PaperlayerView exact revision on arXiv

Source-linked AI summary

SkipGNN: Predicting Molecular Interactions with Skip-Graph Networks

Kexin Huang, Cao Xiao, Lucas Glass, Marinka Zitnik, Jimeng Sun

arXiv:2004.14949v2q-bio.MNcs.LG

TL;DR

Existing GNNs emphasize direct similarity, although second-order similarity is useful for predicting molecular interactions. SkipGNN constructs a skip graph and iteratively fuses it with the original graph. Across four interaction networks, it achieves superior and robust performance, including up to 28.8% PR-AUC improvement, and learns biologically meaningful embeddings.

  • Problem

    Existing GNNs mainly use direct similarity, while molecular interaction prediction also requires information from useful second-order similarity.

  • Method

    SkipGNN constructs a second-order skip graph and uses an iterative fusion architecture to learn from both the skip graph and the original interaction graph.

  • Results

    Across four interaction networks, SkipGNN achieves superior and robust performance, outperforming existing methods by up to 28.8% PR-AUC and learning biologically meaningful embeddings.

  • Takeaways & Limitations

    Skip similarity enables SkipGNN to capture structural and evolutionary forces in molecular interaction networks, including networks with substantial missing interaction information.

  • Takeaways & Limitations

    The study focuses on networks in which all edges are of the same type and suggests adapting SkipGNN to heterogeneous networks as future work.

Abstract

from arXiv · show

Molecular interaction networks are powerful resources for the discovery. They are increasingly used with machine learning methods to predict biologically meaningful interactions. While deep learning on graphs has dramatically advanced the prediction prowess, current graph neural network (GNN) methods are optimized for prediction on the basis of direct similarity between interacting nodes. In biological networks, however, similarity between nodes that do not directly interact has proved incredibly useful in the last decade across a variety of interaction networks. Here, we present SkipGNN, a graph neural network approach for the prediction of molecular interactions. SkipGNN predicts molecular interactions by not only aggregating information from direct interactions but also from second-order interactions, which we call skip similarity. In contrast to existing GNNs, SkipGNN receives neural messages from two-hop neighbors as well as immediate neighbors in the interaction network and non-linearly transforms the messages to obtain useful information for prediction. To inject skip similarity into a GNN, we construct a modified version of the original network, called the skip graph. We then develop an iterative fusion scheme that optimizes a GNN using both the skip graph and the original graph. Experiments on four interaction networks, including drug-drug, drug-target, protein-protein, and gene-disease interactions, show that SkipGNN achieves superior and robust performance, outperforming existing methods by up to 28.8\% of area under the precision recall curve (PR-AUC). Furthermore, we show that unlike popular GNNs, SkipGNN learns biologically meaningful embeddings and performs especially well on noisy, incomplete interaction networks.

1 Introduction

Molecular interaction networks require prediction methods that capture both direct interactions and useful second-order similarity. SkipGNN addresses this gap with a dual-graph GNN and shows robust performance across diverse interaction networks.

  • Motivation: Indirect similarity is valuable in molecular networks because directly interacting entities, such as drugs and target proteins, are not necessarily biologically similar.In DTI networks, drugs binding the same protein can be similar even though a drug and its target are connected without being biologically similar.
  • Present work: SkipGNN constructs a skip graph to capture second-order similarity and fuses it with the original interaction graph for molecular interaction prediction.Its architecture passes neural messages through both direct interactions and second-order relationships.
  • Evaluation: SkipGNN outperforms random-walk, embedding, clustering, network-metric, and state-of-the-art GNN baselines across drug-drug, drug-target, protein-protein, and gene-disease networks.It is evaluated on two homogeneous and two heterogeneous interaction-network types.
  • Robustness: SkipGNN remains consistently stronger than baseline methods when large fractions of interactions are removed, indicating robust performance on highly incomplete networks.The authors attribute this robustness to the skip graph’s ability to preserve network structure under missing interaction information.
  • Embedding analysis: SkipGNN learns biologically meaningful embeddings, whereas a regular GCN does not, including separation of drugs from most proteins while preserving drug-target binding proximity.The embedding analysis is reported for a drug-target interaction network.
  • Related work: Standard GNNs propagate information through neighborhood aggregations, but messages from nodes farther than the central node have discounted propagation power.This limits their ability to capture information important for prediction beyond direct neighborhoods.

2 Methods

SkipGNN constructs a skip graph to represent second-order similarity and iteratively fuses information from the original and skip graphs for molecular interaction prediction. A decoder converts the resulting node embeddings into interaction probabilities.

