acceptodds
Under review as a conference paper at ICLR 2027

PAFDTA: Progressive Adaptive Fusion of Heterogeneous Representations with Cross-Branch Contrastive Alignment for Drug-Target Affinity Prediction

Abstract

Drug-target affinity (DTA) prediction depends on learning representations that preserve complementary information from molecular structure, protein sequence, pair-specific interactions, and relational context. We propose a heterogeneous representation-learning framework, namely PAFDTA, combining graph-based molecular features, pretrained protein representations, drug-target pair context, and topology-aware relational priors. Multi-scale Local-Global Aggregation captures complementary molecular patterns, while Cross-Branch Contrastive Alignment uses relational context to regularize the molecular latent representation before fusion. Progressive Adaptive Fusion then coordinates the molecular, protein, and relational representations through pair-conditioned weighting and relation-level interaction modeling, while the joint drug-target representation is retained as a direct pair-context pathway for affinity regression. We evaluate PAFDTA on Davis and KIBA using five-fold warm evaluation, component ablations, and cold-start settings with unseen drugs and/or targets. Additional analyses examine prediction consistency, uncertainty-aware candidate prioritization, structural interpretation of learned protein representations, and a BRAF-oriented repositioning case. The results support heterogeneous representation learning and adaptive representation integration as effective principles for DTA prediction.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Drug-Target Affinity Prediction, Molecular Representation Learning, Multi-Modal Fusion, Contrastive Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_bnguOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PAFDTA, a multi-source representation learning framework for drug–target affinity prediction. It combines drug graphs, protein sequences, and drug–target relationships using contrastive alignment and adaptive fusion. The authors evaluate the framework on the Davis and KIBA benchmark datasets and conduct ablation studies and case studies. The results show that the framework achieves competitive performance in affinity prediction. However, its main methodological contribution appears to be the integration and adaptation of existing representation learning, contrastive learning, and fusion techniques. Overall, the paper presents a comprehensive DTA framework, but the novelty and specific contributions of its components need to be better demonstrated.

Critical strengths

The paper integrates molecular graph features, pretrained protein representations, drug-target pair context, and relational topology to capture complementary information for affinity prediction.

Critical weaknesses

1.The research motivation is unclear, as it does not specify what concrete limitation of existing DTA methods PAFDTA addresses. 2. No code is released, making key implementation details difficult to verify. 3. The methodological novelty is limited, since the framework mainly combines existing techniques such as GIN, ESM2, Transformer pooling, Affinity2Vec meta-paths, and gated fusion, without clearly articulating the core novel contribution. 4. PAFDTA does not outperform baselines on most metrics. 5. The paper does not clarify whether the experimental settings and data splits are consistent with those of the baselines. 6. The task is conventional and the metrics are standard, so the empirical contribution may need stronger justification.

Questions for authors
  1. The results for the comparison methods in Table 1 are all taken from their respective published papers. The authors are requested to clarify whether these methods employed the same machine training and data partitioning as PAFDTA.
  2. For proteins longer than 1,024 residues, how are the residue embeddings truncated or selected for the downstream model after ESM2 feature extraction?
  3. Why do removing the topological prior and CBCA improve CI over the full model, and how do the authors justify the contributions of these components? 4.What value of the temperature parameter τ is used in Eq. (5)?
  4. It remains unclear whether the performance gains stem from the fusion strategy itself or simply from the additional input information.
  5. The data partitioning might be flawed, the training and test sets could contain the same proteins or small molecules.
  6. The task addressed in this paper is relatively conventional within the DTA field. The evaluation metrics adopted are fairly standard and straightforward.
  7. In the KIBA double-cold setting, PAFDTA achieves a higher CI than LLMDTA, but its MSE and R² are substantially worse. How do the authors reconcile this with their claims about generalization to unseen drugs and targets?
4 comments
AuthorsOpenReview ↗

Response to Reviewer bngu

We sincerely thank the reviewer for the constructive feedback and for recognizing the integration of molecular, protein, pair-context, and relational representations in PAFDTA. We would like to clarify several aspects of the experimental protocol and the specific contributions of the framework.

Q1. Were the baselines trained using the same machine and data partitioning as PAFDTA?

The baseline results in Table 1 are taken from their respective publications, following the reporting convention of previous DTA benchmark studies. They were not all re-trained under an identical hardware and preprocessing environment. We therefore interpret Table 1 as a literature-based comparison rather than a strictly controlled common-protocol comparison.

Our own PAFDTA results are obtained using five-fold pair-wise cross-validation with a fixed seed of 42, a 200-epoch schedule, and final-epoch evaluation. The five test folds are mutually exclusive and collectively cover every observed drug-target pair exactly once.

Q2. How are proteins longer than 1,024 residues processed?

Protein features are extracted using the frozen ESM2-t30-150M-UR50D model, producing 640-dimensional residue embeddings. During feature extraction, long sequences are processed in windows to accommodate the ESM2 token limit. The downstream DTA model receives at most 1,024 residue embeddings per sample, which are projected into the model's 192-dimensional hidden space.

Q3. Why can removing CBCA or the relational prior improve CI? What is the contrastive temperature?

CBCA and the relational prior contribute to the model through representation regularization and contextual information, respectively. Their contribution is not limited to maximizing pairwise ranking.

As shown in Table 3, the Full model achieves the lowest MSE (0.2018) and highest R² (0.7376). Removing CBCA increases MSE to 0.2045, while removing the relational prior increases MSE to 0.2051. Both variants also have lower R² than Full.

