acceptodds
Under review as a conference paper at ICLR 2027

ECMLF: An Explainable Cross-Modal Learning Framework for Cold-Start Protein-Protein Interaction Prediction

Abstract

Cold-start protein-protein interaction (PPI) prediction aims to identify potential interactions involving newly discovered proteins that lack sufficient experimentally observed interaction records, thereby facilitating the functional characterization of these proteins. Its central challenge is to capture the similarity between new and known proteins using information from diverse modalities. However, existing methods have three limitations in capturing such similarity: (1) only partial modalities are considered, resulting in incomplete or biased similarity modeling; (2) the similarity between new and known proteins is computed offline and independently across modalities, causing inconsistency between similarity modeling and cold-start PPI prediction; (3) current explanations are largely single-modal and feature-oriented, overlooking biologically grounded evidence for interaction mechanisms. To address these issues, we propose an explainable cross-modal learning framework (ECMLF) with three components. First, comprehensive modalities of intrinsic amino acid sequences and four external biomedical knowledge graphs are leveraged to capture similarity between new and known proteins. Second, an adaptive gated fusion module is designed to combine similarity modeling and cold-start PPI prediction within an end-to-end learning framework. Finally, a dual-pathway explanation mechanism is proposed to provide biologically grounded explanations of interaction mechanisms. Extensive experiments on two real datasets show that ECMLF achieves superior prediction accuracy and explanation quality compared with existing methods for cold-start PPI prediction. The source code and datasets are submitted as Supplementary Materials for Reproducibility.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Cold-start protein–protein interaction prediction, cross-modal learning, biomedical knowledge graphs, graph neural networks, explainable artificial intelligence

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 4 reviews, above 85% of the others.
Reviewer_cMSMOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This study proposes an Explainable Cross-Modal Learning Framework (ECMLF) for cold-start protein–protein interaction (PPI) prediction. ECMLF integrates amino acid sequence information with four external biomedical knowledge graphs to model similarities between novel and known proteins. An adaptive gated fusion module jointly learns protein similarity and PPI prediction, while a dual-pathway explanation mechanism provides biologically grounded insights into predicted interactions. Experiments on two real-world datasets demonstrate improved predictive performance and explanation quality over existing methods.

Critical strengths
  1. This study proposes ECMLF, which enriches the contextual information available for protein–protein interaction (PPI) prediction by incorporating four biologically meaningful graph attention networks alongside protein sequence representations. This multimodal design constitutes a valuable methodological contribution.

  2. Beyond its technical contributions, this research work examines the model’s attention scores for protein sequence and graph representations. This interpretability analysis offers potential insight into the mechanisms underlying its PPI predictions and strengthens the presentation of the work.

  3. ECMLF demonstrates strong performance across extensive experimental evaluations, as measured by accuracy (ACC), recall (REC), precision (PRE), F1 score, area under the ROC curve (AUC), area under the precision–recall curve (AUPR), and Matthews correlation coefficient (MCC). Together, these metrics provide broad evidence of the constructed model’s predictive advantages.

Critical weaknesses
  1. ECMLF’s performance gains rely substantially on the availability and completeness of four types of external biomedical associations. However, the paper does not systematically evaluate the model under missing modalities, sparse annotations, or distribution shifts. Consequently, its robustness to poorly annotated proteins and incomplete knowledge graphs in real-world scenarios remains unclear.

  2. Considering the paper as a whole, ECMLF introduces additional multimodal information from knowledge graphs to help the model better capture the biological mechanisms of proteins. Subsequently, the subsequent gated fusion mechanism combines representations from four knowledge graphs and a protein sequence encoder before passing them to the PPI prediction head. The pipeline in stage II does not appear to differ substantially from conventional approaches or offer a clear algorithmic innovation. Could the performance gains over the baselines models therefore be attributed primarily to the additional multimodal information?

  3. The included interpretability analyses of individual examples can shed light on plausible mechanisms underlying the model’s correct predictions and thus complement the main findings. However, they do not constitute a large-scale statistical analysis or provide robust evidence for those mechanisms. (minor)

Questions for authors

I recommend that the authors further clarify the core algorithmic innovations underlying the proposed method’s performance gains and provide additional evidence of the generalizability of these innovations in real-world application scenarios.

4 comments

W1: We acknowledge that ECMLF’s performance depends on the availability of the four types of external biomedical associations. To further assess their impact, we conducted two additional experiments. (1) To assess how the number of available graph modalities affects prediction performance for new proteins, we conducted a modality coverage experiment. Specifically, unlike the modality ablation experiment in Table 3 of the original manuscript, which trains a separate model for each modality combination, this experiment uses the full model trained with all four graph modalities. During the testing phase, we selected new proteins covered by all four graph modalities as the test set and masked their graph modality information to simulate new proteins with modality coverage of nn (n=4,3,2,1n = 4, 3, 2, 1). The results (see revised manuscript, Appendix A.6 "Modality Coverage Ablation") indicate that broader graph modality coverage generally improves prediction performance for new proteins. However, even when only one graph modality is available, the model retains a degree of predictive ability for new proteins. (2) To assess whether the model’s advantage persists with few biomedical graph associations, we conducted a graph association ablation experiment comparing the full model with its sequence-only variant. Specifically, we progressively removed 25%, 50%, 75%, and 100% of each new protein’s biomedical graph associations and reported performance stratified by the amount of graph information remaining. Based on the experimental results, as graph associations decrease, the model’s performance gradually declines. But it generally outperforms the sequence-only variant, even for new proteins with few or no remaining graph associations (see revised manuscript, Appendix A.6 "Graph Association Ablation").

