arxiv: 2605.09964 · v2 · submitted 2026-05-11 · 💻 cs.AI · q-bio.QM

Recognition: no theorem link

Learning the Interaction Prior for Protein-Protein Interaction Prediction: A Model-Agnostic Approach

Ziqi Gao , Chenyi Zi , Zijing Liu , Ziqiao Meng , Yu Li , Jia Li

Authors on Pith no claims yet

Pith reviewed 2026-05-13 03:52 UTC · model grok-4.3

classification 💻 cs.AI q-bio.QM

keywords protein-protein interaction predictiongraph prompt learningL3 rulebiological priorspair classificationplug-and-play modulecomplementarity prior

0 comments

The pith

L3-PPI turns protein pair classification into graph classification by adding controlled virtual length-3 paths that encode the complementarity prior.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows that many standard protein-protein interaction datasets follow the biological L3 rule, in which the number of length-3 paths between two proteins correlates with their likelihood of interacting. It introduces L3-PPI, a lightweight module that builds a prompt graph containing virtual L3 paths derived from existing protein embeddings and then solves the original pair-prediction task as a graph-level classification problem. The module can be inserted into any existing PPI predictor without changing its representation learning stage. If the approach works, it supplies a biologically grounded classification head that improves accuracy across different base models and datasets while remaining computationally cheap.

Core claim

The authors establish that the L3 rule constitutes a usable interaction prior of complementarity, demonstrate its presence in popular PPI benchmarks, and show that an L3-path-regularized graph prompt can be generated on the fly from protein representations to convert pair classification into graph classification, yielding consistent gains when attached to existing predictors.

What carries the argument

The L3-path-regularized graph prompt generator that creates a small auxiliary graph containing a controlled number of virtual length-3 paths and feeds it to a graph classifier.

Load-bearing premise

That the L3 rule supplies a generalizable prior that remains useful after the prompt graph is constructed from learned embeddings, without introducing artifacts or overfitting to the training distribution.

What would settle it

Running L3-PPI on a PPI dataset deliberately constructed or filtered to show no positive correlation between L3-path count and interaction labels, then measuring whether performance gains disappear or reverse.

Figures

Figures reproduced from arXiv: 2605.09964 by Chenyi Zi, Jia Li, Yu Li, Zijing Liu, Ziqiao Meng, Ziqi Gao.

**Figure 1.** Figure 1: (A). The L3 rule. This illustrates a typical pattern in PPI networks: two proteins, FAM161A and PRPF31, with similar interfaces, share many common neighbors. In a pattern graph, multiple L3 paths (e.g., the red path) appear between the query proteins, suggesting a high probability of PPI. (B). The limitation of using #L3 paths as a handcrafted feature. If the training and testing clusters have low connecti… view at source ↗

**Figure 2.** Figure 2: Empirical evidences supporting the L3 rule. (A). We compute the number of Li paths (#Li Paths, i = 2–7) for 500 positive (‘+’) and 500 negative (‘-’) links sampled from the SHS27k dataset. (B). We report the Pearson correlation coefficient (ρ) and mutual information (MI) between the number of Li paths (i = 2–7) and link labels (0/1) across two datasets. represents interactions, we compute node embeddings e… view at source ↗

**Figure 3.** Figure 3: The method overview of proposed L3-PPI. Our method operates in a plug-and-play manner, enabling lightweight integration as a classification head after any leading PPI predictor while keeping the predictor frozen. The whole framework contains a pre-trained Graph Neural Networks (GNNs) serving as a surrogate for L3 pattern recognition, learnable prompts representing virtual proteins within the initial prompt… view at source ↗

**Figure 4.** Figure 4: We present an example with K = 3. Query nodes (u, v), prompt nodes (v P 0 , vP 1 , vP 2 , vP 3 ), prompt structures (black dashed lines), and inserting manner (red solid lines) form an L3 pattern that contains three L3 paths. The structure is fixed. However, the vectorized embeddings of the K + 1 prompt nodes are learnable and shared across all query pairs, where K is a hyperparameter. 4.2.2. GATING NETWOR… view at source ↗

