Distilling and Adapting: A Topology-Aware Framework for Zero-Shot Interaction Prediction in Multiplex Biological Networks

Alana Deng; Sugitha Janarthanan; Yan Sun; Zihao Jing; Pingzhao Hu

cs.LGFeb 20, 2026

Distilling and Adapting: A Topology-Aware Framework for Zero-Shot Interaction Prediction in Multiplex Biological Networks

Authors: Alana Deng, Sugitha Janarthanan, Yan Sun, Zihao Jing, Pingzhao Hu

Organizations: Department of Computer Science, Western University, London, ON, Canada · Department of Biochemistry, Western University, London, ON, Canada

Abstract

Multiplex Biological Networks (MBNs), which represent multiple interaction types between entities, are crucial for understanding complex biological systems. Yet, existing methods often inadequately model multiplexity, struggle to integrate structural and sequence information, and face difficulties in zero-shot prediction for unseen entities with no prior neighbourhood information. To address these limitations, we propose a novel framework for zero-shot interaction prediction in MBNs by leveraging context-aware representation learning and knowledge distillation. Our approach leverages domain-specific foundation models to generate enriched embeddings, introduces a topology-aware graph tokenizer to capture multiplexity and higher-order connectivity, and employs contrastive learning to align embeddings across modalities. A teacher-student distillation strategy further enables robust zero-shot generalization. Experimental results demonstrate that our framework outperforms state-of-the-art methods in interaction prediction for MBNs, providing a powerful tool for exploring various biological interactions and advancing personalized therapeutics.

Figures & tables

Figure 1: Illustration of an example multiplex network and its supra-adjacency matrix.

A_{1}

and

A_{2}

are the layer-specific adjacency matrices for layer

1

and

2

C

is the inter-layer adjacency matrix, and

\hat{A}

is the supra-adjacency matrix that represents the example multiplex network

\mathcal{G}

Figure 2: (a) The overall training workflow of CAZI-MBN. (b) Two inference settings in CAZI-MBN: Transductive and zero-shot.

Figure 3: Illustration of the CAE module.

	Context	# of Nodes	ITs	# of ITs
DGIdb	Drug-Gene	1846	Interaction mechanisms	5
ChEMBL	Compound-Bacteria	9368	Response phenotypes	3
PINNACLE	Protein-Protein	7044	Interaction in different cell types	12
MetaConserve	Protein-Metabolite	329	Conservation in bacteria strains	4
TRRUST	Gene-Gene	2,862	Gene regulartory roles	3

Table 1: Dataset summary (ITs: Interaction Types)

Type	Benchmark Models
Sequence-based	MultiLayer Perceptron (MLP), XGBoost (XGB)
Single graph-based	GCN , GraphSAGE, Graph Transformer, DGI
Multiplex graph-based/Relation-Aware	MultiplexSAGE, DMGI, HDMI, R-GCN, R-GAT
Domain-Specific	(1) Drug-Gene & Compoud-Bacteria: CoSMIG, DGCL (2) Protein-Protein: PIPR, xCAPT5 (3) Protein-Metabolite: DrugBAN, MolTrans (4) Gene-Gene: DL-GGI, GENER

Table 2: Benchmark models

Setting	Model	AUROC	AUPRC	HS	SA
T	Graph Transformer	0.505 $\pm$ 0.007	0.514 $\pm$ 0.005	0.493 $\pm$ 0.004	0.508 $\pm$ 0.003
	HDMI	0.551 $\pm$ 0.006	0.557 $\pm$ 0.007	0.540 $\pm$ 0.012	0.511 $\pm$ 0.006
	DGCL	0.519 $\pm$ 0.005	0.531 $\pm$ 0.004	0.527 $\pm$ 0.007	0.512 $\pm$ 0.006
	CAZI-MBN	0.715 $\pm$ 0.007	0.729 $\pm$ 0.009	0.687 $\pm$ 0.011	0.684 $\pm$ 0.015
ZS	Graph Transformer	0.498 $\pm$ 0.011	0.502 $\pm$ 0.004	0.486 $\pm$ 0.004	0.504 $\pm$ 0.007
	DMGI	0.524 $\pm$ 0.008	0.528 $\pm$ 0.016	0.529 $\pm$ 0.004	0.502 $\pm$ 0.006

Table 3: Evaluation of transductive and zero-shot multiplex interaction prediction on DGIdb (top model per category).

