chirality mechanism? Table 2 gives a slightly better F-rank to "w/o motif" (4.396) than to "w/o chirality" (4.479), although the full model is better than both (2.021); this comparison alone establishes neither motif redundancy nor chirality-specific necessity.

  4. Does the model support an empty chiral-center set? If so, how are the chirality token and the retained feature in Eq. (12) handled? If results are available, how does CSIL behave on the full DrugBank dataset or excluded pairs, separating achiral–achiral from mixed chiral–achiral cases?

  5. The reference to Chawla (2025) appears to omit the coauthors. Could you please check the author list against the following record?

Chawla S, Gupta R, Jha SK, Shamim, Kashid S, Jha KT. Stereoisomerism in Chemistry and Drug Development: Optical, Geometrical, and Conformational Isomers. Med Chem. 2025 Oct 3. doi: 10.2174/0115734064366389250923044201. Epub ahead of print. PMID: 41051049.

5 comments

Thank you for your valuable comments.

Q1: Matched Stereoisomer Statistics and Performance

A: (1) Requiring the same parent, partner drug and #Drug1/#Drug2 position, CDDI has 4,903 matched pairs of fully specified configurations (9,690 records), of which 491 (10.0%) carry different event labels: 125 of 2,642 enantiomer pairs and 366 of 2,261 diastereomer pairs. Table 4 pools the two positions, hence its 18.1%. We have added this to Appendix A.2.

(2) Both members of a pair fall in the same fold, so neither label is seen during training; we repeated the five folds with a variant whose inputs no longer distinguish the two configurations ("blind"). On the 125 enantiomer pairs with different labels, CSIL gets both members right in 75.2% of pairs against 62.4% for the blind variant, and separates the two members in 76.8% of them against 1.2% of the 4,412 pairs whose labels agree. Diastereomer pairs show no such gap, so only enantiomers read as evidence on configuration, and a quarter of them are still missed. The per-subset results are reported in the new Appendix A.11 (Table 11), summarized at the end of Section 5.4.

Q2: Feature Construction and Chirality Ablation Details

A: (1) Input features. All features come from RDKit on heavy-atom graphs with ETKDGv3 conformers: 47-dim atom and 29-dim edge features over bonded atoms and the 16 nearest atoms within 4.5 Å. We have added Appendix A.6, cited in Section 4.2, which lists every feature with its dimension, encoding and RDKit function in Table 6(a)(b).

(2) Chirality descriptors. R/S labels come from RDKit's FindMolChiralCenters. For each center, three difference vectors over its CIP-ranked neighbors form a matrix MM with unit-length columns; det⁡M\det M flips sign under reflection and is positive for R in 98.9% of centers, and the signed chiral volume is v=s⋅min⁡(∣det⁡M∣,1)v = s\cdot\min(|\det M|, 1) with s=±1s=\pm1 for R/S. We have given the full construction in Appendix A.6, with all 20 descriptors in Table 6(c).

(3) Chirality-branch ablation. The variant removes only the branch (XχX^{\chi}, chiral cross-attention, gated aggregation, stereo-state embedding, chirality token) and keeps the role embeddings, stereochemical atom and edge features, and candidate configurations, so it measures the branch rather than stereochemical information as a whole. We have stated this in Section 5.4.

Q3: Motif-Level Contribution to Chirality Modeling

A: We have clarified that F-rank is the Friedman mean rank over 24 task-metric combinations, where a lower value indicates better overall performance rather than the importance of an individual component. Each variant in Table 2 removes one component and retains the rest. The motif and chirality branches are complementary, the former capturing structural context and the latter configuration-sensitive information, so the small F-rank gap between w/o motif and w/o chirality does not make either branch the more important one.

To separate their contributions, we have added component-wise addition experiments in Appendix A.10: starting from a common base model, each component is added independently. Both branches improve F1 on all four tasks, and the full model attains the highest F1 on every task and the best overall F-rank, which further supports the complementary contributions of motif and chirality modeling to CDDI prediction.