The contrastive temperature in Equation 5 is τ = 0.07.

Q4. Do the performance gains come from fusion itself or from additional input information?

Our Sum and Concat experiments retain the representation branches while changing the fusion operation. Compared with these fixed alternatives, PAF achieves better numerical regression performance:

PAF: MSE = 0.2018, R² = 0.7376.

Sum: MSE = 0.2084, R² = 0.7313.

Concat: MSE = 0.2064, R² = 0.7339.

This controlled comparison supports the value of the learned progressive fusion mechanism beyond simply providing multiple input representations.

Q5. Could training and test sets contain the same drugs or proteins?

Our primary warm-start evaluation uses pair-wise partitioning. The five held-out folds are mutually exclusive and cover all observed interactions exactly once. Therefore, no identical drug-target pair occurs in both the training and its corresponding test set.

Drug or target identities may recur across training and test sets in warm-start evaluation, since the purpose is to predict unseen interactions involving known entities. In the separate KIBA cold-start experiments, drug and/or target identities are explicitly held out.

Q6. What additional value does PAFDTA provide beyond conventional DTA prediction?

In addition to affinity regression, we investigate uncertainty-aware candidate prioritization through split-conformal calibration.

We use 5,010 calibration pairs and a separate 5,010-pair test set. At a nominal 90% coverage level, the empirical coverage reaches 89.5%.

The screening pipeline progressively reduces the test candidates from 5,010 to 365, then 347, and finally 166 conformal-trusted candidates. Among the final 166 candidates, 154 are active, giving 92.8% precision and 36.9% recall. The selected set has approximately 11.1 times the active-pair rate of the original test set.

This extends the framework beyond point-wise affinity prediction toward uncertainty-aware candidate selection.

Q7. How do you interpret the KIBA double-cold results?

We distinguish ranking performance from numerical regression performance. CI evaluates whether the relative ordering of affinity pairs is preserved, while MSE and R² assess numerical prediction accuracy.

PAFDTA achieves the highest CI among the compared methods in all three KIBA cold-start settings:

Drug-cold: 0.791. Target-cold: 0.751. Double-cold: 0.687.

The drug-cold setting also achieves the best MSE and R² among the compared methods. Our interpretation is therefore that PAFDTA demonstrates competitive ranking transfer under entity-level cold-start conditions, rather than claiming universal superiority on every metric in every setting.

We hope these clarifications better communicate the motivation, evaluation protocol, and practical contribution of PAFDTA.

Reviewer_bnguOpenReview ↗

Follow-up Comments to the Authors

Thank you for the detailed rebuttal. These responses address several of my technical questions. However, some of my main concerns remain unresolved.

Q1. Were the baselines trained using the same machine and data partitioning as PAFDTA?

The authors confirm that the baseline results in Table 1 were taken from different publications rather than reproduced under a unified experimental protocol. While this clarification is helpful, it also confirms that the reported comparisons may be affected by differences in data partitioning, preprocessing, and training configurations. Could the authors clarify whether the published baselines used exactly the same Davis and KIBA splits as PAFDTA? Ideally, a controlled comparison with representative baselines under identical data splits would be needed to substantiate the claimed performance improvements.

Q2. How are proteins longer than 1,024 residues processed?

The explanation of window-based ESM2 feature extraction is helpful. However, the authors have not specified how the final 1,024 residue embeddings are selected when a protein exceeds this length. Please clarify the selection or aggregation strategy. I also note that the original concern regarding code availability has not been addressed. A public or anonymous implementation, including preprocessing and data-splitting scripts, would substantially improve reproducibility.

Q4. Do the performance gains come from fusion itself or from additional input information?

The Sum and Concat ablations provide useful evidence that the proposed fusion mechanism improves regression performance over simple aggregation. However, they do not fully establish the novelty of the overall framework, which integrates several established representation learning components. I would appreciate a clearer explanation of what constitutes the central methodological innovation beyond combining existing techniques. In addition, the improvements in MSE and R² appear relatively small. Are these differences consistent across multiple random seeds or statistically significant?

Q5. Could training and test sets contain the same drugs or proteins?

I appreciate the distinction between ranking performance and numerical prediction accuracy. Nevertheless, the higher CI but substantially worse MSE and R² in the KIBA double-cold setting remains a concern. Since the primary task is affinity regression, strong ranking performance alone does not establish superior generalization to unseen drugs and targets. Could the authors provide a more detailed analysis of the numerical prediction errors in this setting, including whether the model exhibits systematic bias or calibration issues?

AuthorsOpenReview ↗

Response to Reviewer bngu's Follow-up Comments— Part 1 (Q1–Q2)

We thank the reviewer for the careful follow-up and appreciate the opportunity to clarify the experimental protocol and implementation details.

Q1. Baseline comparisons and data partitioning

The baseline results in Table 1 were taken from the respective publications. Therefore, we do not claim that every baseline was trained on exactly the same hardware or identical fold assignments as PAFDTA. Table 1 provides a comparison with representative published methods, rather than a fully controlled reproduction of all competing approaches.

For PAFDTA, we used five-fold pairwise cross-validation with a fixed random seed of 42 on Davis and KIBA. The five held-out test folds are mutually disjoint and collectively cover all observed drug–target pairs, with no pair appearing in both the training and test partitions of the same fold.

