netDx: interpretable patient classification using integrated patient similarity networks

Shraddha Pai, Shirley Hui, Ruth Isserlin, Muhammad A Shah, Hussam Kaka, Gary D. BaderMolecular 2019 · DOI

TL;DR

This paper introduces netDx, a novel machine learning framework for patient classification that leverages "patient similarity networks" to integrate diverse data types. netDx not only achieves superior or comparable performance to existing methods in cancer survival prediction but also offers enhanced interpretability by visualizing patient relationships and identifying key biological pathways driving predictions.

What Problem Does This Paper Address?

The field of precision medicine aims to tailor treatments and diagnoses to individual patients. This requires accurate, generalizable, and interpretable prediction algorithms that can handle the complexity and sparsity of multi-modal patient data (genomics, clinical records, etc.). Traditional machine learning models often struggle with integrating diverse data types, can be "black boxes" making it hard to understand why a prediction is made, and may require extensive data preprocessing like imputation. Furthermore, for genomic data, interpretability is crucial for driving hypothesis-driven research into new treatments.

Methods

netDx tackles these challenges by building upon the concept of Patient Similarity Networks (PSNs). Here's how it works:

  1. PSN Construction: For each data type (e.g., gene expression, clinical measurements), netDx constructs a network where each patient is a node, and edges represent the similarity between pairs of patients for that specific data type. This allows for the integration of diverse data types into a common format.
  2. Feature Selection: netDx uses a modified version of the GeneMANIA algorithm to identify which PSNs (i.e., which data types or combinations thereof) are most predictive of a given patient class (e.g., high vs. low survival risk). This is done through a cross-validation process to ensure generalizability.
  3. Pathway-Level Features: A key innovation is the ability to define PSNs based on biological pathways rather than individual genes. This reduces the number of features, improves interpretability, and can enhance predictive power by capturing coordinated gene activity.
  4. Integrated Network and Classification: The top-scoring PSNs are integrated into a single "integrated patient similarity network." New patients are then classified based on their similarity to known patients within this integrated network.

Datasets

The primary dataset used for benchmarking netDx is the PanCancer Survival Prediction benchmark dataset. This dataset includes data from four tumor types:

  • Renal clear cell carcinoma (KIRC)
  • Ovarian serous cystadenocarcinoma (OV)
  • Glioblastoma multiforme (GBM)
  • Lung squamous cell carcinoma (LUSC)

For each tumor type, the data includes clinical variables, mRNA, miRNA, protein expression, DNA methylation, and somatic copy number aberrations. The paper also uses a breast cancer dataset for Luminal A subtype classification and an asthma dataset for case/control prediction.

Key Insights and Ideas

  • Patient Similarity as a Unifying Framework: The core idea of representing patient relationships as networks provides a powerful and intuitive way to integrate heterogeneous data. It mirrors how clinicians often reason by comparing new patients to those they've seen before.
  • Interpretability Through Networks: Instead of a black-box model, netDx provides a visualizable network of patient similarities. This allows researchers to see why a patient is classified a certain way, by observing their connections to other patients and identifying the data types that contribute most to that similarity.
  • Pathways as Predictive Features: Using biological pathways as features significantly enhances interpretability and can improve classification performance. This suggests that understanding coordinated biological processes is more informative for prediction than looking at individual genes in isolation.
  • Bridging the Gap from Cohort to Individual: By identifying predictive pathways and visualizing their correlation with outcomes, netDx helps translate population-level findings into actionable insights for individual patient care.

Key Results

  • Superior Classification Performance: netDx significantly outperforms most other machine-learning approaches on the PanCancer survival benchmark across most cancer types, achieving comparable or better mean AUROC scores. For KIRC, OV, and GBM, netDx's performance was statistically significantly higher than other methods.
  • Effective Pathway Identification: In breast cancer and asthma studies, netDx successfully identified biologically relevant pathways that were predictive of disease subtypes or status. For instance, in breast cancer, it highlighted pathways related to cell cycle regulation and DNA repair, consistent with known disease mechanisms.
  • Improved Performance with Pathway Features: Using pathway-level gene expression data for classification in breast cancer resulted in slightly, but significantly, better performance (AUROC of 0.97) compared to using raw gene expression as a single feature (AUROC of 0.96).
  • Software Availability: A free, open-source software implementation of netDx with automation workflows is provided, facilitating its adoption by the research community.

Limitations and Future Work

  • Need for Large Datasets: Like most supervised learning methods, netDx requires a substantial number of patient samples for training. The authors suggest that alternative methods might be needed for queries involving smaller patient cohorts.
  • Data Sharing: netDx requires computing similarities for both training and test patients, which may necessitate sharing private data. The authors propose secure sharing, differential privacy, or privacy-preserving data mining techniques as potential solutions.
  • Non-linear Effects: While netDx performs well, the authors acknowledge that a single support vector machine model in one of their benchmarks outperformed netDx, possibly due to identifying a non-linear decision boundary. Future work could explore incorporating non-linear similarity measures.
  • Integration with Other Tools: The authors suggest that netDx could be used in tandem with tools like DIABLO, which focuses on gene-level features, to provide complementary views of predictive multi-omic features.

Summary updated

Abstract

Patient classification has widespread biomedical and clinical applications, including diagnosis, prognosis, and treatment response prediction. A clinically useful prediction algorithm should be accurate, generalizable, be able to integrate diverse data types, and handle sparse data. A clinical predictor based on genomic data needs to be interpretable to drive hypothesis-driven research into new treatments. We describe netDx, a novel supervised patient classification framework based on patient similarity networks, which meets these criteria. In a cancer survival benchmark dataset integrating up to six data types in four cancer types, netDx significantly outperforms most other machine-learning approaches across most cancer types. Compared to traditional machine-learning-based patient classifiers, netDx results are more interpretable, visualizing the decision boundary in the context of patient similarity space. When patient similarity is defined by pathway-level gene expression, netDx identifies biological pathways important for outcome prediction, as demonstrated in breast cancer and asthma. netDx can serve as a patient classifier and as a tool for discovery of biological features characteristic of disease. We provide a free software implementation of netDx with automation workflows.
BibTeX
@inproceedings{pai2019netdx,
  author = {Shraddha Pai and Shirley Hui and Ruth Isserlin and Muhammad A Shah and Hussam Kaka and Gary D. Bader},
  title = {{netDx: interpretable patient classification using integrated patient similarity networks}},
  booktitle = {Molecular},
  year = {2019},
  doi = {10.15252/msb.20188497},
}
Figure