**Figure 5.** Figure 5: (A) Performance comparison of our method vs. 2 baselines on BS, ES and NS categories (Random partition). (B) We set the upper limit for the number of prompt L3 paths, denoted as K, to 50, and we demonstrate the relationship between the number of inferred (predicted) #L3 paths and the actual #L3 paths in the PPI network. ρ refers to the Pearson correlation coefficient. of PPI, accurately inferring this int… view at source ↗

**Figure 6.** Figure 6: The trend of performance with respect to changes in K. 6. Conclusion We propose L3-PPI, a model-agnostic framework that incorporates biologically grounded complementarity priors into protein-protein interaction (PPI) prediction through L3- path-regularized graph prompt learning. Our method generates virtual L3 paths and dynamically regulates path counts based on interaction likelihood, which is consisten… view at source ↗

**Figure 7.** Figure 7: Examples of different testset construction strategies for binary PPI prediction. Clustered scenario (BFS): To simulate the case where unknown proteins form densely connected clusters ( [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗

read the original abstract

Protein-protein interactions (PPIs) are fundamental to cellular function and disease mechanisms. Current learning-based PPI predictors focus on learning powerful protein representations but neglect designing specialized classification heads. They mainly rely on generic aggregating methods like concatenation or dot products, which lack biological insight. Motivated by the biological "L3 rule", where multiple length-3 paths between a pair of proteins indicate their interaction likelihood, our study addresses this gap by designing a biologically informed PPI classifier. In this paper, we provide empirical evidence that popular PPI datasets strongly support the L3 rule. We propose an L3-path-regularized graph prompt learning method called L3-PPI, which can generate a prompt graph with virtual L3 paths based on protein representations and controls the number of paths. L3-PPI reformulates the classification of protein embedding pairs into a graph-level classification task over the generated prompt graph. This lightweight module seamlessly integrates with PPI predictors as a plug-and-play component, injecting the interaction prior of complementarity to enhance performance. Extensive experiments show that L3-PPI achieves superior performance enhancements over advanced competitors.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

L3-PPI adds a prompt-graph wrapper around existing PPI models using the L3 rule, but the gains likely come from extra graph capacity rather than a fixed biological prior.

read the letter

The paper's core move is to take the L3 rule (more length-3 paths between proteins signal higher interaction chance) and build a lightweight module that turns a pair of embeddings into a small prompt graph with virtual L3 paths. It then runs graph-level classification on that prompt graph instead of a simple head. They first check that several standard PPI datasets actually follow the L3 pattern, which is a small but concrete piece of supporting evidence. The module is meant to plug into any base predictor without retraining the whole thing from scratch.

Referee Report

2 major / 2 minor

Summary. The paper claims that popular PPI datasets empirically support the biological L3 rule (multiple length-3 paths between a protein pair indicate interaction likelihood). It proposes L3-PPI, a model-agnostic L3-path-regularized graph prompt learning method that generates a prompt graph with virtual L3 paths from learned protein representations, controls path count via regularization, and reformulates pair classification as graph-level classification over the prompt graph. This lightweight plug-and-play module injects the complementarity prior into any base PPI predictor, with experiments showing superior performance gains over advanced competitors.

Significance. If the results hold, the work is significant for providing a biologically motivated, model-agnostic classifier head that can be added to existing PPI representation learners. It directly addresses the reliance on generic aggregation methods by incorporating the L3 rule as an inductive bias via prompt graphs, which could improve accuracy in downstream applications like disease mechanism modeling. The plug-and-play design and empirical dataset validation of the L3 rule are strengths if supported by rigorous controls.

major comments (2)