Importantly, the relational path features are constructed using the training interactions of the corresponding fold, rather than the held-out interaction labels. This preserves the separation between training and test interactions when constructing the relational representations.

We also provide controlled architectural comparisons in Table 3, where the model variants are evaluated under the same data partition and input setting. These experiments are intended to isolate the contribution of the proposed components, independently of the cross-publication comparisons in Table 1.

We agree that reproducing additional published baselines under identical partitions would provide a stronger basis for direct model-to-model comparisons. Accordingly, we interpret Table 1 as evidence of competitive performance relative to published results, while distinguishing it from the controlled comparisons within PAFDTA.

Q2. Processing proteins longer than 1,024 residues and implementation availability

We thank the reviewer for requesting a more precise explanation.

PAFDTA uses frozen ESM2-t30-150M-UR50D to extract 640-dimensional residue-level representations. For protein sequences exceeding the ESM2 window limit, feature extraction is performed using consecutive, non-overlapping windows of 1,022 residues. The resulting embeddings are concatenated in their original sequence order to reconstruct the residue-level representation.

The downstream model applies a maximum protein length of 1,024 residues. Specifically, when a protein exceeds this length, we retain the first 1,024 residue embeddings from the N-terminus. No random residue sampling, center cropping, or full-sequence pooling is used for this truncation. Shorter sequences are padded with corresponding masks.

Thus, the 1,022-residue window length concerns ESM2 feature extraction, whereas the 1,024-residue limit controls the computational cost of downstream sequence processing. These are two distinct operations.

Regarding reproducibility, the anonymized supplementary materials contain the reference model and training implementation, configuration files, dataset resources, fold files, and fold-specific relational features. We agree that complete preprocessing and exact split-generation scripts would further facilitate end-to-end reproduction, and appreciate the reviewer's emphasis on this point.

AuthorsOpenReview ↗

Response to Reviewer bngu's Follow-up Comments— Part 2 (Q3–Q4)

We appreciate the reviewer's further questions regarding the methodological contribution and the interpretation of numerical prediction performance.

Q4. Methodological novelty and the contribution of fusion

We agree that individual representation-learning components such as GIN, ESM2, and Transformer encoders are established techniques. The central contribution of PAFDTA is not the invention of these components independently, but their coordinated learning through topology-guided cross-branch alignment and progressive interaction modeling.

Regarding the contribution of fusion itself, the Sum and Concat variants in Table 3 retain the underlying representation branches while replacing the proposed fusion strategy. PAFDTA achieves an MSE of 0.2018 and an R² of 0.7376, compared with 0.2084/0.7313 for Sum and 0.2064/0.7339 for Concat. These results indicate improved numerical regression performance under the evaluated setting without introducing additional representation sources.

We additionally performed targeted single-fold experiments examining the two fusion stages. Under the same Davis fold-0 protocol, the full model achieved an MSE of 0.2078, compared with 0.2185 when Stage-I weights were fixed uniformly, 0.2148 when Stage-II expert weights were fixed uniformly, and 0.2182 when the multiplicative expert was removed. These results provide further descriptive evidence concerning the roles of the two stages.

We distinguish these controlled observations from statistical significance across independent training runs. The fusion-specific comparisons have not established multi-seed statistical significance, and test-pair bootstrap uncertainty should not be interpreted as training-seed variability. Our methodological claim therefore concerns the coordinated alignment and integration framework, supported by the reported architectural comparisons, rather than universal superiority of an individual fusion operator.

Q5. Numerical prediction errors in the KIBA double-cold setting

We thank the reviewer for emphasizing numerical affinity prediction. We agree that CI alone is insufficient to characterize regression generalization, and therefore conducted an additional analysis of the 4,637 KIBA double-cold test pairs.

To assess whether the model retains meaningful numerical predictive information under simultaneous drug and target novelty, we additionally compared it with a constant predictor based on the training-set mean affinity. This reference predictor obtains an MSE of 0.6737 and an MAE of 0.5596. PAFDTA reduces these errors by 19.3% and 8.8%, respectively.

This comparison demonstrates that the model provides nontrivial numerical predictions beyond a constant reference, even when both entity types are unseen during training.

We also examined the residual distribution to investigate potential systematic bias. The overall mean signed error, defined as prediction minus ground truth, is −0.0752, while the median absolute error is 0.3467. The standard deviation of predicted affinities is 0.7225, compared with 0.8185 for the ground-truth values.

The errors are not distributed uniformly across affinity ranges. In the middle affinity ranges, the mean absolute errors are approximately 0.35, whereas the highest-affinity quintile has an MAE of 0.9584 and a mean signed error of −0.7065. This indicates a tendency toward prediction-range compression, particularly underestimating high-affinity interactions.

These observations help explain why the model retains useful pairwise ranking performance while its absolute regression accuracy does not reach the best competing result in the double-cold setting. The residual analysis is descriptive and does not constitute a formal guarantee of probabilistic calibration.

Overall, the additional analysis establishes that PAFDTA retains meaningful numerical predictive information under simultaneous drug and target novelty, while identifying high-affinity estimation as the primary source of remaining regression error.

We appreciate the reviewer's constructive comments and hope these additional implementation details and quantitative analyses clarify the contribution and empirical behavior of PAFDTA.

