TL;DR
This paper introduces a novel "Multi-type Data Joint Learning" (MDJL) approach for cancer survival prediction. MDJL effectively integrates diverse genomic data types by learning both sample similarity and correlation representations, leading to significantly improved prediction accuracy compared to existing methods.
What Problem Does This Paper Address?
Accurate prediction of cancer patient survival is crucial for effective treatment planning. While integrating multiple types of genomic data (like DNA methylation, mRNA, and miRNA expression) has shown promise, existing methods often fail to simultaneously leverage both the feature information (e.g., gene expression levels) and the structural information (e.g., how samples relate to each other within and across data types). This limitation hinders the ability to capture a comprehensive understanding of cancer, especially with the increasing availability of high-throughput genomic data. The paper aims to bridge this gap by proposing a method that jointly learns these representations.
Methods
The core of the MDJL approach lies in its four main components:
- Correlation Representations Extraction: Instead of just concatenating data or looking for common features across all data types, MDJL learns "correlation representations" between every pair of data types. This is achieved by first learning feature representations for each data type using neural networks and then computing an "interactive map" between pairs of these representations. These maps are then projected into an embedded space using another set of neural networks to capture deep correlation representations. This aims to exploit diverse and complementary information.
- Discriminant Representations Generation: The correlation representations from all pairwise data types are then fused (concatenated) to form a single, comprehensive discriminant feature representation.
- Similarity Matrices Construction: This component focuses on capturing the "structure information." It constructs similarity matrices not from the raw data, but from the learned correlation representations. This process aims to preserve both global structures (similarities across different data types) and local structures (neighborhood information within data types). The method iteratively updates these matrices, incorporating K-nearest neighbor information to enhance robustness against noise.
- Graph Convolutional Network (GCN) with Cox Loss: Finally, the learned discriminant representations and the fused similarity matrix are fed into a three-layer GCN. The GCN learns to predict survival risk by minimizing a Cox proportional hazards loss function, which is standard for survival analysis and can handle censored data.
Datasets
The study evaluates MDJL on four public cancer datasets:
- Glioblastoma Multiforme (GBM)
- Kidney Renal Clear Cell Carcinoma (KRCCC)
- Lung Squamous Cell Carcinoma (LSCC)
- Breast Invasive Carcinoma (BIC)
For each cancer type, three types of genomic data were used: DNA methylation, mRNA expression, and miRNA expression. These datasets were obtained from http://compbio.cs.toronto.edu/SNF/ and were preprocessed for missing data and normalization.
Key Insights and Ideas
- Pairwise Correlation is Key: The paper argues that learning correlations between pairs of data types is more effective than simply concatenating all data or finding common features across all types. This allows for capturing richer, complementary information.
- Structure from Correlation: Constructing similarity matrices from learned correlation representations, rather than raw data, makes them more robust to noise and outliers, leading to better structural understanding.
- Joint Learning of Feature and Structure: The simultaneous learning of discriminant features (via correlation) and sample structure (via similarity matrices) is a central theme, aiming for a more holistic representation.
- GCN for Survival: The application of a GCN, leveraging the learned similarity structure, provides a powerful framework for integrating diverse information for survival prediction.
Key Results
- Superior Performance: MDJL consistently outperformed several state-of-the-art methods (MKL+Cox loss, MDNNMD-Cox, DLMR, CrossAE, VAECox, DeepSurv) across all four cancer datasets in terms of the average concordance index (C-index).
- Significant Improvement: MDJL improved the average prediction performance by 4.40% to 7.2% over the second-best method on the tested datasets.
- Component Effectiveness: Further investigations confirmed that both the correlation representation extraction and the structure learning components contribute significantly to MDJL's performance.
- Robustness: The approach showed robustness to hyperparameter choices (K and T) within a reasonable range.
Limitations and Future Work
- Computational Cost: While deemed acceptable, the paper notes the computing time for MDJL and other methods. Future work could focus on optimizing computational efficiency.
- Hyperparameter Sensitivity: Although generally insensitive, the paper investigates the impact of hyperparameters K and T. Further exploration into automated hyperparameter tuning could be beneficial.
- Generalizability: While tested on four cancer types, extending the approach to a wider range of cancers and data modalities could be a future direction.
- Interpretability: While the model achieves high accuracy, further work could focus on improving the interpretability of the learned representations and predictions.
Summary updated