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

1. Problem Statement

The paper addresses two critical challenges in computational drug discovery:

1. Predictive Generalization: Existing Deep Learning (DL) models for Drug Response Prediction (DRP) often struggle to generalize to unseen drugs or cell lines (zero-shot or leave-one-out scenarios) and frequently rely on "black box" architectures that lack transparency.
2. Interpretability: While DL models achieve high accuracy, their internal decision-making processes are often opaque. In biomedical contexts, high performance alone is insufficient; models must provide biologically plausible explanations (e.g., identifying specific gene-drug mechanisms) to be useful for hypothesis generation and experimental validation.

Current state-of-the-art (SOTA) models often use homogeneous graphs (drug-cell only) or lack mechanism-oriented interpretability, failing to capture the complex, non-linear dependencies between drugs, cell lines, and the underlying gene regulatory networks.

2. Methodology: The drGT Framework

The authors propose drGT, an interpretable deep learning framework based on Graph Transformers (GTs) operating over a heterogeneous network.

A. Data Construction & Heterogeneous Graph

The model integrates three distinct data modalities into a unified graph structure:

• Nodes: Drugs, Cell Lines, and Genes.
• Edges (Adjacency Matrices):
  • Drug-Cell ($A_{dc}$): Represents drug response (IC50 values for regression, binary labels for classification).
  • Cell-Gene ($A_{cg}$): Derived from gene expression data (standardized and thresholded to retain positive correlations).
  • Drug-Gene ($A_{dg}$): Derived from Drug-Target Interaction (DTI) databases (e.g., DrugBank). Unknown interactions are zero-padded or assigned soft values for interpretability.
• Node Features:
  • Drugs: Morgan fingerprints (2048-bit) derived from SMILES strings.
  • Cell Lines & Genes: Gene expression profiles.
  • Transformation: Features are transformed using Radial Basis Function (RBF) kernels to compute pairwise similarity matrices, which are then projected into a unified latent space via linear layers.

B. Model Architecture: Graph Transformer

drGT utilizes a Graph Transformer (GT) layer (based on the TransformerConv operator) to process the heterogeneous graph:

• Global Attention: Unlike standard Graph Neural Networks (GNNs) that aggregate local neighbors, GTs apply global attention mechanisms across the graph.
• Edge-Aware Attention: The attention mechanism incorporates edge features (e.g., specific drug-cell response values or interaction types) into the Key and Value projections.
• Attention Coefficients (ACs): The core innovation. The model learns Attention Coefficients for every edge. These coefficients quantify the importance of specific gene nodes in predicting the response of a specific drug-cell pair.
• Prediction Head: The final node embeddings are concatenated (Drug + Cell) and passed through Fully Connected (FC) layers with a Sigmoid activation for classification (Binary Cross-Entropy loss) or linear regression for IC50 prediction (Mean Squared Error loss).

C. Interpretability Strategy

The model leverages the learned Attention Coefficients (ACs) for mechanism-oriented interpretation:

1. Gene Ranking: For a given drug, genes are ranked by their ACs.
2. Literature Validation: Top-ranked genes are cross-referenced with PubMed co-occurrence data (using NCBI ESearch) to verify biological plausibility.
3. External Model Validation: Predictions are compared against DeepPurpose, a SOTA DTI predictor, to assess computational consistency.
4. Pathway Enrichment: Over-Representation Analysis (ORA) is performed on top-ranked genes to identify enriched biological pathways (e.g., apoptosis, DNA repair) associated with drug sensitivity.

3. Key Contributions

1. Unified Heterogeneous Graph: drGT is the first to integrate drugs, cell lines, and genes into a single heterogeneous graph using a Graph Transformer, enabling joint reasoning across all three entity types.
2. Mechanism-Oriented Interpretability: By utilizing Attention Coefficients, drGT provides fine-grained, edge-level interpretability, directly attributing drug response to specific gene contributions, unlike previous models that offer only coarse pathway-level or feature-level explanations.
3. Superior Generalization: The model demonstrates robust performance in challenging settings where training data for specific drugs or cell lines is missing (Leave-One-Out) or entirely absent (Zero-Shot).
4. Biological Discovery: The model successfully recovers known Drug-Target Interactions (DTIs) and identifies novel, literature-supported drug-gene associations that are not present in current databases.

4. Experimental Results

The model was evaluated on four major pharmacogenomic datasets: NCI60, GDSC1, GDSC2, and CTRP.

A. Predictive Performance

• Random 5-Fold Cross-Validation: drGT achieved SOTA regression performance ($R^2$ up to 0.690 on NCI60), outperforming all baselines (DeepDSC, GraphCDR, MOFGCN, etc.). It achieved competitive classification accuracy (AUROC up to 0.945).
• Leave-One-Out (Unseen Drugs/Cells):
  • Unseen Drugs: drGT was the only model to yield a positive $R^2$ (0.022) for unseen drugs, whereas all baselines produced negative $R^2$ values. It achieved an AUROC of 0.844.
  • Unseen Cells: Achieved an AUROC of 0.706 and $R^2$ of 0.692.
• Zero-Shot Transfer: Trained on GDSC1 and tested on GDSC2 (different assay protocols), drGT achieved the highest overall AUROC (0.786) and regression $R^2$ (0.334), demonstrating superior robustness to distribution shifts compared to NIHGCN and MOFGCN.

B. Interpretability & Biological Plausibility

• DTI Recovery: Among 976 drugs with known DTIs, 36.9% of the top-5 predicted gene links matched established DTIs.
• Literature Support: 63.7% of the predicted drug-gene links were supported by either PubMed abstracts or the DeepPurpose model.
• Novel Discoveries: The model identified 182 novel drug-gene pairs (not in known DTI databases) that were supported by PubMed literature, suggesting potential new targets for understudied compounds.
• Pathway Analysis: Enrichment analysis of top-ranked genes successfully mapped drugs to known mechanisms of action (e.g., Kinase inhibitors linked to PI3K/AKT signaling; DNA damaging agents linked to p53 and apoptosis pathways).

5. Significance

The drGT model represents a significant advancement in translational bioinformatics by bridging the gap between high-accuracy prediction and biological interpretability.

• Scientific Impact: It moves beyond "black box" predictions to provide actionable hypotheses about why a drug works, identifying specific gene targets and pathways.
• Clinical Utility: Its ability to generalize to unseen drugs and cell lines makes it a powerful tool for repurposing existing drugs or predicting responses for rare cancer types where data is scarce.
• Methodological Impact: It demonstrates that Graph Transformers applied to heterogeneous biological networks can effectively capture complex, multi-modal dependencies, setting a new standard for interpretable drug response prediction.

The code and processed datasets are publicly available, facilitating further research and validation in the community.