W2: (1) We acknowledge that Stage II does not appear to differ substantially from conventional approaches. However, in Stage I, we introduce bridge reconstruction as an auxiliary supervision task to learn latent similarities between known and new proteins, enabling interaction patterns learned from known proteins to transfer to new proteins. This specific bridge reconstruction objective is the core design of our model and is absent from existing conventional PPI methods like KGF-GNN and HEENN. Specifically, bridge reconstruction is designed to address the absence of PPI records for new proteins by learning similarities through biomedical associations. (2) The reconstruction ablation experiment has already demonstrated the importance of this component for cold-start PPI prediction (see revised manuscript, Appendix A.6 "Reconstruction Ablation"). To exclude the influence of sequence information and further assess the contribution of bridge reconstruction, we conducted an additional reconstruction ablation experiment without the sequence branch. The results further support the contribution of bridge reconstruction to cold-start prediction (see revised manuscript, Appendix A.6 "Further Reconstruction Ablation"). (3) Beyond the performance changes shown by the reconstruction ablation above, we have added another experiment to further examine how bridge reconstruction mechanistically addresses the absence of PPI records for new proteins. In the Graph-Pathway Explainability section of the original manuscript, we evaluated the model’s ability to capture key bridge entities that connect new proteins to known proteins sharing similar interaction patterns, thereby enabling the model to transfer interaction patterns learned from known proteins to new proteins. For new proteins lacking PPI records, our bridge reconstruction task provides auxiliary supervision, guiding the model to capture these key bridge entities to uncover latent associations between new and known proteins. Therefore, we compared the full model with its variant without bridge reconstruction in their ability to capture these key entities. The results also show that removing bridge reconstruction reduces the model’s ability to identify key entities connecting known and new proteins (see revised manuscript, Appendix A.7 "Bridge Entity Identification Ablation").

W3: In Section 5.4, both sequence-pathway and graph-pathway explanations are evaluated for every protein pair in the full test set, providing large-scale quantitative evidence of their biological relevance. The case study in Section 5.4 is presented to specifically illustrate sequence-pathway explainability for a selected protein pair, offering biological evidence for a plausible mechanism underlying the correct prediction. However, due to page limits, we did not conduct case study for all protein pairs and the existing case study already supports the sequence-pathway explanation analysis. The case study analyses could be extended to all protein pairs, but this would require substantial time and computational resources.

The following summarizes our additional experiments. Numerical results available in the manuscript tables are reproduced below; for other experiments, we refer to the corresponding sections of the revised manuscript.

W1(1):Modality Coverage Ablation. (see revised manuscript, Appendix A.6 "Modality Coverage Ablation")

SCS

CoverageModalityACCRECPREF1AUCAUPRMCCF-rank
1M160.9187.6558.7764.9874.8866.6037.8611.14
1M261.7587.2858.8465.2473.4865.0639.028.43
1M360.9987.5758.6764.9975.9268.1937.5910.29
1M461.0387.2458.8764.9371.9463.6137.5812.14
2M1+M261.4688.4458.6065.4575.1666.8039.567.00
2M1+M360.9288.4558.4965.2477.4169.7338.168.14
2M1+M460.9588.1258.7265.1874.3567.1538.089.86
2M2+M361.5988.0858.4765.4076.4968.6238.937.00
2M2+M461.7787.8058.7865.4672.4763.9939.187.57
2M3+M460.9888.0158.5165.1376.0368.5737.639.71
3M1+M2+M361.3289.1958.2265.5877.6869.9339.614.86
3M1+M2+M461.4588.8958.5165.6174.6067.3439.695.86
3M1+M3+M460.9188.8458.3465.3577.5370.1238.227.57
3M2+M3+M461.5688.4958.3165.5176.5168.8238.926.29
4M1+M2+M3+M461.3089.6358.0165.6977.7570.2339.634.14

FCS

CoverageModalityACCRECPREF1AUCAUPRMCCF-rank
1M172.9977.7462.8766.2685.7976.4149.5211.57
1M278.4972.4168.2168.9886.0276.3253.867.86
1M373.8578.0263.1067.0385.6075.6950.3511.00
1M474.1677.5663.8067.0885.8076.0250.7910.29
2M1+M277.8575.6266.7869.3086.4877.3854.215.71
2M1+M372.1479.2661.4466.2086.0876.7448.8311.43
2M1+M472.3778.8062.0866.2186.3177.1549.1810.29
2M2+M378.1375.4766.8469.5986.2476.6254.316.00
2M2+M478.2474.5267.3869.4286.3476.8754.225.86
2M3+M473.1679.2362.2566.9685.9876.2749.9710.57
3M1+M2+M377.1878.3065.3069.6186.7177.6854.194.57
3M1+M2+M477.3877.2965.8569.4786.8578.0054.145.14
3M1+M3+M471.7280.6460.8566.4486.4977.4048.899.29
3M2+M3+M477.6976.8965.9169.6886.4677.0654.145.86
4M1+M2+M3+M476.5479.4564.2369.4986.9778.1953.654.57

W1(2):Graph Association Ablation. (see revised manuscript, Appendix A.6 "Graph Association Ablation")

W2(2):Further Reconstruction Ablation. (see revised manuscript, Appendix A.6 "Further Reconstruction Ablation")

W2(3):Bridge Entity Identification Ablation. (see revised manuscript, Appendix A.7 "Bridge Entity Identification Ablation")

I thank the authors for their response. After the discussion, I understand that the incorporated bridge reconstruction mechanism constitutes the paper’s primary methodological innovation. After holistic consideration, I decide to maintain my original rating score.

Thank you for your understanding. We appreciate your time and consideration.