[Abstract and §3 (prompt graph generation and regularization)] The load-bearing claim that L3-PPI injects a non-redundant L3 prior of complementarity (rather than merely adding capacity) is not yet secured. The prompt graph and virtual L3 paths are constructed from protein embeddings learned by the base predictor on the same PPI data, and path count is controlled by a jointly optimized regularization term. This construction risks functioning as an adaptive graph classifier whose inductive bias may not match the empirical L3 rule; the performance gains could arise from the graph reformulation or extra parameters instead. Provide a specific ablation (e.g., replacing L3-specific virtual paths with random or generic paths while keeping the same capacity) and report the resulting performance delta in the experiments section.
[Abstract and Experiments section] The abstract asserts that datasets support the L3 rule and that L3-PPI yields superior enhancements, yet supplies no quantitative details on baseline comparisons, statistical significance, ablation studies, or explicit control of virtual path numbers. This leaves the central empirical claim without verifiable support. Include tables with exact metrics (e.g., AUC, F1, p-values), ablation results isolating the L3 component, and a description of how path count is parameterized and fixed across runs.

minor comments (2)

[Method section] Clarify notation for the prompt graph construction and the exact regularization term controlling virtual paths; ensure all equations are numbered and referenced consistently.
[Figures] Improve figure clarity for the L3-path prompt graph illustration and add captions that explicitly link visual elements to the L3 rule and complementarity prior.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. The comments highlight important ways to strengthen the validation that L3-PPI injects a biologically motivated prior rather than merely increasing model capacity. We address each major comment below and will incorporate the requested changes in the revised manuscript.

read point-by-point responses

Referee: [Abstract and §3 (prompt graph generation and regularization)] The load-bearing claim that L3-PPI injects a non-redundant L3 prior of complementarity (rather than merely adding capacity) is not yet secured. The prompt graph and virtual L3 paths are constructed from protein embeddings learned by the base predictor on the same PPI data, and path count is controlled by a jointly optimized regularization term. This construction risks functioning as an adaptive graph classifier whose inductive bias may not match the empirical L3 rule; the performance gains could arise from the graph reformulation or extra parameters instead. Provide a specific ablation (e.g., replacing L3-specific virtual paths with random or generic paths while keeping the same capacity) and report the resulting performance delta in the experiments section.

Authors: We agree that isolating the contribution of the L3-specific virtual paths is essential to substantiate the claim. The current experiments demonstrate overall gains, but do not yet include a direct control for path structure. In the revision we will add an ablation that replaces the learned L3 paths with random paths of identical length and count while preserving model capacity and the same regularization framework. Performance deltas (AUC and F1) between the L3 and random variants will be reported in the experiments section to show that the biological structure, rather than the graph reformulation or added parameters, drives the improvement. revision: yes
Referee: [Abstract and Experiments section] The abstract asserts that datasets support the L3 rule and that L3-PPI yields superior enhancements, yet supplies no quantitative details on baseline comparisons, statistical significance, ablation studies, or explicit control of virtual path numbers. This leaves the central empirical claim without verifiable support. Include tables with exact metrics (e.g., AUC, F1, p-values), ablation results isolating the L3 component, and a description of how path count is parameterized and fixed across runs.

Authors: We acknowledge that the abstract and experiments section would be strengthened by explicit quantitative support. We will revise the abstract to report key aggregate metrics (average AUC/F1 gains and associated p-values). In the experiments section we will add comprehensive tables containing all baseline comparisons, statistical significance tests, ablation results that isolate the L3 component, and a precise description of path-count control: the regularization coefficient is tuned on a validation split and then held fixed for all reported runs. revision: yes

Circularity Check

0 steps flagged

No circularity: L3 prior is external and method performance is empirically validated

full rationale

The paper motivates L3-PPI from the externally documented biological L3 rule and supplies separate empirical counts showing dataset support for that rule. The prompt-graph construction uses upstream protein embeddings as input to define virtual paths, but the claimed performance gains are measured on held-out test sets via standard cross-validation against baselines; no equation or definition equates the output classifier score to a fitted parameter or to the input embeddings by algebraic identity. The regularization term is a standard capacity-control hyperparameter, not a self-referential re-labeling of the target labels. The derivation chain therefore remains self-contained and non-circular.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on the L3 rule as an external biological prior and on the assumption that generating virtual L3 paths from protein embeddings faithfully encodes interaction complementarity without circular dependence on the prediction target.

