Source-linked AI summary
DeepDDS: deep graph neural network with attention mechanism to predict synergistic drug combinations
J. Wang, X. Liu, S. Shen, L. Deng, H. Liu*
TL;DR
The paper addresses the difficulty of screening the enormous space of anticancer drug combinations experimentally. DeepDDS combines drug molecular graphs and cancer-cell gene-expression profiles through GNN, attention, MLP, and fully connected components, and outperforms competing methods across validation and independent testing. The authors conclude that it can prioritize synergistic combinations for further experimental validation, while noting limitations from the small benchmark size.
Problem
The huge number of possible drug combinations makes wet-lab screening difficult, creating a need for computational prioritization of synergistic combinations.
Method
DeepDDS encodes drug molecular graphs with GAT or GCN, cell-line gene-expression profiles with an MLP, and concatenates the embeddings for synergy classification.
Results
DeepDDS achieved higher performance than competing methods in five-fold validation, leave-one-out evaluations, and an independent test set; its top 100 predicted pairs included 98 experimentally validated synergistic combinations.
Takeaways & Limitations
DeepDDS is presented as a tool for prioritizing synergistic drug combinations for further wet-lab validation.
Takeaways & Limitations
Independent-test predictive accuracy remains limited, which the authors mainly attribute to the small training set containing 38 drugs and 39 cancer cell lines.
Abstract
from arXiv · showhide
Drug combination therapy has become a increasingly promising method in the treatment of cancer. However, the number of possible drug combinations is so huge that it is hard to screen synergistic drug combinations through wet-lab experiments. Therefore, computational screening has become an important way to prioritize drug combinations. Graph neural network have recently shown remarkable performance in the prediction of compound-protein interactions, but it has not been applied to the screening of drug combinations. In this paper, we proposed a deep learning model based on graph neural networks and attention mechanism to identify drug combinations that can effectively inhibit the viability of specific cancer cells. The feature embeddings of drug molecule structure and gene expression profiles were taken as input to multi-layer feedforward neural network to identify the synergistic drug combinations. We compared DeepDDS with classical machine learning methods and other deep learning-based methods on benchmark data set, and the leave-one-out experimental results showed that DeepDDS achieved better performance than competitive methods. Also, on an independent test set released by well-known pharmaceutical enterprise AstraZeneca, DeepDDS was superior to competitive methods by more than 16\% predictive precision. Furthermore, we explored the interpretability of the graph attention network, and found the correlation matrix of atomic features revealed important chemical substructures of drugs. We believed that DeepDDS is an effective tool that prioritized synergistic drug combinations for further wet-lab experiment validation.
Materials and methods
DeepDDS integrates molecular-graph representations of drugs with cancer-cell gene-expression features to classify drug-pair synergy. It uses GNN-based drug encoders, an MLP for cell-line features, and fully connected layers for prediction.
- Cell Line Feature Extraction based on MLP: CCLE gene-expression profiles are reduced to 954 selected genes and encoded by an MLP with two hidden layers.The selected genes come from the intersection of CCLE profiles and the LINCS Landmark gene set after redundant and non-coding transcripts are removed.
- Predicting the synergistic effect of drug combinations versus cell lines: Drug and cell-line embeddings are concatenated and passed through spindle-shaped fully connected layers for end-to-end binary classification of synergistic versus antagonistic combinations.The output probability is computed with a softmax function.
- Drug Representation based on GNN: Drug SMILES are converted into molecular graphs whose nodes are atoms and edges are chemical bonds for GNN-based representation learning.RDKit performs the conversion, and atomic features include atom symbol, adjacent-atom count, adjacent-hydrogen count, implicit value, and aromaticity.
- Graph Convolutional Network (GCN): The GCN receives a node feature matrix and adjacency matrix, then applies graph layers to learn node representations.The model adds self-connections and uses learnable parameters and activation functions during iterative propagation.
- Graph Convolutional Network (GCN): Three ReLU-activated GCN layers are followed by global max pooling to produce graph-level drug features.Max pooling performed better than sum and average pooling in the GCN-based DeepDDS model.
- Graph Attention Network (GAT): The GAT computes node features with multi-head attention, assigning normalized coefficients to each node and its first-order neighbors.The attention architecture uses separate parameters for each head and normalizes neighbor contributions with softmax.
Result
DeepDDS achieved strong predictive performance across cross-validation, leave-one-out evaluations, and an independent AstraZeneca test set. Its attention-based representations also exposed chemical-substructure patterns, although independent-test accuracy and mechanistic interpretation remained limited.
- Cross-validation: 0.93 ROC AUC and PR AUC, with 0.85 accuracy, balanced accuracy, precision, sensitivity, and specificity, were achieved by DeepDDS-GAT in five-fold cross-validation.DeepDDS-GAT and DeepDDS-GCN outperformed the compared methods across the reported performance measures.
- Cross-validation: 98 of the top 100 drug pairs ranked by DeepDDS-GAT had experimental validation as synergistic combinations across cancer cell lines.The pairs were selected by their highest predicted synergy scores.
- Leave-one-out evaluation: 0.89 AUC was achieved by DeepDDS-GAT in leave-one-drug-combination-out evaluation, outperforming the other compared methods.DeepDDS also performed better than competing methods when one drug was excluded from training.
- Leave-one-out evaluation: 0.84, 0.867, 0.821, 0.828, 0.843 and 0.775 ROC AUC were obtained by DeepDDS-GAT across breast, colon, lung, melanoma, ovarian and prostate leave-tissue-out evaluations, respectively.DeepDDS-GAT was reported as better than DeepSynergy and TranSynergy in this evaluation.
- Independent test set: 421/668 drug pairs were correctly predicted by DeepDDS-GAT on the AstraZeneca independent test set, versus 317/668 for DeepSynergy, a 16% difference.DeepDDS-GCN correctly predicted 402/668 pairs, corresponding to a 13% difference from DeepSynergy.
- Interpretability: Atom correlation matrices changed from no clear clustering before training to linking blocks after training, suggesting candidate chemical subgroups while in vivo pharmacological mechanisms remained unclear.The authors describe this interpretation as speculative.
- Limitations: Predictive accuracy remained limited on the independent test set, and the physicochemical meaning of molecular graphs and atom-level attention weights was not fully understood.The authors identify these as continuing limitations despite performance exceeding competing methods.
Author contributions statement
The authors divided responsibility for the study’s conception, manuscript preparation, data collection, experiments, supervision, funding, and review.
- J.W. and H.L. conceived the main idea and manuscript framework.
- J.W. drafted the manuscript, while J.W. and X. L. collected data and performed the experiments.
- L.D. and H.L. improved the idea and manuscript, supervised the study, and provided funding.
- S.S. reviewed drafts, and all authors read and commented on the paper.