Reviewer_KfSBOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper introduces PAFDTA, a model that predicts drug–target binding affinity by combining four sources of information 1) Molecular structure: a graph neural network (GIN), augmented with multi-scale local–global aggregation 2) Protein sequence: frozen ESM2 pretrained residue embeddings and Transformer encoders, 3) Drug–target pair context: joint attention-based processing of molecular and protein tokens, and 4) Relational information: drug–drug, target–target, and drug–target relations summarized through meta-path features.

It proposes Cross-Branch Contrastive Alignment (CBCA), which aligns molecular and relational representations, and Progressive Adaptive Fusion (PAF), which combines the different representations using learned weights and interaction experts. The model is evaluated on Davis and KIBA, including conventional pair-wise evaluation, cold-start tests on KIBA, ablations, uncertainty-aware screening, and a BRAF docking case study.

My biggest concerns are:

  1. lack of novelty- most papers are using a hybrid of sequence transformers on smiley codes, GCNs, convolutional and dense layers, DD,TT and DT correlation matrices, etc. with hybrid combination. There is a flood of papers that evaluate the performance on KIBA and DAVIS with CIs in recent papers exceeding 0.9 on warm splits. Some papers are listed below:

MGraphDTA: https://pubs.rsc.org/sc/article/13/3/816/746808/MGraphDTA-deep-multiscale-graph-neural-network-for DgraphDTA: https://pubs.rsc.org/ra/article/10/35/20701/701287/Drug-target-affinity-prediction-using-graph-neural HiSIF-DTA: https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=10323144 GTAMP-DTA: https://www.sciencedirect.com/science/article/pii/S1476927123001731#sec0050 EMHGCL-DTA: https://ieeexplore.ieee.org/abstract/document/11557120 NHGNN-DTA: https://academic.oup.com/bioinformatics/article-abstract/39/6/btad355/7186502 Cold-DTA: https://www.sciencedirect.com/science/article/pii/S0010482523008375 Fusion-DTA: https://academic.oup.com/bib/article/23/1/bbab506/6470967 DMFF-DTA: https://www.nature.com/articles/s41746-025-01464-x TransVAE-DTA: https://www.sciencedirect.com/science/article/pii/S0169260723006697 MolTrans: https://academic.oup.com/bioinformatics/article/37/6/830/5929692 Perceiver CPI: https://academic.oup.com/bioinformatics/article/39/1/btac731/6842322 GraphCL-DTA: https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=10382475

  1. The reported CIs on warm split are lower compared to most of the recent papers. Same is the case with the cold split scenario.

  2. The paper includes more analyses, but reproducibility, statistical reliability, stronger generalization tests, and evidence for the proposed components are inadequate.

  3. Overall, I would say it is a useful engineering combination, but its stronger implications about representation learning, generalization, and practical drug discovery are not yet demonstrated convincingly.

Critical strengths
  1. CBCA trains the molecular encoder to match the pair’s relational prior, whereas earlier DTA contrastive methods match a molecule against altered copies of itself; conformal calibration is rarely used in DTA, and PAFDTA turns candidate selection into a calibrated decision rather than a plain ranking.

  2. The warm protocol is specified precisely: five-fold pairwise CV with seed 42, mutually exclusive folds covering every observed pair once. Relational priors are reconstructed per fold from that fold’s training interactions (Section 3.1.3, Appendix C), which is the correct instinct.

  3. Implementation details are reported usefully: D = 192, six heads, Adam at 5e-5, dropout 0.25, 200 epochs, β = 0.01, λ= 0.05, and frozen ESM2-t30-150M-UR50D capped at 1024 residues.

  4. Appendix E provides a robustness test, and the case study appropriately frames docking poses as hypotheses rather than evidence.

Critical weaknesses
  1. First of all, there are a number of methods for DTA prediction. See the list below. Many have not been compared with in tables

MGraphDTA: https://pubs.rsc.org/sc/article/13/3/816/746808/MGraphDTA-deep-multiscale-graph-neural-network-for DgraphDTA: https://pubs.rsc.org/ra/article/10/35/20701/701287/Drug-target-affinity-prediction-using-graph-neural HiSIF-DTA: https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=10323144 GTAMP-DTA: https://www.sciencedirect.com/science/article/pii/S1476927123001731#sec0050 EMHGCL-DTA: https://ieeexplore.ieee.org/abstract/document/11557120 NHGNN-DTA: https://academic.oup.com/bioinformatics/article-abstract/39/6/btad355/7186502 Cold-DTA: https://www.sciencedirect.com/science/article/pii/S0010482523008375 Fusion-DTA: https://academic.oup.com/bib/article/23/1/bbab506/6470967 DMFF-DTA: https://www.nature.com/articles/s41746-025-01464-x TransVAE-DTA: https://www.sciencedirect.com/science/article/pii/S0169260723006697 MolTrans: https://academic.oup.com/bioinformatics/article/37/6/830/5929692 Perceiver CPI: https://academic.oup.com/bioinformatics/article/39/1/btac731/6842322 GraphCL-DTA: https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=10382475

  1. Table 1 claims the lowest MSE on both datasets, but this holds only over the six selected baselines. MGraphDTA reports 0.128 MSE / 0.902 CI on KIBA and 0.207 MSE / 0.900 CI on Davis; the Zeng et al. (2024) review cited in the paper’s own introduction tabulates DGraphDTA at 0.126 MSE / 0.904 on KIBA and 0.202 MSE / 0.904 CI, HiSIF-DTA at 0.120 MSE / 0.904 CI on KIBA and 0.191 MSE / 0.906 CI on Davis, and GTAMP-DTA at 0.123 MSE / 0.917 MSE and 0.177 MSE / 0.923 CI for KIBA and Davis, respectively. PAFDTA’s KIBA MSE of 0.1500 is worse than all of these, and worse than GraphDTA’s reported MSE of 0.139. MGraphDTA and GraphDTA are discussed in Section 2 but excluded from Table 1 without explanation. The selection criterion must be stated clearly, or the table should be expanded.