axioms (1)

domain assumption Multiple length-3 paths between a protein pair indicate higher interaction likelihood (L3 rule)
Invoked as motivation and supported by empirical evidence on popular PPI datasets

invented entities (1)

Prompt graph containing virtual L3 paths no independent evidence
purpose: To convert protein-pair classification into a graph-level classification task that injects the L3 interaction prior
Generated from protein representations with control over the number of paths; no independent evidence provided outside the method itself

pith-pipeline@v0.9.0 · 5505 in / 1297 out tokens · 126102 ms · 2026-05-13T03:52:39.626614+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Reference graph

Works this paper leans on

19 extracted references · 19 canonical work pages · 2 internal anchors

[1]

Independent se (3)- equivariant models for end-to-end rigid protein docking

Ganea, O.-E., Huang, X., Bunne, C., Bian, Y ., Barzilay, R., Jaakkola, T., and Krause, A. Independent se (3)- equivariant models for end-to-end rigid protein docking. arXiv preprint arXiv:2111.07786,

work page arXiv
[2]

Hierarchical graph learn- ing for protein–protein interaction.Nature Communica- tions, 14(1):1093, 2023a

Gao, Z., Jiang, C., Zhang, J., Jiang, X., Li, L., Zhao, P., Yang, H., Huang, Y ., and Li, J. Hierarchical graph learn- ing for protein–protein interaction.Nature Communica- tions, 14(1):1093, 2023a. Gao, Z., Niu, Y ., Cheng, J., Tang, J., Li, L., Xu, T., Zhao, P., Tsung, F., and Li, J. Handling missing data via max- entropy regularized graph autoencoder. ...

work page arXiv
[3]

and Rost, B

Hamp, T. and Rost, B. Evolutionary profiles improve protein–protein interaction prediction from sequence. Bioinformatics, 31(12):1945–1950,

work page 1945
[4]

Semi-Supervised Classification with Graph Convolutional Networks

Kipf, T. Semi-supervised classification with graph con- volutional networks.arXiv preprint arXiv:1609.02907,

work page internal anchor Pith review Pith/arXiv arXiv
[5]

Deep neural network based predictions of protein interactions using primary sequences.Molecules, 23(8):1923,

Li, H., Gong, X.-J., Yu, H., and Zhou, C. Deep neural network based predictions of protein interactions using primary sequences.Molecules, 23(8):1923,

work page 1923
[6]

Graphprompt: Uni- fying pre-training and downstream tasks for graph neural networks

Liu, Z., Yu, X., Fang, Y ., and Zhang, X. Graphprompt: Uni- fying pre-training and downstream tasks for graph neural networks. InProceedings of the ACM web conference 2023, pp. 417–428,

work page 2023
[7]

Learning un- known from correlations: Graph neural network for inter-novel-protein interaction prediction.arXiv preprint arXiv:2105.06709,

Lv, G., Hu, Z., Bi, Y ., and Zhang, S. Learning un- known from correlations: Graph neural network for inter-novel-protein interaction prediction.arXiv preprint arXiv:2105.06709,

work page arXiv
[8]

Inversiongnn: A dual path network for multi-property molecular optimization.arXiv preprint arXiv:2503.01488,

Niu, Y ., Gao, Z., Xu, T., Liu, Y ., Bian, Y ., Rong, Y ., Huang, J., and Li, J. Inversiongnn: A dual path network for multi-property molecular optimization.arXiv preprint arXiv:2503.01488,

work page arXiv
[9]

