Affinity Network Fusion and Semi-supervised Learning for Cancer Patient Clustering

Tianle Ma, Aidong Zhang

TL;DR

This paper introduces Affinity Network Fusion (ANF), a novel method for integrating multi-omic data to cluster cancer patients and discover disease subtypes. ANF achieves high accuracy in unsupervised clustering and, when combined with a neural network, demonstrates remarkable few-shot learning capabilities, achieving over 90% accuracy with less than 1% of labeled data.

What Problem Does This Paper Address?

Personalized medicine relies on accurately stratifying patients into distinct subtypes, especially for complex diseases like cancer. While multi-omic data (like gene expression, methylation, etc.) holds immense potential for this, integrating these diverse and often noisy data types is a significant challenge. Existing methods often perform "shallow" integration by combining pre-computed clustering results, or struggle with the heterogeneity of data. The authors aim to develop a more robust and effective method for integrating multi-omic data to achieve better patient clustering and disease subtype discovery.

Methods

The core of the paper is the Affinity Network Fusion (ANF) framework. ANF takes a different approach than methods that directly cluster patient-feature matrices. Instead, it first constructs an affinity network (or similarity graph) for each individual omic data type. These networks represent the similarity between patients based on a specific data modality.

Key steps in ANF include:

  • Patient kNN Affinity Networks: For each omic view, ANF calculates pairwise distances between patients and then constructs a kNN Gaussian kernel-based affinity matrix. This process is designed to be more robust to noise than raw similarity measures.
  • Fusion of Affinity Networks: ANF then fuses these individual affinity networks into a single, more comprehensive fused affinity matrix. This fusion can be done through weighted averaging or more complex recursive formulas, inspired by methods like Similarity Network Fusion (SNF) but aiming for computational efficiency.
  • Spectral Clustering: The fused affinity matrix is then fed into a spectral clustering algorithm to group patients into potential subtypes. This unsupervised approach allows for discovery without prior labels.
  • Semi-supervised Learning: To leverage limited labeled data, the authors propose a semi-supervised model. This model uses a neural network that takes the affinity matrices generated by ANF as input. This architecture allows for learning patient representations and can achieve high accuracy even with very few labeled examples ("few-shot learning").

Datasets

The primary dataset used is a harmonized cancer dataset from the Genomic Data Commons (GDC) Data Portal. This dataset comprises 2193 cancer patients from four primary sites (adrenal gland, lung, kidney, uterus), with each patient having available gene expression, miRNA expression, and DNA methylation data. The authors note that this dataset is well-harmonized, making it suitable for large-scale integration.

Key Insights and Ideas

  • Multi-omic Integration is Crucial: Combining data from different omic sources significantly improves clustering accuracy compared to using single data types, highlighting the complementary nature of these data.
  • Affinity Networks Capture Richer Relationships: ANF's approach of building patient similarity networks before fusion is more effective than methods that directly integrate raw feature matrices, especially when dealing with heterogeneous data.
  • ANF is Computationally Efficient: The paper emphasizes that ANF achieves comparable or better results than existing methods like SNF with significantly reduced computational requirements, often requiring only one iteration for fusion.
  • Few-Shot Learning is Highly Effective for Cancer Subtyping: The semi-supervised model demonstrates that even with a tiny fraction of labeled data (less than 1%), high classification accuracy can be achieved, thanks to the robust patient representations learned by ANF.
  • "Noisy" Labels Can Still Be Useful: Even labels generated by unsupervised ANF clustering (which might not be perfectly accurate) can be effectively used for training the semi-supervised model, demonstrating the robustness of the learned representations.

Key Results

  • Unsupervised Clustering: ANF achieved high clustering accuracies (measured by NMI and ARI) for four cancer types, often outperforming single-omic approaches. For instance, for uterus cancer, integrating two data types improved NMI from near 0 to 0.30.
  • Semi-supervised Few-Shot Learning: The model achieved over 90% accuracy on test sets with less than 1% of training data. For adrenal gland cancer, training with just two "clean" examples resulted in an NMI > 0.8 (approximately 97% accuracy).
  • Robustness: The method showed robustness across different data combinations and was effective even when trained on limited or "noisy" labels.

Limitations and Future Work

  • Limitations: The paper acknowledges that while integrating three data types is often beneficial, it doesn't always yield better results than two. The performance can also vary depending on the specific omic data types and cancer. The authors also note that for some cancer types, DNA methylation data alone performed better than other single omic types.
  • Future Work: The authors suggest that ANF can be extended to discover subtypes in other cancers and for general complex object clustering with multi-view features. They also highlight the potential for combining supervised and unsupervised learning through effective representation learning. The framework's ability to learn robust representations makes it a promising candidate for transfer learning.

Summary updated

Abstract

Defining subtypes of complex diseases such as cancer and stratifying patient groups with the same disease but different subtypes for targeted treatments is important for personalized and precision medicine. Approaches that incorporate multi-omic data are more advantageous to those using only one data type for patient clustering and disease subtype discovery. However, it is challenging to integrate multi-omic data as they are heterogeneous and noisy. In this paper, we present Affinity Network Fusion (ANF) to integrate multi-omic data for patient clustering. ANF first constructs patient affinity networks for each omic data type, and then calculates a fused network for spectral clustering. We applied ANF to a processed harmonized cancer dataset downloaded from GDC data portal consisting of 2193 patients, and generated promising results on clustering patients into correct disease types. Moreover, we developed a semi-supervised model combining ANF and neural network for few-shot learning. In several cases, the model can achieve greater than 90% acccuracy on test set with training less than 1% of the data. This demonstrates the power of ANF in learning a good representation of patients, and shows the great potential of semi-supervised learning in cancer patient clustering.
BibTeX
@inproceedings{ma2018affinity,
  author = {Tianle Ma and Aidong Zhang},
  title = {{Affinity Network Fusion and Semi-supervised Learning for Cancer Patient Clustering}},
  booktitle = {Methods},
  year = {2018},
  doi = {10.1016/j.ymeth.2018.05.020},
}
Figure