Setting	Model	AUROC	AUPRC	HS	SA
T	DGI	0.651 $\pm$ 0.019	0.719 $\pm$ 0.005	0.738 $\pm$ 0.012	0.673 $\pm$ 0.011
	HDMI	0.663 $\pm$ 0.012	0.762 $\pm$ 0.014	0.789 $\pm$ 0.012	0.730 $\pm$ 0.006
	CoSMIG	0.661 $\pm$ 0.021	0.732 $\pm$ 0.011	0.755 $\pm$ 0.009	0.702 $\pm$ 0.025
	CAZI-MBN	0.812 $\pm$ 0.008	0.863 $\pm$ 0.006	0.889 $\pm$ 0.014	0.757 $\pm$ 0.011
ZS	Graph Transformer	0.628 $\pm$ 0.011	0.720 $\pm$ 0.008	0.725 $\pm$ 0.013	0.628 $\pm$ 0.014
	DMGI	0.652 $\pm$ 0.021	0.745 $\pm$ 0.015	0.756 $\pm$ 0.007	0.711 $\pm$ 0.009

Table 4: Evaluation of transductive and zero-shot multiplex interaction prediction on ChEMBL (top model per category).

Setting	Model	AUROC	AUPRC	HS	SA
T	Graph Transformer	0.701 $\pm$ 0.007	0.704 $\pm$ 0.015	0.700 $\pm$ 0.005	0.726 $\pm$ 0.024
	HDMI	0.773 $\pm$ 0.022	0.798 $\pm$ 0.023	0.776 $\pm$ 0.017	0.766 $\pm$ 0.016
	xCAPT5	0.781 $\pm$ 0.011	0.804 $\pm$ 0.014	0.726 $\pm$ 0.012	0.752 $\pm$ 0.015
	CAZI-MBN	0.831 $\pm$ 0.018	0.845 $\pm$ 0.011	0.751 $\pm$ 0.009	0.772 $\pm$ 0.013
ZS	Graph Transformer	0.687 $\pm$ 0.019	0.690 $\pm$ 0.017	0.680 $\pm$ 0.018	0.705 $\pm$ 0.019
	HDMI	0.748 $\pm$ 0.025	0.776 $\pm$ 0.021	0.757 $\pm$ 0.024	0.742 $\pm$ 0.014

Table 5: Evaluation of transductive and zero-shot multiplex interaction prediction on PINNACLE (top model per category).

Setting	Model	AUROC	AUPRC	HS	SA
T	DGI	0.668 $\pm$ 0.020	0.673 $\pm$ 0.019	0.665 $\pm$ 0.010	0.669 $\pm$ 0.010
	HDMI	0.726 $\pm$ 0.015	0.729 $\pm$ 0.021	0.715 $\pm$ 0.006	0.718 $\pm$ 0.015
	DrugBAN	0.737 $\pm$ 0.010	0.714 $\pm$ 0.012	0.710 $\pm$ 0.019	0.711 $\pm$ 0.008
	CAZI-MBN	0.752 $\pm$ 0.020	0.744 $\pm$ 0.012	0.779 $\pm$ 0.008	0.656 $\pm$ 0.014
ZS	Graph Transformer	0.651 $\pm$ 0.007	0.660 $\pm$ 0.020	0.652 $\pm$ 0.015	0.648 $\pm$ 0.015
	HDMI	0.709 $\pm$ 0.017	0.710 $\pm$ 0.014	0.704 $\pm$ 0.014	0.707 $\pm$ 0.009

Table 6: Evaluation of transductive and zero-shot multiplex interaction prediction on MetaConserve (top model per category).

