Table of Contents
Fetching ...

Interpretable Perturbation Modeling Through Biomedical Knowledge Graphs

Pascal Passigan, Kevin Zhu, Angelina Ning

TL;DR

The paper tackles predicting drug-induced transcriptional perturbations by moving beyond binary drug–disease links to a mechanistic, transcriptome-level view. It fuses PrimeKG++ with LINCS L1000 data into a seven-type heterogeneous graph and leverages a GATv2-based message-passing framework to predict a perturbation delta ${\Delta}$ that augments baseline cell expression ${\mathbf{b}_c}$, yielding ${\hat{\mathbf{y}} = \mathbf{b}_c + {\Delta}}$. The authors demonstrate that graph-based context improves generalization to novel chemical scaffolds, with ablation studies confirming the essential roles of graph topology and pretrained node features, and they provide interpretability through attention analyses and a drug-specific reasoning case study. The work highlights a pathway toward mechanistic drug modeling and practical applications in drug repurposing, while recognizing current knowledge gaps and suggesting directions for richer graphs and architectures to bolster perturbation prediction.

Abstract

Understanding how small molecules perturb gene expression is essential for uncovering drug mechanisms, predicting off-target effects, and identifying repurposing opportunities. While prior deep learning frameworks have integrated multimodal embeddings into biomedical knowledge graphs (BKGs) and further improved these representations through graph neural network message-passing paradigms, these models have been applied to tasks such as link prediction and binary drug-disease association, rather than the task of gene perturbation, which may unveil more about mechanistic transcriptomic effects. To address this gap, we construct a merged biomedical graph that integrates (i) PrimeKG++, an augmentation of PrimeKG containing semantically rich embeddings for nodes with (ii) LINCS L1000 drug and cell line nodes, initialized with multimodal embeddings from foundation models such as MolFormerXL and BioBERT. Using this heterogeneous graph, we train a graph attention network (GAT) with a downstream prediction head that learns the delta expression profile of over 978 landmark genes for a given drug-cell pair. Our results show that our framework outperforms MLP baselines for differentially expressed genes (DEG) -- which predict the delta expression given a concatenated embedding of drug features, target features, and baseline cell expression -- under the scaffold and random splits. Ablation experiments with edge shuffling and node feature randomization further demonstrate that the edges provided by biomedical KGs enhance perturbation-level prediction. More broadly, our framework provides a path toward mechanistic drug modeling: moving beyond binary drug-disease association tasks to granular transcriptional effects of therapeutic intervention.

Interpretable Perturbation Modeling Through Biomedical Knowledge Graphs

TL;DR

The paper tackles predicting drug-induced transcriptional perturbations by moving beyond binary drug–disease links to a mechanistic, transcriptome-level view. It fuses PrimeKG++ with LINCS L1000 data into a seven-type heterogeneous graph and leverages a GATv2-based message-passing framework to predict a perturbation delta that augments baseline cell expression , yielding . The authors demonstrate that graph-based context improves generalization to novel chemical scaffolds, with ablation studies confirming the essential roles of graph topology and pretrained node features, and they provide interpretability through attention analyses and a drug-specific reasoning case study. The work highlights a pathway toward mechanistic drug modeling and practical applications in drug repurposing, while recognizing current knowledge gaps and suggesting directions for richer graphs and architectures to bolster perturbation prediction.

Abstract

Understanding how small molecules perturb gene expression is essential for uncovering drug mechanisms, predicting off-target effects, and identifying repurposing opportunities. While prior deep learning frameworks have integrated multimodal embeddings into biomedical knowledge graphs (BKGs) and further improved these representations through graph neural network message-passing paradigms, these models have been applied to tasks such as link prediction and binary drug-disease association, rather than the task of gene perturbation, which may unveil more about mechanistic transcriptomic effects. To address this gap, we construct a merged biomedical graph that integrates (i) PrimeKG++, an augmentation of PrimeKG containing semantically rich embeddings for nodes with (ii) LINCS L1000 drug and cell line nodes, initialized with multimodal embeddings from foundation models such as MolFormerXL and BioBERT. Using this heterogeneous graph, we train a graph attention network (GAT) with a downstream prediction head that learns the delta expression profile of over 978 landmark genes for a given drug-cell pair. Our results show that our framework outperforms MLP baselines for differentially expressed genes (DEG) -- which predict the delta expression given a concatenated embedding of drug features, target features, and baseline cell expression -- under the scaffold and random splits. Ablation experiments with edge shuffling and node feature randomization further demonstrate that the edges provided by biomedical KGs enhance perturbation-level prediction. More broadly, our framework provides a path toward mechanistic drug modeling: moving beyond binary drug-disease association tasks to granular transcriptional effects of therapeutic intervention.
Paper Structure (18 sections, 4 equations, 7 figures, 1 table)

This paper contains 18 sections, 4 equations, 7 figures, 1 table.

Figures (7)

  • Figure 1: Model performance across split types and graph ablations. Under scaffold split (leftmost cluster), GAT (red) achieves DEG correlation of 0.708, significantly outperforming MLP baselines ($\sim$0.68). Edge shuffling and node randomization (middle clusters) cause GAT performance to collapse below baseline levels, demonstrating the critical role of both graph topology and pretrained features. Under random split (rightmost), all models converge to similar performance ($\sim$0.69), indicating that graph structure provides advantages only when generalizing to novel chemical scaffolds.
  • Figure 2: Baseline attention distribution (scaffold split, no ablation). In the full-information setting, 98.8% of attention routes through protein nodes, mirroring the biological reality that drugs act primarily via protein targets. This protein-centric pattern emerges without explicit supervision, demonstrating that the model learns mechanistically coherent reasoning from graph structure alone.
  • Figure 3: Attention distribution under edge shuffling. Destroying biological relationships while preserving connectivity causes attention to fragment, with 23.2% now flowing directly between drug nodes rather than through mechanistic intermediaries. This structural incoherence correlates with the performance collapse to 0.537 DEG correlation.
  • Figure 4: Attention distribution under node feature randomization. Topology preserves the protein-centric attention flow, but without valid semantic embeddings, performance falls to 0.572 DEG correlation. This demonstrates that both graph structure and meaningful node features are required for effective reasoning.
  • Figure 5: Attention distribution under random dataset split. The model relies more heavily on drug-to-drug similarity when test compounds are chemically similar to training data. This explains why graph structure provides no advantage when chemical memorization suffices---the model bypasses mechanistic reasoning in favor of similarity matching.
  • ...and 2 more figures