Q4: Empty Chiral-Center Handling and Non-Chiral DDI Evaluation

A: CSIL needs no architectural change for an empty chiral-center set. With a clamped normalizer, the gated aggregation returns a zero chirality vector in Eq. (9). The chirality token fed to Eq. (12) is the layer norm of that vector plus a learned embedding of the drug's chirality state, which the original text omitted. For a drug with no chiral centers the token therefore reduces to a constant, the layer norm of the embedding for the unspecified state. Achiral drugs are thus unmasked and distinguished by their motif tokens.

Re-adding their interactions (full DrugBank: 1,136,882 DDIs, 3,604 drugs), we compared CSIL with MSHGCL, the best-ranked baseline in Table 1, under Task 1 with five-fold CV on identical folds. CSIL exceeds 97.9% accuracy in every pair group and outperforms MSHGCL on all metrics for chiral–chiral pairs; on mixed and achiral–achiral pairs the two models are within one standard deviation. The per-group accuracy and F1 are reported in the new Appendix A.7 (Table 7).

We have defined the chirality state (Section 3.2) and the chirality token (Section 4.3), described the empty-set case (Section 4.2), rewritten Eq. (12) with the chirality token (Section 4.4), and summarized the full DrugBank results at the end of Section 5.2.

Q5: Correction of Reference Author List

A: We apologize for the incomplete author information. We have checked the author list against the provided record and corrected the reference to list all coauthors in the revised PDF. We thank the reviewer for catching this.

Thank you for the response. I still have two concerns: (1) If the blind inputs no longer distinguish the two configurations, how does this variant achieve a pair accuracy of 0.624 on enantiomer pairs with different labels? (2) I understand that motif and chirality representations can be complementary. However, the reported gains could also be interpreted as benefits of a more general "motif-structural interaction learning framework." The current evidence does not yet clearly establish the distinctive benefit of the proposed chirality module.

Thank you for your response and valuable insights.

Q: The non-zero score of the configuration-blind variant; the distinctive benefit of the chirality module

A (1). Our wording was imprecise and we have corrected it. The variant merges the CIP R/S one-hots and zeroes the signed chirality channels, but it does not give the two members identical inputs, because each drug's conformers are embedded independently with a seed derived from its own identifier: even the best-matching conformer pair differs in pairwise heavy-atom distances by up to 2.42 Å. These residuals are parity-even and carry no R/S information, but they identify the molecule, and each member is seen in training against many other partners, so 0.624 is a memorisation floor rather than configuration reading. The bound of zero we had in mind holds for exactly mirrored inputs, which we now verify: under a parity transform of one member's own conformers the blinded inputs are bit-identical, so the variant returns the same logits for a molecule and its mirror image for any weights, whereas CSIL changes its predicted class in 33 of 152 cases (Figure 3d). Our conclusion rests on the paired contrast on the same pairs (+0.128 pair accuracy, p = 0.001) and on CSIL discriminating 0.768 of the label-discordant pairs against 0.012 of the label-concordant ones, not on the ablated arm's absolute level.

A (2). The ablated variant is precisely this control, which we had not made clear: it keeps the motif branch, the chiral branch, the cross-attention, the gated aggregation and the chirality token at an identical parameter count and training protocol, neutralising only the configuration-bearing inputs, unlike the w/o chirality row of Table 2, which also removes the module. The two arms are indistinguishable in overall accuracy (0.970 vs 0.972), so the general framework accounts for the general classification performance, while the entire difference is confined to the 125 enantiomer pairs whose labels differ (+12.8 points) and is absent on diastereomer pairs, whose configurations are separated by reflection-invariant distances. A general-framework benefit cannot be selective in this way, and Appendix A.10 adds the converse, where the chirality branch alone improves F1 on all four tasks. We have tightened Section 5.4 accordingly: the module is not claimed to explain the overall margin over the baselines, but to supply configuration sensitivity the motif framework alone does not provide. The module operates on chiral-centre descriptors and a drug-level token, so it is not specific to this backbone.

