TL;DR
This paper introduces DeepKEGG, a novel deep learning framework that integrates multiple types of cancer data (genomics, transcriptomics, and proteomics) to predict cancer recurrence and discover biomarkers. By incorporating biological pathway information and a self-attention mechanism, DeepKEGG achieves superior prediction accuracy and provides interpretable insights into the underlying biological mechanisms.
What Problem Does This Paper Address?
Cancer recurrence is a major challenge in cancer treatment, and predicting it accurately is crucial for patient management. Multi-omics data holds immense potential for understanding cancer development and identifying recurrence predictors. However, existing methods often struggle with integrating diverse omics data effectively, lack interpretability (making it hard to understand why a prediction is made), and overlook potential correlations between different patient samples. This paper aims to bridge this gap by developing a deep learning framework that not only integrates multi-omics data for accurate prediction but also provides biological insights and addresses the issue of sample-to-sample correlations.
Methods
DeepKEGG is built upon several key components:
- Biological Hierarchical Module: This module leverages known biological relationships between genes/miRNAs and KEGG pathways. Instead of a fully connected network, it uses this biological knowledge to create local connections between neurons, which helps to alleviate the "curse of dimensionality" (a common problem when dealing with high-dimensional data) and improves model interpretability.
- Pathway Self-Attention Module: To capture correlations between different patient samples, this module uses a self-attention mechanism. It learns how different samples relate to each other within the pathway feature space, enhancing the model's ability to generalize and predict.
- Attribution-Based Feature Importance: To identify biomarkers, the paper employs the DeepLIFT method. This technique quantifies the contribution of each gene or miRNA to the model's prediction, allowing researchers to pinpoint important features and understand their role in cancer recurrence.
- Classification Module: A standard Multi-Layer Perceptron (MLP) is used for the final prediction of cancer recurrence after integrating features from the previous modules.
Datasets
The study utilized publicly available cancer datasets from two major sources:
- The Cancer Genome Atlas (TCGA): Datasets from four cancer types were used: Breast Cancer (BRCA), Liver Cancer (LIHC), Bladder Cancer (BLCA), and Prostate Cancer (PRAD).
- Therapeutically Applicable Research to Generate Effective Treatments (TARGET): Datasets from Acute Myeloid Leukemia (TARGET-AML) and Wilms Tumor (TARGET-WT) were included.
These datasets provided matched multi-omics data (SNV, mRNA, and miRNA) for training and evaluation.
Key Insights and Ideas
- Biological Knowledge is Key: Integrating known biological pathways significantly improves the model's ability to learn from multi-omics data and provides a foundation for interpretability.
- Sample Correlations Matter: Ignoring how samples relate to each other can limit predictive power. The self-attention mechanism effectively captures these inter-sample relationships.
- Interpretability Enhances Trust: By providing feature importance scores and pathway contributions, DeepKEGG moves beyond a "black box" approach, offering insights that can guide further biological research and clinical decisions.
- Multi-omics is Better Than Single-Omics: The study explicitly shows that combining different omics data types leads to better prediction performance than using any single type alone, highlighting the complementary nature of these data sources.
Key Results
- Superior Prediction Accuracy: DeepKEGG consistently outperformed seven other state-of-the-art classification methods (including SGD, LR, DT, SVM, KNN, PathCNN, and MOGONET) across all tested TCGA and TARGET datasets. For example, on the LIHC dataset, DeepKEGG achieved an average AUC of 0.947, significantly higher than the next best method (MOGONET at 0.879).
- Effective Biomarker Discovery: Case studies demonstrated DeepKEGG's ability to identify potential genes and miRNAs associated with cancer recurrence. For instance, it highlighted genes like CYP2E1 and TRAF6 for LIHC, and EGF and ERBB2 for BLCA, with findings supported by existing literature.
- Biological Pathway Insights: The model successfully identified and ranked important biological pathways involved in cancer recurrence, providing a deeper understanding of the underlying mechanisms.
Limitations and Future Work
- Hyperparameter Tuning: While the paper explored the impact of the
kparameter in the self-attention module, further optimization of other hyperparameters might yield even better results. - Integration with Imaging Data: The authors suggest that future work could explore integrating pathological image data with omics data to provide a more comprehensive view for clinical diagnosis and treatment.
- Experimental Validation: While the identified biomarkers are supported by literature, further experimental validation in biological settings would strengthen their significance.
Summary updated