drGAT: Attention-Guided Gene Assessment of Drug Response Utilizing a Drug-Cell-Gene Heterogeneous Network

Yoshitaka Inoue, Hunmin Lee, Tianfan Fu, Augustin Luna

TL;DR

This paper introduces drGT, a novel graph deep learning model that predicts how cell lines will respond to drugs by analyzing complex relationships between drugs, genes, and cell lines. Beyond prediction, drGT uses "attention mechanisms" to highlight which genes are most important for a drug's effect, offering biologically interpretable insights crucial for drug discovery.

What Problem Does This Paper Address?

Drug development is notoriously slow and expensive, with a high failure rate. While machine learning (ML) shows promise in predicting drug responses, a major hurdle is the "black box" nature of these models. In biomedicine, simply predicting an outcome isn't enough; understanding why a model makes a prediction is critical for validating findings, generating hypotheses, and ultimately, developing safer and more effective treatments. This paper tackles the challenge of building ML models that are both accurate in predicting drug response and transparent in explaining their reasoning, bridging the gap between predictive power and biological understanding.

Methods

The core of drGT is a heterogeneous graph neural network (GNN), specifically a Graph Transformer (GT). This means it doesn't just look at individual data points but models relationships between different types of entities.

  1. Heterogeneous Graph Construction: drGT builds a complex network where nodes represent drugs, cell lines, and genes. Edges connect these nodes based on various relationships: drug-cell line response, drug-gene interactions (like drug-target data), and cell line-gene relationships (gene expression).
  2. Feature Engineering: Input data like drug structures (SMILES strings converted to fingerprints) and gene expression are transformed into a common latent space using kernel-based similarity and linear projections.
  3. Graph Transformer (GT) Layers: The GT layers are the heart of the model. They use a "masked multi-head self-attention" mechanism. This allows the model to weigh the importance of different connections (edges) and nodes within the graph. Crucially, these attention mechanisms learn attention coefficients (ACs).
  4. Interpretability via Attention Coefficients (ACs): The ACs are the key to interpretability. They quantify how much influence each gene has on a drug's effect on a cell line. By analyzing these ACs, researchers can identify genes that are most predictive of drug response, offering insights into the underlying biological mechanisms.

Datasets

The study utilizes several publicly available pharmacogenomic datasets:

  • GDSC1 & GDSC2: Genomics of Drug Sensitivity in Cancer datasets.
  • CTRP: Cancer Therapeutics Response Portal.
  • NCI60: A panel of 60 human cancer cell lines.

These datasets are used for both training and evaluating the model's performance in various scenarios, including predicting responses for unseen drugs or cell lines, and transferring knowledge across datasets.

Key Insights and Ideas

  • Bridging Prediction and Interpretation: The most significant insight is drGT's ability to simultaneously achieve state-of-the-art predictive performance and provide biologically meaningful interpretations through attention coefficients. This addresses a major limitation in current ML for drug discovery.
  • Heterogeneous Graph Power: Building a unified graph that integrates drugs, cell lines, and genes allows the model to capture complex, multi-modal interactions that simpler models miss.
  • Attention as a Biological Compass: The attention coefficients learned by the Graph Transformer act as a powerful tool to pinpoint key genes involved in drug response, guiding researchers towards potential biomarkers and mechanisms.
  • Generalization Across Scenarios: drGT demonstrates robust performance not only in standard prediction tasks but also in more challenging scenarios like predicting responses for entirely new drugs or cell lines (zero-shot learning), highlighting its ability to generalize.

Key Results

  • State-of-the-Art Performance: drGT consistently achieves top-tier or competitive performance across multiple benchmark datasets and evaluation settings (random masking, leave-one-out, zero-shot transfer) for both classification (AUROC) and regression (R²) tasks. For instance, it achieved an AUROC of up to 0.945 and an R² of 0.690 in random cross-validation.
  • Interpretable Drug-Gene Links: The AC-derived drug-gene links show strong agreement with known drug-target interactions and are supported by literature (PubMed abstracts). For example, 36.9% of predicted links matched established DTIs, and 63.7% were supported by literature or structure-based models.
  • Pathway-Level Explanations: Over-representation analysis of AC-prioritized genes reveals drug-perturbed biological processes, providing higher-level, pathway-level explanations for drug responses.
  • Novel Target Identification: The model identified novel drug-gene associations that were not present in existing DTI databases but were supported by literature, suggesting potential new targets for drugs.

Limitations and Future Work

  • "Black Box" Nature of Biological Systems: The paper acknowledges that even with interpretable models, the underlying biological mechanisms are complex and not fully understood. The model's explanations are based on available data and may not capture all nuances.
  • Data Sparsity: Performance can be affected by data sparsity, especially in regression tasks for unseen drugs or cell lines. The model's kernel-based similarity representations help mitigate this, but it remains a challenge.
  • Need for More Data: The study highlights the need for more comprehensive drug-gene relationship data, particularly high-confidence DTI data, to further enhance model performance and interpretability.
  • Future Work: The authors suggest incorporating additional multi-omics data (e.g., methylation, copy number variation, mutation data) to further boost model performance and interpretability. They also aim to broaden the approach for more granular interpretability.

Summary updated

Abstract

Drug development is a lengthy process with a high failure rate. Increasingly, machine learning is utilized to facilitate the drug development processes. These models aim to enhance our understanding of drug characteristics, including their activity in biological contexts. However, a major challenge in drug response (DR) prediction is model interpretability as it aids in the validation of findings. This is important in biomedicine, where models need to be understandable in comparison with established knowledge of drug interactions with proteins. drGAT, a graph deep learning model, leverages a heterogeneous graph composed of relationships between proteins, cell lines, and drugs. drGAT is designed with two objectives: DR prediction as a binary sensitivity prediction and elucidation of drug mechanism from attention coefficients. drGAT has demonstrated superior performance over existing models, achieving 78\% accuracy (and precision), and 76\% F1 score for 269 DNA-damaging compounds of the NCI60 drug response dataset. To assess the model's interpretability, we conducted a review of drug-gene co-occurrences in Pubmed abstracts in comparison to the top 5 genes with the highest attention coefficients for each drug. We also examined whether known relationships were retained in the model by inspecting the neighborhoods of topoisomerase-related drugs. For example, our model retained TOP1 as a highly weighted predictive feature for irinotecan and topotecan, in addition to other genes that could potentially be regulators of the drugs. Our method can be used to accurately predict sensitivity to drugs and may be useful in the identification of biomarkers relating to the treatment of cancer patients.
BibTeX
@misc{inoue2024drgat,
  author = {Yoshitaka Inoue and Hunmin Lee and Tianfan Fu and Augustin Luna},
  title = {{drGAT: Attention-Guided Gene Assessment of Drug Response Utilizing a Drug-Cell-Gene Heterogeneous Network}},
  year = {2024},
  eprint = {2405.08979},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2405.08979},
}
Figure