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

The paper studies protein–protein interaction (PPI) prediction when one or both proteins have no PPI labels available during training. ECMLF combines ESM-2 embeddings of annotation-defined sequence regions with representations learned from four protein–entity graphs covering drugs, diseases, miRNAs, and lncRNAs. Graph encoders are trained with an auxiliary four-class bridge-reconstruction objective. Symmetric pair features are combined through a pair-dependent modality gate and an interaction classifier. The paper also presents two explanation pathways: attention-based sequence-region localization and retrieval of shared biomedical entities connecting held-out proteins to known proteins. Experiments on two STRING-derived datasets report improvements over seven baselines in semi-cold-start and full-cold-start settings.

Critical strengths
  1. A coherent use of complementary information for label-cold proteins. Sequence information remains useful when PPI neighborhoods are unavailable, while external protein–entity associations can supply additional context. The symmetric absolute-difference and Hadamard-product features are appropriate for undirected interaction prediction, and joint optimization of prediction and graph reconstruction is a reasonable design. This combination addresses a meaningful practical setting rather than relying exclusively on observed PPI topology.
  2. Promising improvements supported by internal ablations. The reported gains are substantial, particularly in full-cold-start prediction. Importantly, the internal sequence-only comparison also favors multimodal integration: full-cold-start AUC increases from 75.40 to 82.15 on Dataset1 and from 82.90 to 88.99 on Dataset2. The reconstruction, segmentation, encoder, and fusion ablations provide additional evidence that the design choices are useful within the authors' experimental setup. These results make the approach worth further investigation, even though the broader comparative claims need stronger controls.
Critical weaknesses
  1. The evaluation does not yet establish the scope or source of the predictive advantage.

    • Appendix A.5 describes protein-held-out splits, which is preferable to random pair splitting. Nevertheless, essential dataset details are missing: the organism and protein-selection criteria, STRING confidence threshold and evidence channels, numbers of positive PPI pairs, negative-pair construction, class ratios, and overlap between the two STRING-derived datasets. Table 5 reports biomedical association counts rather than the PPI examples being classified. In particular, STRING functional associations should not automatically be interpreted as direct physical binding; the target needs to be specified before the biological claims can be assessed.

    • The random protein split does not control sequence homology, and the manuscript does not clearly specify whether held-out proteins and their biomedical edges participate in graph reconstruction during training.

    • Key ablation results are also missing. For example, on SCS, MsM_s alone already achieves an AUC of 87.80, while the full model reaches only 88.79. This indicates that most of the model's performance comes from sequence information, and it remains unclear which of M1M_1 to M4M_4 contributes the remaining gain of about 1 point. It also suggests that more modalities are not necessarily better.

  2. The explanation experiments support biological plausibility, but not faithful explanations of specific interaction mechanisms.

    • The sequence attention weights in Eqs. (1)–(3) depend on a protein alone, not on its interaction partner. Consequently, a protein receives the same ranked regions for different candidate partners. Although the final modality gate is pair-dependent, the reported sequence explanations do not account for that gate or demonstrate which regions actually drive a particular prediction.

    • The annotation-overlap metric in Eq. (15) measures coverage of an annotated region, not localization precision or recovery of a partner-specific binding interface. A sufficiently long selected segment can therefore score well.

    • Section 4.3 explicitly uses the annotated PPI network to define associated known proteins after prediction, it means the evaluation is an oracle-assisted retrospective analysis, not a demonstrated explanation procedure for proteins whose interactions are unknown. One HDOCK-generated CDK4–PCNA complex provides a plausible structural hypothesis, not independent validation of the proposed interface or mechanism. Together, these issues leave the paper's central "explainable" contribution substantially less supported than its predictive contribution.

Questions for authors
  1. Is the pre-computation of protein similarity actually a problem? Why can the Adaptive Gated Cross-Modal Fusion module resolve this pre-computation issue?
  2. Please specify the STRING release and confidence threshold, protein-selection criteria, positive and negative PPI counts, negative-sampling procedure, and class ratios for each split. Are the labels physical interactions or broader functional associations? How are the knowledge-graph edges of test proteins handled? What information can PnewP_{new} obtain from the knowledge graph?
  3. Can the authors provide ablations of Ms+XM_s + X, where XX is one of M1M_1–M4M_4?
  4. Can the explanation results be shown to reflect specific predictions rather than annotation coverage or graph connectivity? Which annotations independently evaluate the explanations? For the graph pathway, can explanations be constructed without held-out PPI labels, and does removing selected regions or entities affect the relevant prediction more than matched controls?
6 comments

Q1: (1)Independently pre-computed protein similarities cannot be guided or optimized by PPI supervision. (2)This issue is addressed by end-to-end joint optimization(Eq.14), rather than cross-modal fusion itself. Specifically, bridge reconstruction and PPI prediction jointly optimize shared protein representations, allowing the latent similarities to be optimized under PPI supervision. (3)To further evaluate the benefit of joint optimization, we compare the full model with a variant that replaces joint optimization with pre-computed similarity (see revised manuscript, Appendix A.6 "Joint Optimization Ablation").

Q2: (1)We used the publicly available datasets released with MTV-PPI and HEENN as Dataset1 and Dataset2, respectively. The table below summarizes dataset details checked with the original publications and STRING database.

