Source-linked AI summary
Drug Similarity Integration Through Attentive Multi-view Graph Auto-Encoders
Tengfei Ma, Cao Xiao, Jiayu Zhou, Fei Wang
TL;DR
Drug similarity integration must handle heterogeneous, nonlinear, noisy information while balancing predictive accuracy and interpretability, particularly with scarce labels. The paper uses attentive multi-view graph autoencoders with semi-supervised and transductive variants, and reports improved predictive performance with interpretable, adaptive view integration.
Problem
Drug similarity learning must integrate heterogeneous, nonlinear, noisy drug features while remaining accurate and interpretable when labels are scarce.
Method
The paper integrates feature views with attentive multi-view graph autoencoders and extends them to semi-supervised and transductive settings.
Results
The proposed models significantly outperform baselines across single-view and multi-view evaluations, with attentive semi-supervised GAE often achieving the best ROC-AUC.
Takeaways & Limitations
Attentive integration provides task-wise feature relevance for interpretable similarity while adapting similarity weights to the data.
Abstract
from arXiv · showhide
Drug similarity has been studied to support downstream clinical tasks such as inferring novel properties of drugs (e.g. side effects, indications, interactions) from known properties. The growing availability of new types of drug features brings the opportunity of learning a more comprehensive and accurate drug similarity that represents the full spectrum of underlying drug relations. However, it is challenging to integrate these heterogeneous, noisy, nonlinear-related information to learn accurate similarity measures especially when labels are scarce. Moreover, there is a trade-off between accuracy and interpretability. In this paper, we propose to learn accurate and interpretable similarity measures from multiple types of drug features. In particular, we model the integration using multi-view graph auto-encoders, and add attentive mechanism to determine the weights for each view with respect to corresponding tasks and features for better interpretability. Our model has flexible design for both semi-supervised and unsupervised settings. Experimental results demonstrated significant predictive accuracy improvement. Case studies also showed better model capacity (e.g. embed node features) and interpretability.
1 Introduction
Drug similarity integration must accommodate heterogeneous, nonlinear, noisy data while remaining accurate and interpretable, especially when labels are scarce. The paper addresses this with attentive multi-view graph autoencoders that support adaptive fusion, unlabeled-data use, and robustness to noise.
- Motivation: Multiple drug data types create opportunities for more comprehensive similarity, but their heterogeneous relationships complicate integration.Different views can vary in relevance, biomedical relations can be nonlinear, and data may be noisy or sparsely labeled.
- Approach: The model treats each drug feature type as a view and integrates them with multi-view graph autoencoders.Drugs are represented as nodes, while graph convolutional networks embed multi-view node features and edges.
- Approach: An attentive view-selection scheme assigns task-wise relevance to views, producing adaptive and interpretable similarity measures.The attention mechanism is intended to model heterogeneous feature relevance across targeting tasks.
- Benefits: The autoencoder framework uses unlabeled graph information to support accurate prediction when labels are scarce.Its reconstruction loss acts as a regularizer that explicitly models graph structure.
- Benefits: The proposed methods are designed to extract representations relatively robust to noisy data, including positive-unlabeled interaction samples.This reduces the negative impact of unobserved interactions that may not indicate no interaction.
2 Related Work
Existing drug-similarity methods include neighborhood, random-walk, and other graph-based approaches, but they differ in how they use similarity structure. The related work highlights limited treatment of higher-order similarity in some existing methods.
- Prior approaches: Existing approaches include nearest-neighbor methods, random walks, unsupervised methods, and multiple kernel learning.These categories summarize prior strategies for combining or exploiting drug similarity information.
- Nearest-neighbor methods: Nearest-neighbor methods commonly predict from majority cases among neighbors.The cited discussion notes that many such methods construct neighborhoods from first-order similarity.
- Random-walk methods: Random-walk methods use the assumption that data points on the same manifold are likely to share semantic labels.Label propagation is given as an example of this approach.
3 Background
Graph convolutional networks learn node representations from graph features and connections, providing the foundation for graph autoencoders. Later graph autoencoders extend this framework to link prediction and related graph-learning tasks.
- Graph convolutional networks: Graph convolutional networks learn node embeddings from node features and their graph connections.These embeddings can be used for node classification.
- Graph convolutional networks: The GraphCNN layer uses a multi-layer neural network with a layer-wise propagation rule over an undirected graph and adjacency matrix.The graph representation includes nodes X and adjacency matrix A.