We have revised Appendix A.11 to define the variant channel by channel, to state that it is capacity-matched, to explain its non-zero score on enantiomer pairs as a memorisation floor and to report the parity-transform check, and Section 5.4 to use the same description of the variant and to attribute the overall classification performance to the general framework and the configuration-dependent behaviour to the chirality module.

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

This paper proposes CSIL, a framework for chiral drug-drug interaction prediction. The model combines molecular structure and chirality information by learning representations from different molecular conformers and configurations, followed by attention mechanisms to capture interactions between drugs. The authors construct a CDDI dataset from DrugBank and evaluate CSIL across four prediction settings, reporting competitive performance against existing methods.

Critical strengths

I appreciate the authors' attention to molecular chirality in DDI prediction.

Different stereoisomers may exhibit distinct pharmacological behaviors despite sharing the same molecular connectivity.

I consider this problem formulation the most valuable aspect of the paper.

Critical weaknesses
  1. The paper does not report how frequently stereoisomers with identical molecular connectivity exhibit different DDI event labels when paired with the same drug. This information is important for establishing the relevance of the proposed CDDI benchmark. Moreover, although Task 2 evaluates unseen chiral forms, it does not directly assess whether CSIL can correctly distinguish interactions whose labels differ across stereoisomers.

  2. The proposed framework largely follows conventional DDI modeling pipelines. Its main components, including motif-level attention, hierarchical pooling, and cross-drug attention, are established techniques. In particular, the proposed chirality-motif collaborative cross-attention appears to be standard bidirectional cross-attention applied to motif and chirality tokens (Eq. 10). The methodological contribution beyond incorporating explicit chirality features into an existing DDI architecture is not sufficiently clear.

Questions for authors
  1. Among stereoisomers sharing the same molecular connectivity, what proportion exhibit different annotated DDI event labels when paired with the same drug? Could the authors provide the corresponding dataset statistics?

  2. Could the authors evaluate CSIL specifically on matched stereoisomer drug pairs with different DDI labels? Such an analysis would help determine whether the model captures chirality-dependent differences rather than primarily improving general DDI classification.

  3. Why are the Vi×Vi and Vi×Vp / Vp×Vi settings excluded? Are there insufficient samples, or is there another experimental consideration?

3 comments

Thank you for your valuable comments.

Q1: Stereoisomer Label Discrepancy Statistics

A: We thank the reviewer for this question. We pair each fully specified drug did_i with every drug djd_j of the same parent structure but a different configuration, and compare their event labels with each shared partner drug. Overall, 20.6% of these comparisons carry different labels. The results by the chirality status of djd_j are:

Status of djd_jIsomer pairsDifferent labels
fully specified2518.1%
inferred-racemate3819.2%
partial349.1%
unspecified0–
Total6620.6%

We have added these statistics to Table 4 and described them in the "Chirality status" paragraph of Appendix A.2.

Q2: Chirality-Dependent DDI Evaluation

A: We evaluate the 491 matched pairs whose event labels differ (125 enantiomer, 366 diastereomer) under a dedicated five-fold split that places both members of a pair in the test set of the same fold, so neither of their labels is seen during training and the prediction for one cannot be copied from the other; the 4,412 pairs whose labels agree are held out the same way as a control. We repeated the five folds with a variant of the same model whose inputs no longer distinguish the two configurations. A pair is correct only when both members are; discrimination means the two members receive different predictions-required when the labels differ, an error when they agree.

SubsetPairsPair acc. CSILBlindDiscrim. CSILBlind
Enantiomer, labels differ1250.7520.6240.7680.648
Diastereomer, labels differ3660.8330.8330.8690.861
Enantiomer, labels agree2,5170.9610.9700.0120.006
Diastereomer, labels agree1,8950.9860.9870.0120.009

Overall five-fold accuracy on this split: 0.9702 ± 0.0098 (CSIL) vs. 0.9715 ± 0.0070 (blind).