DatasetSTRING ReleaseConfidence ThresholdProteins (Dataset-Specific)Positive(Dataset-Specific)Negative
Dataset1v11.0Combined score from the experimental and database evidence channels ≥0.41648(387)19237(15152)19237
Dataset2v12.0Combined score across all evidence channels >0.94283(3022)52034(43864)52034

All proteins belong to Homo sapiens, and positive labels were inherited from the upstream source datasets. Dataset1 retained 1648 proteins from MTV-PPI after excluding one without an available sequence(this protein had no recorded positive label in the dataset), and Dataset2 retained all 4283 proteins from HEENN. The negative samples were randomly drawn from protein pairs absent from the positive set at a 1:1 ratio. To reduce the risk of including latent true interactions, we screened these candidates against experimentally supported PPI records in the IntAct database. Pairs supported by these records were excluded and replaced through resampling. The ratio(+:-) is approximately 1:1 in the training, validation, and test sets of each fold. (2)After reviewing the original publications and checking with the STRING database, we found that many positive labels were in STRING physical interaction networks(Dataset1:51.31%; Dataset2:71.36%), and the rest encompass broader functional associations. Both physical interactions and broader functional associations have already been uniformly encoded as positive labels in these public datasets. Furthermore, in our explainability analysis, we describe these interfaces as potential or possible sites of binding, presenting them as hypotheses rather than definitive claims of physical binding. (3) During training, with PPI labels involving all new proteins kept hidden, knowledge-graph edges are retained for all proteins, providing auxiliary supervision for new proteins without PPI records through bridge reconstruction to learn similarities between known proteins and new proteins. Specifically, PnewP_{\mathrm{new}} obtains information about its biomedical associations, as well as relational evidence of similarity to known proteins through shared biomedical entities.

Q3: We extended the modality ablation study to evaluate all configurations of the form Ms+∑i∈mMiM_s + \sum_{i \in m} M_i, where m⊆{1,2,3,4}m \subseteq \{1,2,3,4\} (see revised manuscript, Section 5.3 "Ablation Study"). The results show that combining graph modalities generally improves predictive performance, with larger gains under FCS. Through bridge reconstruction over biomedical knowledge graphs, the model learns latent similarities between known and new proteins connected by shared entities, facilitating the transfer of interaction patterns learned from known proteins to new proteins.

Q4: (1)To assess whether the sequence explanations reflect specific predictions, we used SHAP to quantify the contribution of every sequence region to the predicted interaction for each test pair (see revised manuscript, Appendix A.8 "SHAP Analysis of Sequence Region Contributions"). Based on the results, the top two regions in each protein collectively accounted for 51.66% of the total attribution, supporting their relevance to the model’s predictions. (2)The annotations we use to evaluate the explanations are protein domains, motifs, regions, repeats, and compositionally biased regions. (3)Held-out PPI labels are used only after prediction as reference evidence for validation. The graph-pathway analysis aims to explain how the model uses learned latent similarities to transfer interaction patterns from known proteins and correctly predict PPIs involving new proteins. For unknown PPIs, wet-lab experiments are needed to verify the predictions and assess the explanations. And most existing models also rely on known PPI information to validate their explanations. Furthermore, we conducted experiments that remove key sequence regions and graph entities to assess their effects on the predictions (see revised manuscript, Appendix A.6 "Key Region and Entity Ablation").

W1: (1)As detailed in our response to Q2, we clarified the organism and protein-selection criteria (proteins were taken directly from the upstream public datasets), STRING confidence threshold and evidence channels, numbers of positive PPI pairs, negative-pair construction, class ratios, and overlap between the two STRING-derived datasets. (2)To control sequence homology, we constructed five-fold splits based on sequence similarity. Specifically, MMseqs2 was used to compare protein sequences against one another and group proteins with sequence identity ≥40% and alignment coverage ≥80% for both sequences into the same group, thereby limiting sequence similarity between proteins in different groups. These groups were assigned intact to five folds with similar protein counts, with proteins in one fold treated as new proteins and those in the other four as known proteins in each round. Then, we compared our model with the baselines on these homology-controlled splits (see revised manuscript, Appendix A.9 "Performance under Homology-Controlled Splits"). The experimental results show that our model still achieved the best predictive performance among the compared methods. Moreover, as clarified in Q2, held-out proteins and their biomedical association edges participate in bridge reconstruction during training, while their PPI labels are withheld. (3)As detailed in our response to Q3, we extended the modality ablation study to assess the additional contributions of different graph modalities beyond sequence information. Although MsM_s alone performs strongly under SCS, the full model outperforms better under more challenging FCS setting. Through bridge reconstruction, biomedical associations provide auxiliary supervision for learning latent similarities between known and new proteins, supporting the transfer of learned interaction patterns to new protein pairs. Therefore, more modalities are not necessarily better, as their benefit depends on whether associated entities provide informative connections between known and new proteins.

W2: (1) We added an experiment comparing the original model with a variant using protein-pair-dependent region attention. As shown in the table below, the two models achieve similar predictive performance.

TaskModelACCRECPREF1AUCAUPRMCC
SCSOriginal81.3178.4183.6180.8988.7988.3862.83
SCSPair-dependent attention81.4179.6183.4981.5088.4088.2462.91
FCSOriginal74.4068.4777.4372.3482.1582.0048.85
FCSPair-dependent attention74.8663.5276.3969.3682.9580.4748.95

Protein interactions can be mediated by specific domains or motifs, and some binding regions are reused across different partners. Thus, even though a protein’s region ranking remains unchanged across partners, its top-ranked regions may remain relevant to multiple interactions. Furthermore, we conducted a comparative SHAP analysis of the original model and the variant with protein-pair-dependent region attention to assess how their top-kk regions contribute to PPI predictions (see revised manuscript, Appendix A.8 "SHAP Analysis of Sequence Region Contributions"). The results show that the top two regions in each protein accounted for similar proportions of total absolute SHAP attribution: 51.66%(original model) and 52.06%(protein-pair-dependent variant).

(2) Eq.(15) contains a typographical error: the denominator should be the selected region’s length, not the annotated region’s length. We have corrected it in the revised manuscript:

Hit⁡(sij)=I[max⁡Aik∈Ai∣sij∩Aik∣∣sij∣≥τ].\operatorname{Hit}(s_{ij})=\mathbb{I}\left[\max_{A_{ik}\in\mathcal{A}_i}\frac{|s_{ij}\cap A_{ik}|}{|s_{ij}|}\geq\tau\right].