S., Smith, A

Salwinski, L., Miller, C. S., Smith, A. J., Pettit, F. K., Bowie, J. U., and Eisenberg, D. The database of interacting pro- teins: 2004 update.Nucleic acids research, 32(suppl 1): D449–D451,

work page 2004
[10]

Saprot: Protein language modeling with structure-aware vocabulary.BioRxiv, pp

Su, J., Han, C., Zhou, Y ., Shan, J., Zhou, X., and Yuan, F. Saprot: Protein language modeling with structure-aware vocabulary.BioRxiv, pp. 2023–10,

work page 2023
[11]

L., Fang, T., Doncheva, N

Szklarczyk, D., Kirsch, R., Koutrouli, M., Nastou, K., Mehryary, F., Hachilif, R., Gable, A. L., Fang, T., Doncheva, N. T., Pyysalo, S., et al. The string database 10 Learning the Interaction Prior for Protein-Protein Interaction Prediction: A Model-Agnostic Approach in 2023: protein–protein association networks and func- tional enrichment analyses for an...

work page 2023
[12]

T., Yang, X

Wang, G., Liu, X., Wang, K., Gao, Y ., Li, G., Baptista-Hon, D. T., Yang, X. H., Xue, K., Tai, W. H., Jiang, Z., et al. Deep-learning-enabled protein–protein interaction anal- ysis for prediction of sars-cov-2 infectivity and variant evolution.Nature Medicine, 29(8):2007–2018,

work page 2007
[13]

V ., and Li, S

Wu, L., Tian, Y ., Huang, Y ., Li, S., Lin, H., Chawla, N. V ., and Li, S. Z. Mape-ppi: Towards effective and efficient protein-protein interaction prediction via microenvironment-aware protein embedding.arXiv preprint arXiv:2402.14391,

work page arXiv
[14]

How Powerful are Graph Neural Networks?

Xu, K., Hu, W., Leskovec, J., and Jegelka, S. How powerful are graph neural networks?arXiv preprint arXiv:1810.00826,

work page internal anchor Pith review Pith/arXiv arXiv
[15]

arXiv preprint arXiv:2203.06125 , year=

Zhang, Z., Xu, M., Jamasb, A., Chenthamarakshan, V ., Lozano, A., Das, P., and Tang, J. Protein representa- tion learning by geometric structure pretraining.arXiv preprint arXiv:2203.06125,

work page arXiv
[16]

L., and Li, X

Zhao, Z., Qian, P., Yang, X., Zeng, Z., Guan, C., Tam, W. L., and Li, X. Semignn-ppi: Self-ensembling multi- graph neural network for efficient and generalizable protein-protein interaction prediction.arXiv preprint arXiv:2305.08316,

work page arXiv
[17]

This appendix is outlined as follows: • Section A provides a detailed introduction to the dataset, including statistical information and collection methods

11 Learning the Interaction Prior for Protein-Protein Interaction Prediction: A Model-Agnostic Approach Technical Appendix In the technical appendix, we present the data processing techniques, detailed implementations of the experiments, and additional experimental results. This appendix is outlined as follows: • Section A provides a detailed introduction...

work page 2023
[18]

It comprises 2,497 proteins forming 11,188 PPIs, equally split into positive and negative cases

collection, is a widely adopted benchmark. It comprises 2,497 proteins forming 11,188 PPIs, equally split into positive and negative cases. Positive cases are derived from the Database of Interacting Proteins (DIP) (Salwinski et al., 2004), excluding proteins with fewer than 50 amino acids or ≥40% sequence identity. Full protein sequences are sourced from...

work page 2004
[19]

The selected solution has the highest overall training efficiency and the best generalization performance. 14 Learning the Interaction Prior for Protein-Protein Interaction Prediction: A Model-Agnostic Approach Dataset Partition BS ES NS Average Scheme GNN-PPI GNN-PPI+ GNN-PPI GNN-PPI+ GNN-PPI GNN-PPI+ GNN-PPI GNN-PPI+ SHS27k Random 86.02 88.93 68.71 73.4...

work page 1982