  2. Every component has a direct predecessor: the GIN encoder from GraphDTA, multi-scale molecular modeling from MGraphDTA, the variational drug latent from TransVAE-DTA (a Table 1 baseline), frozen protein-LM features from LLMDTA, joint token streams from MolTrans and PerceiverCPI, the six meta-path schemas from Affinity2Vec, contrastive regularization from GraphCL-DTA and EMHGCL-DTA, and gated fusion plus mixture-of-experts from the general multimodal literature. The contribution therefore rests entirely on the assembly and on two mechanisms, CBCA and PAF. Table 3 shows that removing CBCA raises CI from 0.8980 to 0.9043, and removing the relational prior raises it to 0.9047, while Sum and Concat fusion both also exceed the full model’s CI.

  3. All baseline values are literature-reported across different splits and preprocessing steps.

  4. The cold-start claim is overstated, and the strongest comparators are absent. PAFDTA wins on all three metrics only in drug-cold. Under target-cold, LLMDTA has lower MSE (0.312 vs 0.319) and higher R² (0.485 vs 0.474); under double-cold, it leads by 0.061. MSE and 0.092 R². PAFDTA is ahead only on CI, by 0.004 and 0.006, from a single split with no variance reported. Methods that report cold-start results, such as ColdDTA, FusionDTA, NHGNN-DTA, and DMFF-DTA, are not compared.

  5. Davis and KIBA contain near-duplicate target sequences. The paper describes no sequence-identity or scaffold-based deduplication, so a held-out “unseen” target may have a near-identical training counterpart. The same applies to drug scaffolds in drug-cold.

  6. On Davis, removing the relational prior gives CI 0.9047 and removing CBCA gives CI 0.9043, both above the full model’s 0.8980; removing the joint pair representation also improves CI (0.9002). Of the eight rows in Table 3, the full model has the second-lowest CI; only removing the molecular branch is worse. The MSE spread across all eight rows is 0.0075, from a single seed. The full-model values also differ between Table 1 (0.2025 / 0.8994 / 0.7368) and Table 3 (0.2018 / 0.8980 / 0.7376) without explanation.

  7. The MLGA encoder is the first claimed contribution, yet discarding its output leaves performance essentially intact, suggesting the drug-identifying signal already lives in the pair branch and the meta-path prior.

  8. In Section 3.1.1, the local branch convolves over an atom ordering produced by SMILES parsing, which carries no chemical meaning; the GIN already captures genuine locality through bonds. No ablation isolates MSConv, the kernel sizes, the gate, or the variational projection.

  9. For the binding site analysis, "high-attention residues” is undefined; the model contains protein self-attention across six heads and multiple layers plus a masked attention-pooling operator, and the paper does not say from which or for which drug partner. Three to four residues are named per structure, with no enrichment statistic. The identities of 4Y85, 3GC8, 6EG9, and 8SSN, whether they are dataset targets, and how PDB numbering maps onto the truncated sequences are all unstated.

  10. For the docking study, no grid box center or size, exhaustiveness, run count, seeds, receptor preparation (chain, protonation, waters, cofactors), ligand preparation (tautomers, stereochemistry, conformers), or self-docking control on the native 4RZV ligand have been specified. “Midostaurin” and “M Midostaurin” occupy ranks 2 and 4 and appear closely related, so Figure 3(b) shows near-identical poses rather than independent evidence. No approved BRAF inhibitor appears in the top 10, which should be reported and discussed.

  11. The contrastive temperature tau is never given, nor is the number of GIN layers L, nor is it stated whether the variational molecular representation is sampled or set to μ at inference. Since ‘u’ depends only on the drug, while 'v' depends on the pair, the anchor comes from the drug alone, but the positive comes from the drug-target pair. A drug appearing in hundreds of pairs is therefore pulled toward hundreds of different positives at once, and any two pairs sharing that drug within a batch are treated as negatives of each other.

  12. Parameter counts, training time, memory, and inference cost are not reported for PAFDTA or the baselines.

Questions for authors
  1. What is the selection criterion for Table 1, and what are PAFDTA’s results against MGraphDTA, GraphDTA, DGraphDTA, HiSIF-DTA, and GTAMP-DTA under a common protocol?

  2. What is the performance of a meta-path-only model (six scores plus an MLP) under the same five-fold protocol on both datasets?

  3. Are the drug-drug and target-target relations derived from structure and sequence alone, or do they use affinity profiles? Are they recomputed per fold?

  4. Can all tables be reported with standard deviations over folds and seeds, with significance tests, and the cold-start experiments repeated over multiple entity-level splits?

  5. How do the authors reconcile the full model having the second-lowest CI of the eight rows in Table 3 with the claim that CBCA and the relational prior are contributions?

  6. Why do the Davis full-model values differ between Tables 1 and 3?

  7. Can MSConv, the kernel sizes, the gate, and the variational projection be ablated separately, and the ablations repeated on KIBA and in cold start?