Furthermore, although our explanation cannot precisely pinpoint binding interfaces, it can highlight broader regions that may contain potential binding sites.

(3) As detailed in our response to Q4, our explanation analysis uses known PPI records and functional annotations as reference evidence to examine how the model identifies regions potentially involved in binding and captures latent similarities through key entities bridging known and new proteins. Similarly, most existing methods rely on known PPI information or functional annotations to validate their explanations. Validating previously unknown PPIs and confirming their binding interfaces require wet-lab experiments. But given our current experimental constraints, we used HDOCK docking as a supplementary assessment of structural plausibility.

The following summarizes our additional experiments. Numerical results available in the manuscript tables are reproduced below; for other experiments, we refer to the corresponding sections of the revised manuscript.

Q1:Joint Optimization Ablation. (see revised manuscript, Appendix A.6 "Joint Optimization Ablation")

Q3/W1(3):Modality Ablation. (see revised manuscript, Section 5.3 "Ablation Study" and Appendix A.6 "Modality Ablation")

Dataset1–SCS

ModalityACCRECPREF1AUCAUPRMCCF-rank
M162.3339.6772.7751.2167.6667.8027.7526.71
M263.4450.1868.5257.8068.1266.9628.0026.86
M358.2629.7069.5141.4760.6361.2620.2029.14
M456.8289.4854.2367.4754.6352.0517.8625.43
M1+M267.8068.1867.7767.8973.2070.8535.6821.86
M1+M367.5264.7068.8266.5773.1871.4735.2322.71
M1+M463.4179.8360.2168.5768.3564.9628.4722.57
M2+M364.4149.2270.9857.9869.5868.8130.3824.57
M2+M464.2260.3465.5562.7568.5666.6628.6126.29
M3+M460.1071.7158.4164.2663.5560.9020.7827.43
M1+M2+M370.3671.0970.1970.5578.2576.7440.8318.71
M1+M2+M469.6869.9869.8469.7775.7573.4239.5119.71
M1+M3+M467.6460.7670.6565.2173.0271.7535.7222.00
M2+M3+M466.4668.9965.7567.2671.7669.3433.0423.43
M1+M2+M3+M472.4267.4575.0670.9578.7377.5945.1718.00
Ms79.5673.0084.2978.1187.8087.2859.8112.57
Ms+M179.3873.4883.5178.0788.0987.7059.3213.86
Ms+M279.5673.9883.3578.3288.1087.6059.5613.36
Ms+M379.3673.5783.4378.0788.0488.0459.2714.00
Ms+M479.2673.2183.4877.9087.9488.0059.0915.00
Ms+M1+M280.6575.6284.2279.6088.3588.3561.724.50
Ms+M1+M380.3175.1783.9479.2288.3588.2761.066.64
Ms+M1+M479.9574.4983.7878.7788.0787.8460.3810.43
Ms+M2+M380.3474.9984.0879.2088.2788.0761.127.14
Ms+M2+M479.8974.3783.8678.7088.2488.0360.329.86
Ms+M3+M479.7073.9883.6878.4488.1588.0459.9010.64
Ms+M1+M2+M381.1177.1983.9180.3288.6888.1462.533.29
Ms+M1+M2+M480.9476.5184.0780.0388.4488.0862.234.00
Ms+M1+M3+M480.6375.9983.9079.6788.4588.1561.644.86
Ms+M2+M3+M480.3775.7783.6579.4188.4287.8661.147.57
Full81.3178.4183.6180.8988.7988.3862.832.86

Dataset1–FCS

ModalityACCRECPREF1AUCAUPRMCCF-rank
M169.6570.0669.6369.7570.0666.6439.4015.14
M255.2437.0558.4145.2160.2459.3311.3327.57
M352.0825.2054.8434.3353.5253.225.0529.57
M448.7886.6749.4362.8852.0852.38-6.2025.29
M1+M265.3762.2166.5664.2173.2172.0230.8919.14
M1+M362.0350.2266.0956.9169.3668.2724.9222.57
M1+M450.1787.9950.2463.8858.5357.97-2.2323.86
M2+M360.5550.0763.6555.9065.3664.2221.7124.29
M2+M457.7546.6660.2352.4363.1561.6416.0126.00
M3+M453.0224.4557.3834.0958.6057.377.4729.00
M1+M2+M365.9757.1869.5562.6574.1372.9732.5419.57
M1+M2+M463.0658.6764.6061.3571.0170.0226.3621.43
M1+M3+M459.9942.3765.6051.3762.6461.9821.4425.29
M2+M3+M459.7260.4459.7359.9866.8865.7419.5123.71
M1+M2+M3+M466.6155.5671.4362.4275.5074.5234.1318.71
Ms70.1466.1172.0168.8575.4073.5340.4914.71
Ms+M172.4065.9875.8570.4878.3577.2445.289.57
Ms+M270.9163.9974.4168.7177.0476.2042.3213.57
Ms+M369.5261.9973.3467.0276.4575.0239.6715.43
Ms+M468.4659.9872.3065.5075.7375.0237.5216.71
Ms+M1+M272.9867.0176.3171.2580.0879.1346.436.00
Ms+M1+M372.7166.5276.1170.8979.4078.6445.907.14
Ms+M1+M472.4466.1775.8870.5878.8277.6745.388.29
Ms+M2+M372.1165.5075.5570.1078.2077.0144.6710.57
Ms+M2+M471.5664.8375.1069.4877.5676.3743.6411.71
Ms+M3+M470.9464.0374.4268.7477.0175.6842.3613.14
Ms+M1+M2+M374.0268.1177.4272.3681.6280.8548.492.29
Ms+M1+M2+M473.7267.7877.1472.0480.6879.4047.923.43
Ms+M1+M3+M473.5067.4376.8671.7680.1079.2447.444.43
Ms+M2+M3+M473.0766.9876.4871.3079.1377.6546.626.29
Full74.4068.4777.4372.3482.1582.0048.851.57
Reviewer_euAmOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes ECMLF for cold-start protein–protein interaction prediction involving proteins with no PPI records used for training. It combines sequence-region features from a pretrained protein language model with four biomedical association graphs to model similarity between new and known proteins. Bridge reconstruction assists representation learning, while adaptive gated fusion is used for PPI prediction. The authors also propose sequence- and graph-based explanations. On two datasets under semi- and full-cold-start settings, they report improvements over seven baselines and evaluate the proposed explanations.

