IBPGNET: lung adenocarcinoma recurrence prediction based on neural network interpretability

Zhanyu Xu, Haibo Liao, Liuliu Huang, Qingfeng Chen, Wei Lan, Shikang LiBriefings in Bioinformatics 2024 · DOI · OpenReview

TL;DR

This paper introduces IBPGNET, a novel interpretable neural network that leverages biological pathway hierarchies to predict lung adenocarcinoma (LUAD) recurrence. By integrating multi-omics data and providing interpretable insights into molecular mechanisms, IBPGNET achieves superior prediction performance and identifies key genes like PSMC1 and PSMD11 as potential therapeutic targets.

What Problem Does This Paper Address?

Lung adenocarcinoma (LUAD) is a significant subtype of lung cancer with a high probability of recurrence even in early stages. Predicting this recurrence and understanding the underlying molecular mechanisms are crucial for improving patient outcomes and developing targeted therapies. While deep learning models have shown promise in cancer prediction, their "black box" nature hinders biological interpretation, making it difficult to translate findings into clinical applications. Existing methods often struggle to effectively integrate diverse multi-omics data and provide interpretable insights into complex biological pathways.

Methods

The core of this work is the Interpretable Biological Pathway Graph Neural Networks (IBPGNET) framework. IBPGNET is built upon the Reactome pathway hierarchy, compressing this complex biological information into a structured neural network. Key methodological components include:

  • Pathway Hierarchy Compression: The Reactome pathway hierarchy is used to construct a multi-layer network where nodes represent biological entities and edges represent their relationships. This structure is designed to mimic the flow of information from lower-level biological entities to higher-level ones.
  • Graph Neural Network (GCN) for Latent Relationships: A GCN is employed to predict latent pathway relationships, addressing the sparsity in existing pathway databases and enhancing the network's ability to capture complex interactions. This step is crucial for improving prediction performance.
  • Multi-omics Data Integration: IBPGNET integrates various omics data types (SNV, CNV - amplification and deletion) to provide a comprehensive view of the biological landscape.
  • Interpretability through DeepLIFT: The DeepLIFT algorithm is used to assign importance scores to input features and biological entities, allowing for a visual and hierarchical understanding of the factors driving recurrence prediction. This addresses the "black box" problem of deep learning.
  • Data Preprocessing: Standard techniques like noise removal, feature selection (using chi-square test), and handling missing values are applied to prepare the multi-omics data.

Datasets

The study utilized data from The Cancer Genome Atlas (TCGA) database for lung adenocarcinoma (LUAD). Specifically, the following data types were used:

  • Copy Number Variant (CNV) data: GISTIC-focal score by gene type.
  • Somatic Mutation (SNV) data: VarScan2 Variant Aggregation and Masking type.
  • Clinical data: For LUAD patients.

The dataset was filtered to include only samples with all three data types. CNV data was further divided into amplification (AMP_CNV) and deletion (DEL_CNV) subsets. The paper mentions using a chi-square test to select the top 3000 significant variables as pre-training features.

Key Insights and Ideas

  • Pathway Hierarchy as a Predictive Framework: The paper demonstrates that explicitly incorporating biological pathway hierarchies into a neural network can significantly improve the prediction of complex biological outcomes like cancer recurrence.
  • GCNs for Uncovering Latent Biological Relationships: Using GCNs to predict latent pathway relationships is a powerful way to enrich sparse biological knowledge bases and improve model performance.
  • Interpretability is Key for Clinical Translation: By employing DeepLIFT, the study not only predicts recurrence but also provides actionable biological insights, making the model more trustworthy and useful for researchers and clinicians.
  • PSMC1 and PSMD11 as Novel LUAD Biomarkers: The identification and experimental validation of PSMC1 and PSMD11 as genes significantly associated with LUAD recurrence and drug sensitivity is a major finding.
  • EGFR as a Mediator of Therapeutic Sensitivity: The research suggests that PSMC1 and PSMD11 might influence therapeutic sensitivity (e.g., to afatinib) by modulating EGFR expression, offering a potential mechanism for overcoming drug resistance.

Key Results

  • Superior Prediction Performance: IBPGNET achieved an Area Under the Precision-Recall Curve (AUPR) of 0.790 in 5-fold cross-validation, outperforming other classification methods (e.g., SGD, RF, LR, SVMs, DeepOmix, P-NET, PathCNN). It also reported an AUC of 0.88, accuracy of 0.82, and F1 score of 0.68.
  • Multi-omics Integration Benefits: Integrating SNV, AMP_CNV, and DEL_CNV data yielded the best prediction performance for IBPGNET (AUPR of 0.79), highlighting the value of comprehensive omics data. SNV data alone performed better than CNV data for recurrence prediction.
  • Latent Pathway Relationships Improve Performance: IBPGNET with latent pathway relationships predicted by GCN performed better than the model without them, underscoring the importance of learning these hidden connections.
  • Identification of PSMC1 and PSMD11: These genes were identified as significantly associated with LUAD recurrence. Their expression was higher in LUAD cells than normal cells.
  • Functional Validation of PSMC1 and PSMD11: Knockdown of PSMC1 and PSMD11 in LUAD cells led to reduced migration, invasion, and proliferation. It also increased sensitivity to afatinib and decreased EGFR expression.

Limitations and Future Work

  • Data Limitations: The study acknowledges the limitations of sequencing technology leading to outliers and missing data in multi-omics datasets. While preprocessing steps were taken, these inherent data challenges can still impact model performance.
  • Generalizability: While the study focuses on LUAD, the generalizability of IBPGNET to other cancer types or diseases would require further investigation.
  • Experimental Validation Scope: The experimental validation focused on PSMC1 and PSMD11. Future work could explore other identified biomarkers and pathways in more detail.
  • Therapeutic Target Exploration: The identified therapeutic potential of inhibiting PSMC1 and PSMD11 warrants further preclinical and clinical investigation to develop effective treatments for LUAD patients, particularly those with drug resistance.
  • Broader Application of IBPGNET: The authors suggest that IBPGNET's framework, which leverages pathway hierarchies and interpretability, could be applied to other complex biological prediction tasks beyond LUAD recurrence.

Summary updated

BibTeX
@article{xu2024ibpgnet,
  author = {Zhanyu Xu and Haibo Liao and Liuliu Huang and Qingfeng Chen and Wei Lan and Shikang Li},
  title = {{IBPGNET: lung adenocarcinoma recurrence prediction based on neural network interpretability}},
  journal = {Briefings in Bioinformatics},
  year = {2024},
  doi = {10.1093/bib/bbae080},
}
Figure