  8. Was any sequence-identity or scaffold deduplication applied before the cold-start splits? If a held-out target shares a high degree of identity with a training target, how is this accounted for?

  9. What precisely is a “high-attention residue,” and can a quantitative pocket-enrichment measure (AUROC or precision@k against the 5 Å label, averaged over many targets, with a shuffled-attention null) be reported?

  10. Can the full docking protocol be provided, including a self-docking control for 4RZV and the Vina score distribution over the full 2,138-compound library?

  11. For each of the top 10 BRAF candidates, do the compound, the target, and the specific pair appear in training? Can the analysis be repeated with BRAF fully held out?

  12. What are tau, L, and the inference-time behavior of the variational layer? In Equation 7, "mean" is over the feature dimension, and can the learned distribution of alpha be reported for warm and cold settings?

4 comments
AuthorsOpenReview ↗

Response to Reviewer KfSB — Part 1 (Q1–Q3)

We appreciate the reviewer’s detailed assessment and the recognition of CBCA’s distinction from conventional molecular-view contrastive learning. We provide the following clarifications and additional experimental evidence.

Q1. What is the selection criterion for Table 1, and what are the results against the listed stronger baselines under a common protocol?

Table 1 selects representative DTA methods from different methodological families, including adversarial learning (GANsDTA), Transformer-based modeling (TF-DTA), graph-based modeling (GDilatedDTA), variational representation learning (TransVAE-DTA), attention-based interaction modeling (DTIAM), and pretrained protein-language-model approaches (LLMDTA).

Our PAFDTA results are obtained using five-fold pair-wise cross-validation, while baseline values are taken from their respective publications, following the reporting convention of previous DTA studies. We clarify that Table 1 is a literature-based comparison rather than a common-protocol reimplementation.

Among the listed methods, PAFDTA achieves the lowest MSE on both datasets:

KIBA: MSE = 0.1500, CI = 0.8868, R² = 0.7858.

Davis: MSE = 0.2025, CI = 0.8994, R² = 0.7368.

It also achieves the highest reported R² among the listed methods on both datasets. These results support the effectiveness of coordinating complementary molecular, protein, pairwise, and relational information.

Q2. What is the performance of a meta-path-only model using six scores and an MLP?

We additionally implemented a two-layer MLP using the six meta-path scores as its only input and evaluated it using five-fold experiments.

The results demonstrate that relational topology contains independently predictive information:

Davis: R² = 0.342 ± 0.034, CI = 0.797, AUPR = 0.445.

KIBA: R² = 0.018 ± 0.006, CI = 0.547.

This experiment isolates the predictive information contained in relational features. Importantly, PAFDTA does not simply use the meta-path scores as its prediction output; it coordinates relational information with molecular and protein representations, while retaining a direct drug-target pair pathway. The prior-only experiment therefore provides a complementary reference for understanding the role of relational information within the complete architecture.

Q3. How are DD, TT, and DT relations constructed? Are they recomputed for each fold?

Drug-drug and target-target relations are constructed from their respective similarity information, whereas drug-target relations are derived from observed affinity interactions.

For each experimental fold, the drug-target transition matrix is constructed exclusively from the training interactions. The six meta-path matrices are then recalculated using the corresponding fold-specific relations.

This procedure prevents held-out affinity labels from being used to construct the drug-target relational prior and ensures that the relational representation follows the training graph of each fold.

The six path schemas follow the established Affinity2Vec formulation. Our methodological contribution is their integration into CBCA and progressive multimodal fusion, rather than claiming the path schemas themselves as newly invented.

We hope these clarifications better establish the experimental scope and the role of relational information in PAFDTA.

AuthorsOpenReview ↗

Response to Reviewer KfSB — Part 2 (Q4–Q6)

Q4. Can all tables report standard deviations over folds and seeds, significance tests, and multiple cold-start splits?

Our primary warm-start evaluation uses five-fold pair-wise cross-validation with a fixed seed of 42 and a 200-epoch training schedule. The five held-out folds are mutually exclusive, cover all observed drug-target pairs, and each pair appears in exactly one test fold. Table 1 reports the mean across these five folds.

The different tables have distinct evaluation purposes:

Table 1: five-fold warm-start benchmark results.

Table 2: KIBA entity-level cold-start results using the same fixed partition for all compared methods.

Table 3: controlled component ablations on Davis fold 0.

We keep these evaluation protocols separate rather than combining single-fold, multi-fold, and entity-level results into an artificial common variance estimate. In particular, the five-fold benchmark evaluates pair-wise generalization, whereas the cold-start experiments evaluate transfer to held-out drug and/or target identities.

We also performed additional five-fold evaluation for the meta-path-only MLP and report its fold-level variability. These results provide an additional assessment of the relational representation.

Q5. Why does the Full model have the second-lowest CI in Table 3?

The key point is that the contribution of PAFDTA should not be evaluated using CI alone.

The complete Table 3 results show that Full achieves the best numerical regression performance. Removing CBCA or the relational prior slightly increases CI, but both variants have higher MSE and lower R² than Full. Furthermore, both fixed fusion alternatives produce higher MSE and lower R².

Thus, Full provides the strongest numerical affinity estimation among these variants, while some simplified configurations produce slightly different pairwise rankings. We believe this distinction is important when evaluating a regression framework.

Q6. Why are the Full values different between Table 1 and Table 3?

The two tables report results from different evaluation scopes.