Critical strengths
  1. The problem and evaluation settings are valuable. Predicting interactions involving proteins without PPI records available for training is an important challenge. Holding out proteins and evaluating both known–new and new–new pairs addresses this setting more directly than a random split of protein pairs.

  2. The reported FCS results are promising. ECMLF performs well against the selected baselines. Relative to its sequence-only variant, the full model’s AUC gain is larger under FCS than SCS on both datasets: 6.75 versus 0.99 percentage points on Dataset1, and 6.09 versus 3.07 percentage points on Dataset2. This pattern suggests that the combined modalities may be particularly useful when neither protein in a test pair has PPI records available during training.

Critical weaknesses
  1. The overall method is not sufficiently distinguished as a cold-start-specific contribution. The use of sequence and biomedical graph information is appropriate for predicting interactions involving new proteins. However, the proposed encoding, protein–entity reconstruction, and pairwise fusion can also be applied to general multimodal PPI prediction. At inference, the same encoders and fusion head are applied regardless of whether a protein is known or new. The paper includes component ablations showing performance benefits, but these do not fully establish what is methodologically specific to the absence of PPI records for new proteins. Consequently, I find the claimed cold-start methodological contribution insufficiently substantiated.

  2. The experimental evidence leaves important questions unresolved. The manuscript and appendix explain the protein-level splits, but do not specify which STRING network and score threshold define positive PPI labels, how negative protein pairs are selected, or the resulting positive and negative counts. These choices affect the task definition, difficulty, and interpretation of AUPR. The paper includes ESM2_AMP and a sequence-only variant, but additional sequence-only comparisons would make the assessment of the multimodal gains more convincing. The adaptation of baselines that depend on observed PPI topology is also described only briefly. These gaps do not show that the datasets or comparisons are incorrect, but they limit confidence in the reported gains.

Questions for authors
  1. How were PPI labels constructed in each dataset? Please specify the STRING network type and score threshold, the numbers of positive and negative pairs, and how negative pairs were selected. If these details appear in separate supplementary materials, please summarize them here so that the evaluation protocol can be assessed.

  2. How were baselines that rely on PPI topology adapted when test proteins had no observed training PPI edges, and which edges were available to each model? Under the same protein-level splits, could the authors also report results for a minimal ESM-2-based pairwise classifier and representative sequence-only models such as TUnA and PIPR? If any model is incompatible with the PPI labels used here, please explain the incompatibility.

  3. Does the full model’s advantage over its sequence-only variant persist for new proteins with few biomedical graph associations? Results stratified by the amount of available graph information, including proteins with no graph associations if present, would clarify the scope of the FCS gains.

2 comments

Q1: We used the publicly available datasets released with MTV-PPI and HEENN as Dataset1 and Dataset2, respectively. The table below summarizes dataset details checked with the original publications and STRING database.

DatasetSTRING network typeScore thresholdPositiveNegative
Dataset1v11.0 functional association networkCombined score from the experimental and database evidence channels ≥0.419,23719,237
Dataset2v12.0 functional association networkCombined score across all evidence channels >0.952,03452,034

The negative samples are randomly drawn from protein pairs absent from the positive set at a 1:1 ratio. To reduce the risk of including laten true interactions, we screen these candidates against experimentally supported PPI records in IntAct database. Pairs supported by these records will be excluded and replaced through resampling.

Q2:(1) All comparative experiments used the same training, validation, and test sets. During training and validation, all models had access only to the PPI records of known proteins. PPI records of new proteins were used only for test evaluation. For topology-dependent baselines, we preserved their original encoders and model-specific non-PPI information, such as sequence features or entity associations, ensuring that PPI records of new proteins were unavailable during training and validation. (2) Under the same protein-level splits, we added a minimal ESM-2-based classifier and representative sequence-only models including TUnA and PIPR as new baselines (see revised manuscript, Section 5.2 "Comparative Performance with Baselines").

Q3: To assess whether the model’s advantage persists with few biomedical graph associations, we conducted a graph association ablation experiment comparing the full model with its sequence-only variant. Specifically, we progressively removed 25%, 50%, 75%, and 100% of each new protein’s biomedical graph associations and reported performance stratified by the amount of graph information remaining. Based on the experimental results, as graph associations decrease, the model’s performance gradually declines. But it generally outperforms the sequence-only variant, even for new proteins with few or no remaining graph associations (see revised manuscript, Appendix A.6 "Graph Association Ablation").

