TL;DR
This paper introduces BDKANN, a novel deep neural network that integrates biological knowledge to predict cancer drug response from gene expression data. By structuring the network based on known biological pathways and protein complexes, BDKANN not only achieves competitive prediction accuracy but also provides interpretable insights into drug mechanisms, potentially uncovering new therapeutic hypotheses.
What Problem Does This Paper Address?
Precision oncology aims to tailor cancer treatments to individual patients based on their unique genomic profiles. A critical step in this process is predicting how a patient will respond to a specific drug. While current computational models are improving, they often suffer from two major drawbacks: they are "black boxes," meaning their predictions are difficult to explain, and they largely ignore existing biological knowledge. This lack of interpretability is a significant hurdle for clinical adoption, as doctors need to understand why a treatment is recommended. Furthermore, many models treat genes in isolation, failing to leverage the complex, hierarchical interactions within biological systems (genes forming protein complexes, which then participate in pathways targeted by drugs). This paper seeks to bridge this gap by developing a model that is both accurate and interpretable, explicitly incorporating biological domain knowledge into its architecture.
Methods
The core innovation of this work is the Biological Domain Knowledge-based Artificial Neural Network (BDKANN). The network's architecture is designed to mirror biological organization:
- Gene Layer: Takes gene expression data as input.
- Protein Complex Layer: Nodes represent protein complexes that genes can form.
- Pathway Layer: Nodes represent biological pathways that protein complexes participate in.
- Drug Layer: Nodes represent drugs that target specific pathways.
Connections between these layers are initially defined by existing biological knowledge, primarily sourced from the Reactome database. This ensures that the model's structure reflects known biological interactions.
To go beyond simply encoding existing knowledge, the authors introduce BDKANN+, which adds "discovery connections" (red lines in Figure 1) between layers that are not explicitly supported by current biological databases. These new connections are then subjected to L1 regularization. This technique encourages the weights of these "discovery" connections to shrink towards zero, effectively penalizing connections that don't contribute significantly to accurate predictions. This allows the model to potentially uncover novel, previously unknown associations between biological entities and drug response, while still prioritizing established biological pathways.
For interpreting the model's predictions, the authors employ SHAP (SHapley Additive exPlanations) values. This game-theoretic approach quantifies the contribution of each node (gene, complex, pathway) to a specific drug response prediction. By aggregating these SHAP values, they can extract a "high-attribution sub-network" for each drug, highlighting the most influential biological components driving the prediction.
Datasets
The study utilizes two publicly available datasets for training and evaluation:
- Genomics of Drug Sensitivity in Cancer (GDSC): Contains gene expression data and drug response (Area Above the Curve - AAC) for over 1000 cancer cell lines screened against 265 drugs.
- Cancer Therapeutic Response Portal (CTRPv2): Similar to GDSC, with over 800 cell lines and 481 drugs.
The authors note that while these datasets are large, there's only a partial overlap in cell lines and drugs, and significant missing data. They focus on 11 overlapping drugs for which sufficient pathway information was available in Reactome. The models are evaluated in various configurations: within-dataset cross-validation (e.g., training and testing on GDSC) and cross-dataset evaluation (e.g., training on GDSC and testing on CTRPv2), to assess generalization capabilities.
Key Insights and Ideas
- Biological Structure Enhances Prediction: Explicitly encoding biological knowledge into a neural network's architecture can improve drug response prediction accuracy, suggesting that biological organization is a crucial factor in drug efficacy.
- Interpretability Through Structure: By mirroring biological hierarchies, BDKANN provides a natural framework for understanding why a prediction is made. The extracted sub-networks directly link predictions to specific genes, complexes, and pathways.
- Discovery Through Regularization: The BDKANN+ model's ability to discover novel connections through L1 regularization is a powerful concept. It allows the model to learn beyond existing knowledge, potentially identifying new drug targets or mechanisms.
- Sub-network Specificity: The interpretation results show that different drugs activate distinct biological sub-networks, reflecting their unique mechanisms of action. However, some shared components are also identified, hinting at common underlying biological processes relevant to multiple drugs.
Key Results
- Improved Prediction Accuracy: BDKANN+ consistently outperforms baseline models (including fully connected DNNs and a knowledge-based DNN without discovery regularization) across most of the tested data configurations, particularly in cross-dataset scenarios. It achieves lower RMSE (Root Mean Squared Error) in three out of four configurations.
- Interpretable Sub-networks: The SHAP-based interpretation successfully extracts biologically relevant sub-networks for individual drugs. These sub-networks highlight key genes, protein complexes, and pathways involved in drug response.
- Identification of Shared and Specific Mechanisms: Analysis of sub-networks for Tamoxifen and Selumetinib reveals both distinct components (reflecting different drug targets) and shared elements (like the RAS pathway), suggesting potential cross-talk or common vulnerabilities in cancer.
- Efficiency of Knowledge-Based Models: While fully connected networks (FCv1, FCv2) sometimes performed well, they struggled more in cross-dataset generalization compared to BDKANN+. The DCELL model, another knowledge-based approach, performed poorly on within-dataset tasks.
Limitations and Future Work
- Dataset Limitations: The study acknowledges the challenges posed by the differences in drug response measurements and missing data between the GDSC and CTRPv2 datasets, which can impact model generalization.
- Limited Drug Coverage: The analysis focused on 11 drugs due to the availability of pathway information in Reactome. Expanding this to more drugs is a clear future direction.
- Completeness of Biological Knowledge: The Reactome database, while comprehensive, is not exhaustive. Future work could incorporate additional or alternative sources of biological knowledge.
- Generalizability of Approach: The authors suggest that the BDKANN framework could be extended to other tasks beyond drug response prediction, such as cancer diagnosis or sub-typing, and could incorporate other omics data types. The inclusion of more drugs and pathway data will further enhance the model's predictive power and discovery potential.
Summary updated