  • 2.1 Construction of the skip graph: SkipGNN constructs a skip graph from two-hop neighbors, allowing neural messages to propagate along paths connecting indirectly related nodes.The skip adjacency matrix is obtained from sign(A A^T), converting positive counts of two-hop paths into edges.
  • 2.2.1 Iterative fusion: The model uses two interacting GNNs to propagate and transform embeddings on the original graph and skip graph before fusing them iteratively.Each graph’s update uses the latest embedding from the other graph rather than simply concatenating the two outputs.
  • 2.2.1 Iterative fusion: Two GNN layers capture up to second-order neighbors, and a summation gate produces the final embedding without an additional activation before decoding.The authors report that the summation gate performs best empirically.
  • 2.2.2 SkipGNN decoder: A decoder combines the embeddings of two target nodes and applies a neural binary classifier to estimate their interaction probability.The implementation uses concatenation as the combination function because it consistently performs best across network types.
  • 2.3 The SkipGNN algorithm: SkipGNN initializes node attributes with node2vec and trains its parameters end to end using mini-batched interaction pairs and binary cross-entropy loss.The trained model predicts f(i,j), the probability that entities i and j interact.

3 Results

SkipGNN is evaluated on four molecular interaction datasets against network-science and graph-machine-learning baselines. It achieves the strongest overall predictive ranking while the experiments also assess robustness, embeddings, ablations, and novel predictions.

  • 3 Results: The experiments investigate predictive power, robustness to noise and missing data, skip similarity, component ablations, and novel interaction predictions.The study evaluates these questions across Sections 3.2 through 3.6.
  • 3.1.1 Molecular interaction networks: The evaluation covers DTI, DDI, PPI, and GDI datasets using established train-validation-test splits, balanced negative sampling, repeated random runs, and PR-AUC and ROC-AUC metrics.The four datasets include BIOSNAP-DTI, BIOSNAP-DDI, HuRI-PPI, and DisGeNET-GDI.
  • 3.1.1 Molecular interaction networks: The baselines span direct network embeddings and graph neural networks, including DeepWalk, node2vec, struc2vec, VGAE, and GCN.The comparison includes seven predictors from network science and graph machine-learning fields.

3.2 Predicting molecular interactions

SkipGNN was evaluated across four molecular interaction tasks and compared with baseline methods. It ranked first overall and remained robust when trained on highly incomplete networks.

  • SkipGNN ranked first among 11 methods across all molecular interaction networks, with an average rank of 1.5 and statistically significant improvements over all baselines.It performed best on DTI and PPI and second best on DDI and GDI.
  • Up to 2.7% PR-AUC improvement over GCN, 8.8% over VGAE, 28.8% over DeepWalk, 20.4% over node2vec, and 15.6% over spectral clustering were reported.
  • SkipGNN gave the most robust results as interaction data became incomplete, maintaining strong performance with access to only 10% of interactions.The incomplete-network experiment trained methods with 10%, 30%, 50%, or 70% of edges and predicted the remainder.
  • VGAE fell to around 0.5 PR-AUC on highly incomplete DTI and DDI networks, whereas node2vec and GCN improved steadily as more edges were observed.Spectral clustering was robust but varied substantially across tasks.

3.4 SkipGNN learns meaningful embedding spaces

Embedding visualizations show that SkipGNN separates biomedical entity types while preserving biologically relevant proximity. This behavior is attributed to combining direct and skip similarity rather than aggregating only immediate-neighbor messages.

  • Figure 3 compares five-fold PR-AUC averages with 95% confidence intervals across four tasks and varying fractions of missing edges.The methods shown are node2vec, spectral clustering, VGAE, GCN, and SkipGNN.
  • SkipGNN separated drugs from proteins and diseases from genes in t-SNE embeddings, unlike GCN, which conflated the biomedical entity types.Both models used the same input embedding, isolating skip similarity as their architectural difference.
  • SkipGNN achieved silhouette scores of 0.114 versus 0.014 for GCN on DTI and 0.079 versus 0.018 on GDI.These scores indicate stronger separation of entity clusters for SkipGNN.
  • GCN cannot capture skip similarity because it aggregates messages only from direct neighbors, whereas SkipGNN aggregates from direct and indirect neighbors.

3.5 Ablation studies

Ablation studies examined SkipGNN’s fusion scheme, skip graph, aggregation gates, and related components. The results identify the skip graph and iterative fusion as especially important, while the summation gate performed best among tested alternatives.