Setting	Model	AUROC	AUPRC	HS	SA
T	DGI	0.841 $\pm$ 0.014	0.844 $\pm$ 0.011	0.845 $\pm$ 0.009	0.847 $\pm$ 0.006
	HDMI	0.878 $\pm$ 0.008	0.879 $\pm$ 0.014	0.883 $\pm$ 0.005	0.881 $\pm$ 0.004
	GENER	0.867 $\pm$ 0.012	0.868 $\pm$ 0.006	0.865 $\pm$ 0.010	0.869 $\pm$ 0.006
	CAZI-MBN	0.905 $\pm$ 0.013	0.872 $\pm$ 0.008	0.791 $\pm$ 0.023	0.784 $\pm$ 0.015
ZS	DGI	0.832 $\pm$ 0.011	0.835 $\pm$ 0.014	0.836 $\pm$ 0.007	0.838 $\pm$ 0.006
	HDMI	0.868 $\pm$ 0.006	0.869 $\pm$ 0.010	0.872 $\pm$ 0.006	0.870 $\pm$ 0.004

Table 7: Evaluation of transductive and zero-shot multiplex interaction prediction on TRRUST (top model per category).

Figure 4: Average metric-wise performance drop across all five datasets when each module is ablated.

Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Interaction Types	# Nodes	# Edges	# Interaction Types	Multiplex Rate (%)
DGIdb	Inhibitor (7,004)	1,846	8,443	5	1.34
	Modulator (1,439)
	Agonist (915)
	Activator (104)
	Other (170)
ChEMBL	S (120,265)	9,368	248,582	3	14.45

Appendix

Table 8: Summary Statistics of Datasets.

Component	Value/Detail
Optimizer	Adam
Learning rate	1e-4
Batch size	256
Number of epoches	10000
Early stopping	Yes
Order of UGT	5

Appendix

Table 9: Training details of CAZI-MBN.

Setting	Model	AUROC	AUPRC	HS	SA
T	XGB	0.350 $\pm$ 0.010	0.389 $\pm$ 0.011	0.579 $\pm$ 0.011	0.421 $\pm$ 0.011
	MLP	0.284 $\pm$ 0.013	0.375 $\pm$ 0.005	0.609 $\pm$ 0.009	0.407 $\pm$ 0.011
	XGB (LLM)	0.517 $\pm$ 0.012	0.572 $\pm$ 0.009	0.541 $\pm$ 0.014	0.488 $\pm$ 0.010
	MLP (LLM)	0.511 $\pm$ 0.011	0.561 $\pm$ 0.008	0.536 $\pm$ 0.010	0.503 $\pm$ 0.012
	GCN	0.482 $\pm$ 0.010	0.493 $\pm$ 0.005	0.511 $\pm$ 0.003	0.491 $\pm$ 0.003
	GraphSAGE	0.447 $\pm$ 0.005	0.467 $\pm$ 0.003	0.527 $\pm$ 0.006	0.473 $\pm$ 0.006

Appendix

Table 10: Evaluation of transductive and zero-shot multiplex interaction prediction on DGIdb.

Setting	Model	AUROC	AUPRC	HS	SA
T	XGB	0.515 $\pm$ 0.013	0.607 $\pm$ 0.018	0.699 $\pm$ 0.007	0.587 $\pm$ 0.012
	MLP	0.531 $\pm$ 0.005	0.629 $\pm$ 0.012	0.657 $\pm$ 0.009	0.613 $\pm$ 0.010
	XGB (LLM)	0.648 $\pm$ 0.014	0.694 $\pm$ 0.016	0.735 $\pm$ 0.010	0.721 $\pm$ 0.011
	MLP (LLM)	0.624 $\pm$ 0.008	0.678 $\pm$ 0.013	0.718 $\pm$ 0.011	0.697 $\pm$ 0.009
	GCN	0.632 $\pm$ 0.014	0.700 $\pm$ 0.008	0.709 $\pm$ 0.004	0.641 $\pm$ 0.012
	GraphSAGE	0.631 $\pm$ 0.005	0.698 $\pm$ 0.012	0.734 $\pm$ 0.006	0.663 $\pm$ 0.005

Appendix

Table 11: Evaluation of transductive and zero-shot multiplex interaction prediction on ChEMBL.

