DRPreter: Interpretable Anticancer Drug Response Prediction Using Knowledge-Guided Graph Neural Networks and Transformer

Jihye Shin, Yinhua Piao, Dongmin Bang, Sun Kim, Kyuri JoInternational Journal of Molecular Sciences 2022 · DOI

TL;DR

This paper introduces DRPreter, a novel interpretable model that predicts anticancer drug response by integrating biological pathway knowledge into graph neural networks and transformers. DRPreter not only achieves state-of-the-art prediction accuracy but also highlights key biological pathways involved in drug efficacy, paving the way for more personalized cancer treatments and drug discovery.

What Problem Does This Paper Address?

Predicting how a cancer patient will respond to a specific drug is crucial for personalized medicine, aiming to maximize treatment success and minimize adverse effects. While many computational models exist, they often lack interpretability, making it difficult to understand why a drug is predicted to be effective or ineffective. This "black box" nature hinders clinical adoption and drug discovery efforts, as it doesn't reveal the underlying biological mechanisms. Existing interpretable models often simplify biological pathways, failing to capture the intricate gene-gene interactions within them. The challenge, therefore, is to develop a model that is both highly accurate in predicting drug response and transparent in explaining the biological basis for its predictions, leveraging complex biological knowledge.

Methods

DRPreter employs a multi-faceted approach combining graph neural networks (GNNs) and transformers to achieve its goals:

  • Knowledge-Guided Cell Line Representation: Instead of a single, monolithic graph for a cell line, DRPreter breaks down the cell line's biological network into subgraphs, each representing a specific cancer-related biological pathway. These pathway subgraphs are then encoded using Graph Attention Networks (GATs) to capture intra-pathway gene interactions.
  • Drug Graph Representation: Drugs are represented as molecular graphs, where atoms are nodes and bonds are edges. A Graph Isomorphism Network (GIN) is used to learn embeddings from these drug graphs, capturing structural information.
  • Transformer-Based Fusion and Interpretation: A type-aware transformer encoder is the core of DRPreter's interpretability. It takes the pathway embeddings and drug embeddings as input and learns the complex interactions between them. This allows the model to identify which pathways are most influential for a given drug-cell line pair. The transformer's self-attention mechanism is key here, revealing how pathways relate to each other and to the drug.
  • Gradient-Weighted Gene Node Interpretation: To further pinpoint key biological players, the authors use GradCAM to analyze gene importance within the model's predictions, highlighting specific genes that contribute significantly to the predicted drug response.

Datasets

The study primarily utilizes the Genomics of Drug Sensitivity in Cancer (GDSC) dataset for training and evaluating drug response predictions. For cell line data, they use transcriptomic data from the Cancer Cell Line Encyclopedia (CCLE). The biological pathway information is sourced from the KEGG pathway database.

Key Insights and Ideas

  • Pathways as Building Blocks: Decomposing cell line networks into pathway-specific subgraphs is more effective than using a single, large graph. This allows the model to learn more granular and biologically relevant representations.
  • Transformer for Drug-Pathway Synergy: The transformer architecture is crucial for understanding how drugs interact with specific biological pathways, going beyond simple gene-drug associations.
  • Interpretability as a Feature, Not an Add-on: DRPreter is designed from the ground up to be interpretable, using its architecture (transformer attention, GradCAM) to reveal the "why" behind predictions.
  • Bridging the Gap: The model successfully connects molecular-level drug structures and cellular pathway activity to predict drug response, offering a more holistic view.

Key Results

  • State-of-the-Art Performance: DRPreter outperforms existing graph-based drug-response prediction models on the GDSC dataset, demonstrating improved accuracy in predicting IC50 values. For example, it achieved a significantly lower Mean Squared Error (MSE) compared to other models.
  • Identified Key Pathways: The model successfully identified specific pathways (e.g., TGF-β signaling pathway) that are highly relevant to drug sensitivity for certain drug-cell line pairs, with findings supported by existing literature.
  • Case Study Validation: In a case study, DRPreter accurately predicted missing drug response values and identified known drug-cell line sensitivities, such as Bortezomib for multiple myeloma cell lines. It also highlighted potential novel sensitivities, like Rapamycin for MV-4-11 cells.
  • Gene-Level Insights: Gradient-weighted gene importance analysis identified known drug targets and genes involved in relevant pathways for specific drug-cell line pairs, further validating the model's interpretability.

Limitations and Future Work

  • Data Availability: The model's performance is dependent on the availability of comprehensive omics data and drug sensitivity profiles. The study notes that some cell lines were excluded due to missing data.
  • Generalizability: While validated on GDSC, further testing on diverse cancer types and drug classes would be beneficial.
  • Computational Cost: Integrating multiple GNNs and a transformer can be computationally intensive, although the paper doesn't explicitly detail this as a limitation.
  • Future Directions: The authors suggest that their interpretable model can be further applied to identify novel drug targets, predict drug combinations, and aid in the selection of personalized treatment strategies. The ability to interpret drug mechanisms could accelerate drug discovery and development.

Summary updated

Abstract

Some of the recent studies on drug sensitivity prediction have applied graph neural networks to leverage prior knowledge on the drug structure or gene network, and other studies have focused on the interpretability of the model to delineate the mechanism governing the drug response. However, it is crucial to make a prediction model that is both knowledge-guided and interpretable, so that the prediction accuracy is improved and practical use of the model can be enhanced. We propose an interpretable model called DRPreter (drug response predictor and interpreter) that predicts the anticancer drug response. DRPreter learns cell line and drug information with graph neural networks; the cell-line graph is further divided into multiple subgraphs with domain knowledge on biological pathways. A type-aware transformer in DRPreter helps detect relationships between pathways and a drug, highlighting important pathways that are involved in the drug response. Extensive experiments on the GDSC (Genomics of Drug Sensitivity and Cancer) dataset demonstrate that the proposed method outperforms state-of-the-art graph-based models for drug response prediction. In addition, DRPreter detected putative key genes and pathways for specific drug-cell-line pairs with supporting evidence in the literature, implying that our model can help interpret the mechanism of action of the drug.
BibTeX
@article{shin2022drpreter,
  author = {Jihye Shin and Yinhua Piao and Dongmin Bang and Sun Kim and Kyuri Jo},
  title = {{DRPreter: Interpretable Anticancer Drug Response Prediction Using Knowledge-Guided Graph Neural Networks and Transformer}},
  journal = {International Journal of Molecular Sciences},
  year = {2022},
  doi = {10.3390/ijms232213919},
}
Figure