W1:(1) We introduce bridge reconstruction as an auxiliary supervision task to learn latent similarities between known and new proteins, enabling interaction patterns learned from known proteins to transfer to new proteins. This specific bridge reconstruction objective is absent from existing multimodal PPI methods like KGF-GNN and HEENN. (2) During training, the model learns interaction patterns from PPI records of known proteins only and learns latent similarities between known and new proteins through bridge reconstruction. So, at inference, the model relies on the same encoders and fusion head to transfer the learned interaction patterns to new proteins through latent similarities. (3) Bridge reconstruction is designed to address the absence of PPI records for new proteins by learning similarities through biomedical associations. The reconstruction ablation experiment has already demonstrated the importance of this component for cold-start PPI prediction (see revised manuscript, Appendix A.6 "Reconstruction Ablation"). To exclude the influence of sequence information and further assess the contribution of bridge reconstruction, we conducted an additional reconstruction ablation experiment without the sequence branch. The results further support the contribution of bridge reconstruction to cold-start prediction (see revised manuscript, Appendix A.6 "Further Reconstruction Ablation"). Beyond the performance changes shown by the reconstruction ablation above, we have added another experiment to further examine how bridge reconstruction mechanistically addresses the absence of PPI records for new proteins. In the Graph-Pathway Explainability section of the original manuscript, we evaluated the model’s ability to capture key bridge entities that connect new proteins to known proteins sharing similar interaction patterns, thereby enabling the model to transfer interaction patterns learned from known proteins to new proteins. For new proteins lacking PPI records, our bridge reconstruction task provides auxiliary supervision, guiding the model to capture these key bridge entities to uncover latent associations between new and known proteins. Therefore, we compared the full model with its variant without bridge reconstruction in their ability to capture these key entities. The results also show that removing bridge reconstruction reduces the model’s ability to identify key entities connecting known and new proteins (see revised manuscript, Appendix A.7 "Bridge Entity Identification Ablation").

The following tables summarize our additional experiments. Due to formatting constraints in the response, all results are presented in tabular form here, including those shown as figures in the revised manuscript.

Q2:Additional baselines. (see revised manuscript, Section 5.2 "Comparative Performance with Baselines")

SCS

DatasetModelACCRECPREF1AUCAUPRMCC
D1PIPR65.18±0.0746.15±1.7476.05±1.0057.42±1.0773.29±1.0574.24±0.2233.61±0.27
D1TUnA71.08±0.3861.79±7.6577.28±4.6968.39±2.8879.34±0.1179.23±0.4043.50±0.79
D1ESM2-MLP66.77±0.2164.30±6.0868.69±2.9166.27±1.8973.06±0.1574.36±0.1833.83±0.87
D1ECMLF81.31±1.2878.41±3.1083.61±1.4080.89±1.4088.79±0.9888.38±0.5462.83±2.35
D2PIPR65.90±0.7846.33±2.9075.34±0.3857.36±2.3373.69±1.0673.28±0.9634.15±1.51
D2TUnA75.32±0.6473.83±1.8775.78±1.6174.78±0.1882.44±0.6379.95±0.3850.64±1.29
D2ESM2-MLP71.08±1.1574.36±8.7369.60±2.0671.71±2.9878.69±1.9978.05±1.8642.62±2.88
D2ECMLF84.61±0.5977.44±1.0890.43±0.6083.43±0.5792.27±0.5092.72±0.2869.95±1.06

FCS

DatasetModelACCRECPREF1AUCAUPRMCC
D1PIPR54.00±2.0425.11±3.3065.09±3.9836.10±2.8257.27±0.6060.83±2.2613.13±0.08
D1TUnA68.58±1.5659.34±10.1175.83±6.8466.02±3.7477.85±0.1577.81±1.1439.35±0.11
D1ESM2-MLP64.75±0.7262.03±9.6768.08±5.4764.45±2.8271.89±1.0473.52±2.1130.49±0.16
D1ECMLF74.40±1.8568.47±7.8377.43±3.7972.34±3.7582.15±2.7582.00±2.6148.85±3.59
D2PIPR57.11±2.5829.95±0.0464.05±0.1040.82±0.0661.35±1.3561.20±1.7016.03±2.70
D2TUnA73.56±0.4270.62±3.1474.45±0.2072.47±1.7580.66±1.1477.64±0.9547.00±0.69
D2ESM2-MLP70.48±1.3073.32±6.8169.08±3.0370.97±1.6177.12±1.8075.89±0.5241.42±3.17
D2ECMLF79.56±1.8467.53±3.9288.78±1.0476.65±2.4588.99±1.5289.15±1.3460.90±2.82

Q3:Graph Association Ablation. (see revised manuscript, Appendix A.6 "Graph Association Ablation")

DatasetTaskMetric0%25%50%75%100%Sequence
D1SCSAUC88.8488.7188.3287.8987.0187.80
D1SCSMCC62.2863.1762.3162.3057.8861.91
D1FCSAUC82.9682.0281.6780.8579.7875.40
D1FCSMCC51.1349.7449.5849.1846.4348.15
D2SCSAUC92.6292.5792.1891.5790.4689.20
D2SCSMCC76.4472.4371.1669.5962.6462.84
D2FCSAUC89.3089.0288.7288.5487.3082.90
D2FCSMCC65.2663.5161.8159.1355.7651.43

W1(3):Further Reconstruction Ablation. (see revised manuscript, Appendix A.6 "Further Reconstruction Ablation")

DatasetTaskModelAUCAUPR
D1SCSECMLF73.5074.16
D1SCSNO-Recon63.1562.28
D1FCSECMLF75.1872.01
D1FCSNO-Recon60.8461.90
D2SCSECMLF68.7368.49
D2SCSNO-Recon57.3456.05
D2FCSECMLF68.3367.64
D2FCSNO-Recon56.9757.08

W1(3):Bridge Entity Identification Ablation. (see revised manuscript, Appendix A.7 "Bridge Entity Identification Ablation")

TaskModelCapturedUncaptured
SCSECMLF7079734060
SCSNO-Recon3399770860
FCSECMLF6690034079
FCSNO-Recon4132159658
Reviewer_hY65OpenReview ↗
Rating
4(+1)
Confidence
1
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The authors developed ECMLF, a framework for predicting PPI for newly discovered proteins with no recorded interactions (cold-start scenario). They argue that existing methods use few modalities, compute similarity between new and known proteins offline and per modality, and explain predictions in terms of feature contributions rather than biological mechanisms. To this end, they combine a sequence encoder, which partitions each protein into structural regions and embeds them with a pretrained LLM, with four KGs covering drug, disease, miRNA and lncRNA associations. The graph encoders are trained with a bridge reconstruction task so that similarity modeling is learned jointly with prediction. Explanations are given as attended sequence regions and as shared entities bridging new and known proteins. ECMLF is evaluated on two datasets in semi-cold-start and full-cold-start settings, showing improvement.