- Graph autoencoders: Graph autoencoders use GraphCNN components for node classification and link prediction.The cited prior models extend graph convolutional representations into an autoencoding framework.
- Graph autoencoders: The paper extends prior graph autoencoders by reconstructing links and node embeddings and enabling inductive prediction.The cited limitation of earlier models is that they reconstruct only edges and cannot work on unseen data.
4 Method
The method integrates drug-feature views as graphs, using GraphCNN-based autoencoders to learn embeddings and similarities, with attention for adaptive view weighting. It supports unsupervised, semi-supervised, and transductive formulations that reconstruct features or links while incorporating graph structure.
- Multi-view graph construction: Each drug feature type becomes a graph view, and the model derives an integrated node embedding and similarity matrix across views.Nodes represent drugs, while view-specific similarity matrices represent pairwise drug similarity.
- Similarity matrix fusion: The basic model normalizes and fuses view-specific similarity matrices into one adjacency matrix, then encodes nodes with GraphCNN and decodes embeddings back to the original feature space.The decoder uses a sigmoid transformation of the graph-propagated embedding.
- Unsupervised learning: Without labels, the autoencoder optimizes reconstruction loss and produces an integrated similarity matrix usable for tasks such as node clustering.This formulation is described as unsupervised multi-graph fusion and embedding.
- Attentive view selection: Attention assigns input-dependent weights to graph views, enabling nonlinear fusion and adaptive, task-relevant similarity integration.The scheme projects each adjacency matrix, normalizes weights across views, and uses a diagonal attention matrix to reduce parameters from N∗N to N.
- Semi-supervised extension: The semi-supervised extension combines label-prediction loss with autoencoder reconstruction loss, explicitly modeling graph structure as a regularizing signal.The reconstruction term replaces graph-Laplacian regularization in the stated objective.
- Transductive extension: The transductive extension uses training labels and test labels as latent variables to reconstruct predicted links, with a stability regularizer on the test-label variables.Inference jointly obtains neural-network parameters and latent test labels.
5 Experiment
The experiments evaluate multi-view drug representations for binary and multilabel DDI prediction, using heterogeneous drug data, several baselines, repeated hold-out validation, and case studies of attention-weighted similarity. Across datasets, the proposed models outperform baselines, while attention and multiview integration provide task-adaptive, interpretable similarity measures.
- Data sources: Dataset 1 integrates DDI, side-effect, off-label side-effect, and chemical-structure views for binary DDI prediction.The DDI view contains 645 drugs, 1,318 DDI events, and 63,473 reported drug pairs; chemical structure uses 1,024-bit ECFP6 fingerprints.
- Data sources: Dataset 2 predicts 1,301 specific DDI types among new drug pairs using multiple drug-characterization views.The dataset contains 222 drugs and includes drug indication and chemical protein interactome information.
- Evaluation: 25% and 50% drug hold-outs were repeated 50 times, with mean and standard deviation reported for evaluation.For non-testing data, 90% of drugs were used for training and 10% for validation and model selection.
- Results: The proposed models significantly outperform baselines on both single-view and multiview settings, while multiview models generally outperform corresponding single-view models.The baselines include nearest neighbor, label propagation, GraphCNN, and multiple kernel learning; attention assigns larger weights to more relevant feature types.
- Results: AttSemiGAE often achieves the best ROC-AUC by embedding node features, with a stronger advantage on Dataset 2 where all views have node features.On Dataset 1, where most views lack node features, attTransGAE performs better.
- Case studies: Attention-weighted case studies align feature importance with reported DDI mechanisms and improve integrated similarity for heterogeneous drug pairs.For Acyclovir–Ganciclovir, the integrated score is 0.682 versus 0.551 for label propagation; related examples produce 0.682–0.720 versus 0.551–0.630 for methods weighting views homogeneously.
6 Conclusion
The paper concludes that attention-based multi-view graph autoencoder models integrate drug similarities nonlinearly and adaptively while supporting interpretability. Experiments on two real-world drug datasets demonstrated predictive performance and efficacy, with future work targeting larger databases and direct cross-view similarity computation.
- Multi-view graph autoencoder models use attention to integrate drug similarities and select views.The models support nonlinear, adaptive integration and interpretable view selection.
- Experiments on two real-world drug datasets demonstrated the methods’ predictive performance and efficacy.
- The models were extended to semi-supervised and transductive settings for predicting unknown drug–drug interactions.
- Future work includes testing larger drug databases and computing integrated similarity directly across views.The larger-database direction is intended to exploit deep learning while avoiding overfitting.