Setting	Model	AUROC	AUPRC	HS	SA
T	XGB	0.650 $\pm$ 0.007	0.644 $\pm$ 0.007	0.651 $\pm$ 0.005	0.660 $\pm$ 0.023
	MLP	0.676 $\pm$ 0.008	0.668 $\pm$ 0.008	0.656 $\pm$ 0.011	0.666 $\pm$ 0.016
	XGB (LLM)	0.742 $\pm$ 0.010	0.767 $\pm$ 0.009	0.744 $\pm$ 0.007	0.734 $\pm$ 0.018
	MLP (LLM)	0.733 $\pm$ 0.011	0.751 $\pm$ 0.010	0.699 $\pm$ 0.010	0.708 $\pm$ 0.014
	GCN	0.659 $\pm$ 0.010	0.670 $\pm$ 0.012	0.666 $\pm$ 0.014	0.679 $\pm$ 0.021
	GraphSAGE	0.692 $\pm$ 0.017	0.677 $\pm$ 0.008	0.686 $\pm$ 0.005	0.693 $\pm$ 0.025

Appendix

Table 12: Evaluation of transductive and zero-shot multiplex interaction prediction on PINNACLE.

Setting	Model	AUROC	AUPRC	HS	SA
T	XGB	0.611 $\pm$ 0.006	0.612 $\pm$ 0.023	0.617 $\pm$ 0.021	0.613 $\pm$ 0.010
	MLP	0.617 $\pm$ 0.010	0.614 $\pm$ 0.015	0.620 $\pm$ 0.017	0.628 $\pm$ 0.007
	XGB (LLM)	0.662 $\pm$ 0.009	0.671 $\pm$ 0.018	0.658 $\pm$ 0.020	0.669 $\pm$ 0.013
	MLP (LLM)	0.694 $\pm$ 0.012	0.702 $\pm$ 0.015	0.707 $\pm$ 0.016	0.703 $\pm$ 0.010
	GCN	0.630 $\pm$ 0.007	0.642 $\pm$ 0.022	0.628 $\pm$ 0.011	0.633 $\pm$ 0.018
	GraphSAGE	0.649 $\pm$ 0.013	0.646 $\pm$ 0.011	0.644 $\pm$ 0.021	0.645 $\pm$ 0.012

Appendix

Table 13: Evaluation of transductive and zero-shot multiplex interaction prediction on MetaConserve.

Setting	Model	AUROC	AUPRC	HS	SA
T	XGB	0.786 $\pm$ 0.006	0.791 $\pm$ 0.016	0.794 $\pm$ 0.005	0.788 $\pm$ 0.007
	MLP	0.796 $\pm$ 0.008	0.799 $\pm$ 0.019	0.801 $\pm$ 0.011	0.803 $\pm$ 0.012
	XGB (LLM)	0.854 $\pm$ 0.019	0.863 $\pm$ 0.027	0.852 $\pm$ 0.014	0.867 $\pm$ 0.018
	MLP (LLM)	0.861 $\pm$ 0.022	0.868 $\pm$ 0.031	0.856 $\pm$ 0.017	0.859 $\pm$ 0.021
	GCN	0.810 $\pm$ 0.008	0.815 $\pm$ 0.022	0.812 $\pm$ 0.008	0.817 $\pm$ 0.005
	GraphSAGE	0.822 $\pm$ 0.010	0.826 $\pm$ 0.007	0.819 $\pm$ 0.008	0.824 $\pm$ 0.010

Appendix

Table 14: Evaluation of transductive and zero-shot multiplex interaction prediction on TRRUST.

Figure 5: Interaction type-specific performance analysis across five datasets, evaluated by prediction accuracy.

Figure 6: Module-wise ablation analysis across five MBNs under transductive and zero-shot settings. Bars show the performance drop for four evaluation metrics when removing each module (UGT, LLMs, CAE, MoE). Error bars reflect the variance across interaction types.

Variant	Description
w/o UGT	Remove UGT; use only LLM-derived features as input
w/o LLMs	Replace LLM features with 6-mer and Morgan fingerprints
w/o CAE	Replace CAE with a single GCN per interaction type
w/o MoE	Replace MoE with a plain MLP per interaction type
w/o $L_{dis}$	Discard discriminator
w/o $L_{reg}$	Discard consensus regularizer and replace with attention-only fusion for Z

Appendix

Table 15: Summary of ablation variants.