Table 1 reports the five-fold warm-start mean, whereas Table 3 reports the fold-0 result used for the component ablation study.

Accordingly, the Full results are:

Table 1: 0.2025 MSE / 0.8994 CI / 0.7368 R².

Table 3: 0.2018 MSE / 0.8980 CI / 0.7376 R².

The difference reflects five-fold aggregation versus a single-fold ablation evaluation, rather than two metrics from the same aggregation protocol.

We hope this clarifies the evaluation design and the interpretation of the ablation results.

AuthorsOpenReview ↗

Response to Reviewer KfSB — Part 3 (Q7–Q9)

Q7. Can MSConv, individual kernel sizes, the gate, and the variational projection be evaluated separately?

The molecular representation is designed to combine complementary information from graph connectivity and local-global aggregation. GIN propagates information through molecular graph connectivity, while MSConv extracts local patterns at multiple receptive-field sizes (1, 3, and 5). A dimension-wise gate then combines the global and local representations, followed by variational projection.

We additionally conducted a connectivity-preserving node-reindexing diagnostic to examine the sensitivity of the trained model to atom ordering. For 300 samples per setting, the prediction correlations after random node permutations were:

KIBA drug-cold: 0.999998; KIBA double-cold: 0.999985; Davis drug-cold: greater than 0.999999.

The 95th-percentile absolute prediction changes were below 7.4 × 10⁻³ pKd, and the absolute CI changes were below 5 × 10⁻⁴.

This provides empirical evidence that the trained predictor remains stable under the tested connectivity-preserving atom reindexing operations.

Q8. Was sequence-identity or scaffold deduplication applied before cold-start evaluation?

Our KIBA cold-start evaluation is defined at the entity-identity level:

Drug-cold: test drug identities are held out from training.

Target-cold: test target identities are held out from training.

Double-cold: both drug and target identities are held out.

All compared methods use the same entity-level partition within each setting, and the relational prior is reconstructed from the corresponding training interactions.

We distinguish entity-level cold-start evaluation from scaffold-disjoint or protein-family-disjoint evaluation. Our reported results are intended to characterize transfer to unseen entity identities under the specified splits.

Q9. What precisely is a high-attention residue, and what is its relationship to the binding pocket?

In our model, protein residue representations are contextualized by Transformer encoders and summarized using masked attention pooling. A high-attention residue refers operationally to a residue receiving a high learned pooling weight.

For structural comparison, pocket residues are defined using a 5 Å heavy-atom distance threshold from the bound ligand in the corresponding co-crystal structure.

Figure 3(a) is intended as a qualitative structural interpretation of learned protein representations. It complements the affinity prediction and uncertainty analyses rather than defining PAFDTA as a dedicated binding-site prediction model.

We appreciate the suggestion to use quantitative pocket-enrichment metrics and shuffled-attention controls to further investigate this aspect of model interpretability.

AuthorsOpenReview ↗

Response to Reviewer KfSB — Part 4 (Q10–Q12)

Q10. Can the full docking protocol and a self-docking control be provided?

Our BRAF case study follows a two-stage workflow. First, PAFDTA ranks a library of 2,138 FDA-approved small molecules against BRAF. Subsequently, selected high-ranking compounds are examined using AutoDock Vina.

The model performs the library-wide affinity ranking, whereas docking serves as a structural follow-up for selected candidates. Thus, the two procedures have distinct roles: the model prioritizes candidates at library scale, and docking provides structural hypotheses for representative compounds.

The top-10 ranking and corresponding reported Vina scores are summarized in Appendix H (Table 6). We emphasize that docking is used as complementary structural evidence, not as experimental confirmation of binding affinity.

Q11. Do the top-10 compounds, BRAF, and the corresponding pairs appear in training?

The screening model is trained using KIBA and then applied to the FDA-approved candidate library. The exact candidate–BRAF interaction pairs are not part of the observed KIBA training interactions.

We distinguish this pair-level evaluation from target-held-out generalization: BRAF is represented in the KIBA benchmark, so this case study is intended to examine candidate prioritization for BRAF rather than claim a completely unseen-target evaluation.

We also performed an external evaluation using 21 measured BRAF drug–target pairs. The results show:

Spearman correlation: ρ = 0.5422; Permutation-test p-value: approximately 0.0112; AUROC: 0.944; Average Precision: 0.639.

This provides additional evidence that the model ranking contains useful information about measured BRAF affinity. Unmeasured compounds are not treated as experimentally confirmed negatives.

Q12. What are τ, L, the inference behavior of the variational layer, and the interpretation of the mean in Equation 7?

The contrastive temperature is τ = 0.07, and the GIN implementation uses L = 4 message-passing layers.

For uncertainty estimation, we perform T = 20 stochastic forward passes, with dropout active and the variational drug encoder sampled. The predictive mean is calculated across these passes, and their sample standard deviation is used as the uncertainty scale.

In Equation 7, mean is calculated along the feature dimension of each representation branch. The resulting three scalar summaries are passed through the modality-weighting network and normalized using softmax to obtain α.

For an audited Davis warm fold-0 gate summary, the mean modality weights were approximately:

Molecular: 0.235; Protein: 0.334; Relational: 0.431.

These values show that the learned fusion retains contributions from all three representation sources.

Regarding computational cost, the model has approximately 3.849 million trainable parameters. On our reference hardware, training takes approximately 112 seconds per epoch, and evaluation takes approximately 23 seconds. These measurements characterize the PAFDTA implementation under the recorded experimental configuration.