  • Removing the skip graph reduced performance substantially relative to the fusion ablation, identifying the skip graph as a key driver of improvement.The fusion ablation replaces SkipGNN’s fusion with simple concatenation of GCN embeddings, while the skip-graph ablation degenerates to GCN.
  • The iterative fusion scheme was important, indicating that successful methods need to integrate direct and skip similarity.
  • The summation gate outperformed weighted-L1 and Hadamard aggregation, with Hadamard performing worst.Weighted-L1 had comparable or worse performance than the summation gate.
  • The complete SkipGNN component setup achieved the best ablation result across DDI, PPI, and DTI tasks.
  • Literature review of the top 10 predicted DDI pairs found supporting evidence for novel predictions, including a Warfarin–Clozapine interaction.

4 Discussion

SkipGNN uses skip similarity to improve molecular-interaction prediction and remains effective on incomplete networks, while future work includes broader network types.

  • SkipGNN uses skip similarity to better capture structural and evolutionary forces governing molecular interaction networks than current GNNs.
  • SkipGNN achieves superior, robust performance across key interaction-network prediction tasks, including highly incomplete networks.
  • The study focuses on networks whose edges share one type, leaving adaptation to heterogeneous networks and other biological networks as future directions.

A Experiments on the importance of each layer of GNN for biomedical link prediction

The layer-variation experiments support second-order neighborhood information as especially important for biomedical link prediction, while suggesting that immediate and two-hop information should be balanced.

  • TwoLayers-OriGraph aggregates two-hop information indirectly through one-hop neighbors, whereas OneLayer-OriGraph uses only immediate neighbors.
  • TwoLayers-SkipGraph directly connects nodes to their two-hop neighborhood and also incorporates indirect four-hop neighbors.
  • OneLayer-SkipGraph directly measures skip similarity using only two-hop neighbor information.
  • OneLayer-3Hops tests the significance of higher-order neighbors in the GNN architecture.
  • Two-hop neighborhood information is more predictive than one-hop information alone, while higher-order neighborhoods contribute less to interaction link prediction.OneLayer-SkipGraph substantially improves over OneLayer-OriGraph, whereas the OneLayer-SkipGraph to TwoLayers-SkipGraph improvement is smaller.
  • TwoLayers-OriGraph outperforms TwoLayers-SkipGraph, supporting a balance between immediate-neighbor and two-hop information.The comparison provides support for SkipGNN’s iterative fusion scheme.

B Details about baseline methods

The appendix describes baseline implementations spanning path statistics, random walks, spectral methods, autoencoders, and several GNN architectures, with parameters selected by validation search.

  • The L315 method counts length-3 paths between node pairs and normalizes those counts by pairwise node degrees; baseline parameters are chosen by random search on validation data.
  • Baseline methods include random-walk embeddings, spectral clustering, variational graph autoencoders, and multiple GNN architectures for comparison with SkipGNN.The listed GNN baselines are GCN, GIN, JK-Net, and MixHop.
  • DeepWalk, node2vec, and struc2vec learn node embeddings from random walks and use concatenated embeddings for link prediction.The embeddings are paired with logistic regression classifiers.
  • Spectral Clustering uses the top 16 normalized-Laplacian eigenvectors as node embeddings before producing link probabilities.
  • VGAE reconstructs the adjacency matrix with a two-layer GCN, while GCN directly applies two GCN layers to the original adjacency matrix.
  • GIN uses an MLP aggregation function, JK-Net uses cross-layer skip connections with a GIN backend, and MixHop propagates messages through multiple higher-order adjacency matrices.

C Potential novel hits for PPI, DTI, and GDI

The authors report additional SkipGNN-discovered potential novel interaction hits for protein-protein, drug-target, and gene-disease tasks.

  • Table 8 presents potential novel hits discovered by SkipGNN for PPI, DTI, and GDI tasks.

D A Network Heuristic Explanation

SkipGNN’s robustness on incomplete networks is attributed to skip graphs preserving network topology better than original graphs. This explanation is evaluated by comparing spectral-norm relative errors as edges are removed.

  • The authors hypothesize that skip graphs preserve topology better than original graphs, especially when interaction data are scarce.Other methods use only the original graph, whereas SkipGNN also uses the skip graph.
  • The analysis measures spectral-norm relative error between complete and incomplete graph Laplacians as edges are removed at rate p.Err(A, p) compares the spectral norms of L = A−D and Lp = Ap−Dp.
  • Skip graphs have much lower relative error than original graphs across almost all missing-edge settings, supporting better topology preservation.The comparison uses 100 missing-edge fractions on the DDI task.

E Biomedical Interaction Network Visualization

Figure 7 visualizes four biomedical interaction networks. The figure is presented as a network visualization rather than as a reported quantitative comparison.

  • Figure 7 provides a visualization of four biomedical interaction networks.
  • The supplied passages specify the figure’s visualization scope but do not report an outcome comparison.
  • The biomedical network visualization is identified as Fig. 7 in the accompanying text.
Loading 2004.14949v2…