Dataset	Model	AUROC	AUPRC	HS	SA
DGIdb	w/o UGT	0.666 $\pm$ 0.015	0.705 $\pm$ 0.008	0.639 $\pm$ 0.012	0.634 $\pm$ 0.017
	w/o LLMs	0.563 $\pm$ 0.021	0.579 $\pm$ 0.023	0.591 $\pm$ 0.014	0.582 $\pm$ 0.016
	w/o CAE	0.601 $\pm$ 0.014	0.635 $\pm$ 0.018	0.613 $\pm$ 0.016	0.602 $\pm$ 0.012
	w/o MoE	0.672 $\pm$ 0.019	0.707 $\pm$ 0.011	0.645 $\pm$ 0.020	0.637 $\pm$ 0.018
	w/o $L_{dis}$	0.692±0.024	0.701±0.017	0.673±0.019	0.655±0.022
	w/o $L_{reg}$	0.679 $\pm$ 0.012	0.696 $\pm$ 0.013	0.660 $\pm$ 0.010	0.644 $\pm$ 0.014

Appendix

Table 16: Ablation results in transductive setting across five MBNs.

Dataset	Model	AUROC	AUPRC	HS	SA
DGIdb	w/o UGT	0.633 $\pm$ 0.015	0.664 $\pm$ 0.010	0.637 $\pm$ 0.012	0.625 $\pm$ 0.014
	w/o LLMs	0.527 $\pm$ 0.017	0.553 $\pm$ 0.025	0.599 $\pm$ 0.018	0.573 $\pm$ 0.015
	w/o CAE	0.589 $\pm$ 0.019	0.607 $\pm$ 0.020	0.618 $\pm$ 0.009	0.604 $\pm$ 0.012
	w/o MoE	0.641 $\pm$ 0.012	0.662 $\pm$ 0.015	0.639 $\pm$ 0.011	0.631 $\pm$ 0.011
	w/o $L_{dis}$	0.643 $\pm$ 0.017	0.677 $\pm$ 0.009	0.819 $\pm$ 0.022	0.690 $\pm$ 0.015
	w/o $L_{reg}$	0.593±0.012	0.616±0.013	0.621±0.011	0.601±0.012

Appendix

Table 17: Ablation results in zero-shot setting across five MBNs for CAZI-MBN(KD).

	DGIdb	TRRUST	PINNACLE	MetaConserve
GAPDH ( gapA )	34 / 41 (82.9%)	2 / 3 (66.7%)	103 / 140 (73.6%)	–
PC ( accC )	2 / 2 (100%)	2 / 2 (100%)	–	–
RPS14 ( rpsK )	3 / 5 (60%)	–	340 / 454 (74.9%)	–
OAT ( orn )	1 / 1 (100%)	2 / 2 (100%)	11 / 16 (68.7%)	–
NDUFAB1 ( fabD )	3 / 3 (100%)	–	46 / 57 (80.7%)	5 / 8 (62.5%)
Total	43 / 52 (82.7%)	6 / 7 (85.7%)	500 / 666 (75.1%)	5 / 8 (62.5%)

Appendix

Table 18: Per-Gene/Protein Product Statistics Across Datasets a

	MetaConserve	DGIdb	TRRUST	PINNACLE	ChEMBL
# Nodes	329	1,846	2,862	7,044	9,368
# Interaction Types	4	5	3	12	3
Teacher Total GPU Hours	0.463 h	3.191 h	2.137 h	31.34 h	8.456 h
Teacher Time/Epoch	0.574 s	4.332 s	3.903 s	115.037 s	43.201 s
Student Total GPU Hours	0.00109 h	0.0567 h	0.0098 h	0.4275 h	0.114 h
Student Time/Epoch	0.005 s	0.129 s	0.078 s	3.802 s	1.311 s

Appendix

Table 19: Runtime and parameter statistics across datasets for teacher and student models.

Setting	Model	AUROC	AUPRC	HS	SA
T	XGB	0.786 $\pm$ 0.006	0.791 $\pm$ 0.016	0.788 $\pm$ 0.017	0.792 $\pm$ 0.014
	MLP	0.796 $\pm$ 0.008	0.799 $\pm$ 0.019	0.809 $\pm$ 0.013	0.812 $\pm$ 0.018
	GCN	0.810 $\pm$ 0.008	0.815 $\pm$ 0.022	0.821 $\pm$ 0.011	0.826 $\pm$ 0.010
	GraphSAGE	0.822 $\pm$ 0.010	0.826 $\pm$ 0.007	0.824 $\pm$ 0.014	0.829 $\pm$ 0.015
	Graph Transformer	0.832 $\pm$ 0.005	0.835 $\pm$ 0.015	0.841 $\pm$ 0.010	0.834 $\pm$ 0.012
	DGI	0.841 $\pm$ 0.014	0.844 $\pm$ 0.011	0.851 $\pm$ 0.013	0.853 $\pm$ 0.011

Appendix

Table 20: Evaluation of transductive and zero-shot multiplex interaction prediction on TRRUST (Post-Hoc Per-Label Calibration).