Critical strengths

I am no expert in the field, but to my understanding, the main strength is conceptual: the model learns similarity between new and known proteins during training rather than precomputing it per modality. Bridge reconstruction is a sensible way to do this. I like the dual-pathway framing, since the sequence and graph pathways answer different questions. I also appreciate that the explanations are quantitatively evaluated against annotations at several overlap thresholds.

Critical weaknesses

My background is more in classical computational biology than in deep learning, so I have concentrated on the data sources and the explanation evaluation. One point there is unclear to me and I have raised it as a question below.

Questions for authors

Table 5 shows that knowledge-graph coverage is partial (on Dataset1, 613 of 1,648 proteins have drug associations and 164 have lncRNA associations), so many held-out proteins will appear in only one or two graphs, or none. Is prediction performance for new proteins related to how many graph modalities they appear in? This seems relevant to the cold-start motivation, since the least characterized proteins are both the hardest cases and the ones with the least graph evidence.

1 comment

(1) The knowledge graphs are constructed by integrating existing association records from multiple databases. Therefore, not every protein is covered by all four graph modalities, but each protein in both datasets is covered by at least one modality. The table below shows the distribution of proteins by knowledge-graph coverage. Here, each row shows how many proteins have associations in nn graph modalities (n=1,2,3,4n = 1, 2, 3, 4) in the corresponding dataset.

Number of graph modalitiesNumber of proteins in Dataset1(%)Number of proteins in Dataset2(%)
11,264(76.70%)993(23.18%)
2322(19.54%)1,856(43.33%)
359(3.58%)1,047(24.45%)
43(0.18%)387(9.04%)
Total1,648 (100.00%)4,283 (100.00%)

The modality ablation in Table 3 of the original manuscript has already demonstrated that the model can achieve better predictive performance when trained with multiple graph modalities. Combining all four modalities achieves higher AUC than any evaluated single-modality, two-modality or three-modality configuration.

(2) To further assess how the number of available graph modalities affects prediction performance for new proteins, we conducted a modality coverage experiment. As shown in the table above, only three proteins in Dataset1 are covered by all four graph modalities. We therefore conducted this experiment on Dataset2 to obtain a larger set of test pairs. Unlike the modality ablation experiment in Table 3 of the original manuscript, which trains a separate model for each modality combination, this experiment uses the full model trained with all four graph modalities. During the testing phase, we selected new proteins covered by all four graph modalities as the test set and masked their graph modality information to simulate new proteins with modality coverage of nn (n=4,3,2,1n = 4, 3, 2, 1). Within each task, the test protein pairs and their labels remain fixed across all configurations.

SCS

CoverageModalityACCRECPREF1AUCAUPRMCCF-rank
1M160.9187.6558.7764.9874.8866.6037.8611.14
1M261.7587.2858.8465.2473.4865.0639.028.43
1M360.9987.5758.6764.9975.9268.1937.5910.29
1M461.0387.2458.8764.9371.9463.6137.5812.14
2M1+M261.4688.4458.6065.4575.1666.8039.567.00
2M1+M360.9288.4558.4965.2477.4169.7338.168.14
2M1+M460.9588.1258.7265.1874.3567.1538.089.86
2M2+M361.5988.0858.4765.4076.4968.6238.937.00
2M2+M461.7787.8058.7865.4672.4763.9939.187.57
2M3+M460.9888.0158.5165.1376.0368.5737.639.71
3M1+M2+M361.3289.1958.2265.5877.6869.9339.614.86
3M1+M2+M461.4588.8958.5165.6174.6067.3439.695.86
3M1+M3+M460.9188.8458.3465.3577.5370.1238.227.57
3M2+M3+M461.5688.4958.3165.5176.5168.8238.926.29
4M1+M2+M3+M461.3089.6358.0165.6977.7570.2339.634.14

FCS

CoverageModalityACCRECPREF1AUCAUPRMCCF-rank
1M172.9977.7462.8766.2685.7976.4149.5211.57
1M278.4972.4168.2168.9886.0276.3253.867.86
1M373.8578.0263.1067.0385.6075.6950.3511.00
1M474.1677.5663.8067.0885.8076.0250.7910.29
2M1+M277.8575.6266.7869.3086.4877.3854.215.71
2M1+M372.1479.2661.4466.2086.0876.7448.8311.43
2M1+M472.3778.8062.0866.2186.3177.1549.1810.29
2M2+M378.1375.4766.8469.5986.2476.6254.316.00
2M2+M478.2474.5267.3869.4286.3476.8754.225.86
2M3+M473.1679.2362.2566.9685.9876.2749.9710.57
3M1+M2+M377.1878.3065.3069.6186.7177.6854.194.57
3M1+M2+M477.3877.2965.8569.4786.8578.0054.145.14
3M1+M3+M471.7280.6460.8566.4486.4977.4048.899.29
3M2+M3+M477.6976.8965.9169.6886.4677.0654.145.86
4M1+M2+M3+M476.5479.4564.2369.4986.9778.1953.654.57

The results in the two tables above (see revised manuscript, Appendix A.6 "Modality Coverage Ablation") indicate that broader graph modality coverage generally improves prediction performance for new proteins Even when only one graph modality is available, the model retains predictive ability for new proteins.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.