The two arms are equally accurate overall, so the blind variant is not a weaker DDI classifier; the difference appears only on the 125 enantiomer pairs whose labels differ, where CSIL gains 12.8 points in pair accuracy and is ahead in 6 of the 7 parent families. Within CSIL, discrimination is 0.768 there but 0.012 on the label-concordant pairs, so it separates the members it should while predicting identically elsewhere-which a uniform gain in classification accuracy cannot produce. Diastereomer pairs show no such difference, so only enantiomers are read as evidence on configuration, and a quarter of them are still missed. Added as Appendix A.11 (Table 11), with the result at the end of Section 5.4.

Q3: Excluded Evaluation Settings

A:We thank the reviewer for raising this. We designed the evaluation as four generalization scenarios rather than as an enumeration of the ordered combinations of Vk, Vi, and Vp in Eq. (2), and these three combinations were simply not considered when the tasks were defined. Their absence is an oversight in the task design, not a restriction of CSIL.

We have now evaluated CSIL on all three settings, with no change to the model, over five folds and one seed. Their splits are added to Table 3, and the results to Table 8 of the new Appendix A.8. Macro-F1 is 0.5281 on Vi×Vi, and 0.3688 and 0.4157 on Vi×Vp and Vp×Vi.

Table 8 reports CSIL only, as the baseline runs did not finish in time for this response. We will report all comparison methods on the three settings in the camera-ready version.

I appreciate the detailed response and additional experiments. After reviewing the discussion and other reviews, I will keep my original score.

Thank you for your time and careful consideration of our response and additional experiments. We respect your decision. We would also like to ask whether any concerns remain unclear or have not been fully addressed. If so, we would be happy to provide further clarification or additional explanations.

Reviewer_m8K7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes CSIL, a framework for predicting chiral drug–drug interactions (CDDI) in which stereochemical configuration can influence interaction outcomes. It constructs a CDDI dataset from DrugBank and learns hierarchical motif representations together with configuration-sensitive chirality representations derived from molecular conformers. A pair-aware cross-attention mechanism models dependencies between the structural and chirality-related features of the two drugs. The method is evaluated under four settings that probe different degrees of generalization (known–known, unseen chiral forms of known parents, unseen parent structures, and unseen–unseen parents) and reports the best Friedman mean rank among the compared DDI models. Attention-based analyses are used to examine motif importance, particularly for warfarin enantiomers.

Critical strengths
  1. The problem is important and timely. Most structure-based DDI models rely on molecular connectivity and substructures that can be identical across different chiral configurations, so explicitly addressing chirality is a valuable direction for drug-safety applications.
  2. The representation design is chemically motivated. Explicit stereochemical descriptors (including CIP assignments and geometric features) are combined with motif information in separate branches, which is a reasonable way to preserve both structural and configuration-sensitive signals.
  3. The evaluation goes beyond a single random split and examines several forms of generalization, including to unseen chiral forms and unseen parent structures. This provides a more informative test of whether the method can handle novel stereochemical cases.
Critical weaknesses
  1. The overall improvement is modest and not fully consistent across metrics and tasks. Some simplified ablation variants outperform the full model on individual metrics, which weakens the claim that every component is necessary. Relying primarily on the aggregate F-rank can obscure this variation and complicates assessing which design choices are most responsible for the gains.
  2. The novelty relative to existing stereochemistry-aware DDI methods is not fully established. The related work discusses approaches such as DSMV-DDI that incorporate stereochemical information from three-dimensional conformers, yet these methods are not included in the main experimental comparison. Without direct comparisons under the same data splits and protocol, it remains difficult to determine whether the explicit chirality modeling is the primary source of the observed gains.
  3. The interpretability analysis is based mainly on one drug family (warfarin) and attention patterns. While the overlap with known metabolism-associated regions is interesting, attention weights do not establish that the highlighted motifs are causally responsible for the predicted interaction. A limited case study can illustrate model behavior but does not constitute strong evidence for a mechanistic claim.
Questions for authors
  1. What is the primary performance metric, and are the reported improvements statistically reliable across random seeds or across different groups of drugs?
  2. How does CSIL compare with existing stereochemistry-aware DDI models (for example, those discussed in the related work) under the same data splits and training budget?
  3. Can counterfactual or motif-deletion experiments demonstrate that the reported attention differences are related to the chiral interaction outcome?
1 comment