We hope these additional implementation details clarify the model's computational characteristics and make its design easier to reproduce.

Reviewer_C2TAOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PAFDTA for drug-target affinity prediction. The model combines molecular graph features, pretrained protein representations, drug-target pair context, and relational topology information. Its main components are Cross-Branch Contrastive Alignment and Progressive Adaptive Fusion. The authors evaluate the method on Davis and KIBA, including warm-start and cold-start settings, ablation studies, and downstream screening analyses.

Critical strengths
  1. The paper considers several complementary sources of information for drug-target affinity prediction and integrates them into a unified framework.
  2. The evaluation covers both standard benchmarks and more challenging cold-start settings. The uncertainty and structural analyses are also useful for understanding the model's potential applications.
Critical weaknesses
  1. Section 3.2 aligns drug-level molecular representations with pair-specific relational representations, but the same drug can interact with multiple targets, the current contrastive objective may incorrectly treat related pairs as negatives. The authors should better justify the alignment strategy and its positive and negative sample construction.
  2. The two-stage fusion mechanism in Section 3.3 appears somewhat heuristic. It is unclear why both stages are necessary or why the three specific interaction experts are chosen. The comparison with only ‘sum’ and ‘concat’ is insufficient, stronger fusion baselines or ablations of individual stages would help justify the proposed method.
  3. The experimental results do not consistently support the proposed components. The warm-start baselines are taken from previous studies rather than evaluated under a unified protocol. In target-cold and double-cold settings, improvements in ranking performance are marginal, while regression performance is worse than baseline. Moreover, removing CBCA or the relational prior improves the concordance index in ablation study, raising questions about their effectiveness.
Questions for authors
  1. How does CBCA handle cases where the same drug appears with multiple targets? Could this introduce false negatives during contrastive learning?
  2. Could the authors provide further justification or ablation results for the two-stage fusion design?
  3. Could the authors clarify why removing CBCA or the relational prior improves ranking performance in the ablation study?
1 comment
AuthorsOpenReview ↗

Response to Reviewer C2TA

We sincerely thank the reviewer for the constructive feedback and for recognizing the value of integrating complementary representations and conducting uncertainty-aware evaluation. We would like to further clarify the motivation and empirical contributions of CBCA and PAF.

Q1. How does CBCA handle cases where the same drug appears with multiple targets? Could this introduce false negatives?

CBCA is designed to align drug-centric molecular representations with pair-specific relational contexts. Specifically, the matched molecular and topology representations form positive pairs, while other samples within the mini-batch provide contrastive negatives. We agree that different drug–target pairs sharing the same drug may create potential negative collisions.

To quantify this issue, we conducted a batch-level collision audit and an additional same-drug negative-masking experiment. Across the Davis fold-0 training batches, 1,058,676 out of 72,127,200 in-batch negative entries (1.47%) corresponded to potential same-drug collisions. The masking implementation was verified in both contrastive directions, with no degenerate contrastive rows. The resulting CI remained essentially unchanged (0.9012 for the original objective and 0.9009 for the masked variant). This provides a direct assessment of the sensitivity of CBCA to this potential issue.

Importantly, CBCA does not aim to make molecular and relational representations identical. Instead, it uses the relational context as an auxiliary structural signal to regularize the molecular representation before fusion.

Q2. Why are two fusion stages necessary, and why are these three experts selected?

The two stages address two different levels of representation integration.

Stage I performs modality-wise aggregation, learning the relative contributions of molecular, protein, and relational representations. This is important because these sources contain complementary information and may have different relevance to affinity prediction.

Stage II performs relation-wise interaction modeling. We use three experts with progressively different interaction capabilities: a linear expert for relatively simple combinations, a deep nonlinear expert for complex feature dependencies, and a multiplicative expert for explicit cross-representation interactions.

The motivation is therefore not simply to increase model complexity, but to separate source-level aggregation from interaction-level modeling.

Importantly, the controlled fusion comparisons in Table 3 retain the same representation branches while replacing PAF with fixed Sum and Concat operations. The Full model achieves MSE/R² of 0.2018/0.7376, compared with 0.2084/0.7313 for Sum and 0.2064/0.7339 for Concat. This supports the benefit of learned progressive fusion for numerical affinity prediction.

Q3. Why can removing CBCA or the relational prior increase CI?

MSE, R², and CI measure different aspects of prediction. CI evaluates relative ranking, whereas MSE and R² evaluate numerical affinity estimation.

In Table 3, the Full model achieves the lowest MSE (0.2018) and highest R² (0.7376). Removing CBCA increases MSE to 0.2045 and reduces R² to 0.7363. Removing the relational prior increases MSE to 0.2051 and reduces R² to 0.7355.

Although these two variants obtain slightly higher CI, their numerical regression performance is weaker. Similarly, replacing PAF with Sum or Concat leads to higher MSE and lower R².

Therefore, the contributions of CBCA and relational information should be evaluated through their complementary effects on numerical affinity estimation and representation coordination, rather than CI alone.

Beyond point prediction, our uncertainty-aware analysis achieves 89.5% empirical coverage at a nominal 90% level. The conformal screening stage further identifies 154 active pairs among 166 selected candidates, reaching 92.8% precision. These results demonstrate the practical value of combining affinity prediction with calibrated candidate prioritization.

We hope these clarifications and additional analyses better demonstrate the motivation and contributions of PAFDTA.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.