Figure 7: Attention weight distributions across all five multiplex biological datasets. Each pie chart shows the relative importance assigned to different interaction types within a dataset.

Link Prediction

Link prediction estimates missing or future relationships in graphs, from social ties and knowledge-graph facts to biological interactions.

See all 46 papers

KGOT: Unified Knowledge Graph and Optimal Transport Pseudo-Labeling for Molecule-Protein Interaction Prediction
Dec 10, 2025Jiayu Qin, Zhengquan Luo, Guy Tadmor +3Pseudo-Labeling
Department of Computer Science and Engineering, University at Buffalo, USA · Mohamed bin Zayed University of Artificial Intelligence, UAE · Weizmann Institute of Science, Israel
CrossLLM-Mamba: Multimodal State Space Fusion of LLMs for RNA Interaction Prediction
Feb 23, 2026Rabeya Tus Sadia, Qiang Ye, Qiang ChengCross-Modal Representation Learning State Space Models
Department of Computer Science, University of Kentucky, Lexington, KY, USA · Department of Mathematics, University of Kentucky, Lexington, KY, USA · Institute for Biomedical Informatics, University of Kentucky, Lexington, KY, USA
MKGR: Multimodal Knowledge-Graph Representation Learning for Cold-Start Protein-Protein Interaction Prediction
Jul 2, 2026Wenbo ZhangKG Embedding Protein-Protein Interaction Prediction
College of Computer and Information Science, Southwest University, Chongqing, Chongqing, China

Dec 10, 2025J/K move · Enter open · S save

KGOT: Unified Knowledge Graph and Optimal Transport Pseudo-Labeling for Molecule-Protein Interaction Prediction

Jiayu Qin, Zhengquan Luo, Guy Tadmor, Changyou Chen+2

Department of Computer Science and Engineering, University at Buffalo, USA · Mohamed bin Zayed University of Artificial Intelligence, UAE · Weizmann Institute of Science, Israel

Pseudo-Labeling

Open paper AskPDF arXiv

Predicting molecule-protein interactions (MPIs) is a fundamental task in computational biology, with crucial applications in drug discovery and molecular function annotation. However, existing MPI models face two major challenges. First, the scarcity of labeled molecule-protein pairs significantly limits model performance, as available datasets capture only a small fraction of biological relevant interactions. Second, most methods rely solely on molecular and protein features, ignoring broader biological context such as genes, metabolic pathways, and functional annotations that could provide essential complementary information. To address these limitations, our framework first aggregates diverse biological datasets, including molecular, protein, genes and pathway-level interactions, and then develop an optimal transport-based approach to generate high-quality pseudo-labels for unlabeled molecule-protein pairs, leveraging the underlying distribution of known interactions to guide label assignment. By treating pseudo-labeling as a mechanism for bridging disparate biological modalities, our approach enables the effective use of heterogeneous data to enhance MPI prediction. We evaluate our framework on multiple MPI datasets including virtual screening tasks and protein retrieval tasks, demonstrating substantial improvements over state-of-the-art methods in prediction accuracies and zero shot ability across unseen interactions. Beyond MPI prediction, our approach provides a new paradigm for leveraging diverse biological data sources to tackle problems traditionally constrained by single- or bi-modal learning, paving the way for future advances in computational biology and drug discovery.

Abstract

Link Prediction

Explore similar work

KGOT: Unified Knowledge Graph and Optimal Transport Pseudo-Labeling for Molecule-Protein Interaction Prediction

CrossLLM-Mamba: Multimodal State Space Fusion of LLMs for RNA Interaction Prediction

MKGR: Multimodal Knowledge-Graph Representation Learning for Cold-Start Protein-Protein Interaction Prediction