Thank you for your valuable comments.

Q1: Primary Metric and Statistical Reliability

A: The primary metric is macro-F1. the 82 event classes are highly imbalanced, so accuracy and the micro-averaged AUPR and AUC are dominated by frequent events, whereas macro-F1 weights every event equally. Section 5.1 now states this and how each metric is averaged. To assess reliability across random seeds, we ran CSIL with four additional seeds on each task in addition to the run reported in Table 1, keeping the five outer folds fixed. Across the five runs, the standard deviation of macro-F1 is at most 0.011, and a two-way ANOVA (seed × fold) finds no significant seed effect on any task. The macro-F1 of CSIL exceeds the best baseline in Table 1 in 19 of the 20 runs, with mean margins of 0.011-0.021, and a one-sided one-sample t-test of the five runs against the best baseline value gives p < 0.001, 0.030, 0.034, and < 0.001 on Tasks 1-4. We have revised Section 5.2 accordingly and report all runs in Appendix A.9 (Table 9).

Q2: Comparison with Stereochemistry-Aware Baselines

A: We thank the reviewer for this comment. We have added the two stereochemistry-aware methods from our related work, LSA-DDI and DSMV-DDI, re-trained under the same splits, seeds and training budget as CSIL. Results are in Table 1 and their designs are summarized in Table 5. CSIL outperforms both on all 24 task-metric combinations.

Values are reported as Accuracy / Macro-F1.

TaskLSA-DDIDSMV-DDICSIL (ours)
Task 10.9121 / 0.89360.9723 / 0.94270.9758 / 0.9516
Task 20.6026 / 0.60490.6610 / 0.61340.7456 / 0.7199
Task 30.5669 / 0.51640.6320 / 0.53650.6569 / 0.5659
Task 40.3673 / 0.24220.3328 / 0.23320.4290 / 0.3181

Our LSA-DDI numbers are below those in its original paper because it was evaluated there on a general DrugBank benchmark, whereas our CDDI benchmark covers only chirality-relevant pairs under different splits; this reflects the change of dataset, not an insufficient reproduction.

Q3: Counterfactual Analysis

A: A counterfactual experiment does demonstrate the relation, and we have added it to the paper; the question also made us realise that the counterfactual control already built into our analysis was never stated. The R and S values in Figure 3b each average both anchor reference ensembles, so their difference equals the mean of two within-molecule mirror contrasts, attention(R) - attention(S) = 1/2 [(R - mirror of R) + (mirror of S - S)]. Each mirror is an exact parity transformation of the encoder inputs: distances, connectivity and all parity-even channels are identical and only the parity-odd quantities change sign, so conformational differences between the two drugs cancel by construction. Section 5.3, which previously said only that connectivity and distances are kept comparable, now states this. Figure 3a-c, however, report only where attention falls, so the new Figure 3d reports the outcome pair by pair on the same 152 co-medications. Inverting the chiral center alone shifts the target-class logit by 0.735, with 119/152 pairs sharing the sign of the mean, and changes the predicted top-1 interaction class in 33/152 pairs; a feature-midpoint input, which carries no R/S information by construction and so cannot show a configuration effect, gives only 0.129 and 6/152. Therefore, the attention difference of Figure 3b and the predicted interaction outcome change under one and the same intervention.

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

Related papers

Discussion (0)

Sign in to comment.