[Paper Review] Affinity Network Fusion and Semi-supervised Learning for Cancer Patient Clustering

Tianle Ma, Aidong Zhang|arXiv (Cornell University)|May 22, 2018

Bioinformatics and Genomic Networks6 references3 citations

TL;DR

This paper proposes Affinity Network Fusion (ANF), a novel method for integrating multi-omic data (gene expression, miRNA, methylation) to improve cancer patient clustering and subtype discovery. ANF constructs kNN-based affinity networks from each omic data type, fuses them into a robust consensus network, and enables both unsupervised clustering and semi-supervised learning with high accuracy using less than 1% of labeled data.

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.

Motivation & Objective

• To address the challenge of integrating heterogeneous and noisy multi-omic data for accurate cancer patient clustering and subtype discovery.
• To develop a more efficient and robust alternative to existing network fusion methods like SNF, reducing computational complexity while maintaining or improving performance.
• To enable zero-shot and few-shot learning for cancer patient classification by leveraging the patient representations learned through ANF.
• To demonstrate the effectiveness of ANF in generating reliable, biologically interpretable patient clusters on a large-scale, harmonized cancer dataset from the GDC portal.
• To explore the utility of ANF as a representation learning framework for transfer learning in semi-supervised settings using deep neural networks.

Proposed method

• Construct patient affinity networks using k-Nearest Neighbor (kNN) and Gaussian kernel transformations on each omic data type to reduce noise and capture nonlinear relationships.
• Fuse multiple individual affinity networks into a single consensus network using a weighted combination strategy that emphasizes complementary information across views.
• Apply spectral clustering on the fused affinity network to obtain patient clusters without requiring labeled data, enabling unsupervised, zero-shot learning.
• Integrate the output of ANF as input features for a feedforward neural network to enable semi-supervised learning with minimal labeled examples.
• Fine-tune the neural network classifier by freezing lower-layer weights (learned representations) and updating only the final layers on additional labeled data, enabling transfer learning.
• Use eigengap analysis to validate the quality of the fused affinity network and ensure it reflects true biological subtypes.

Experimental results

Research questions

• RQ1 Can Affinity Network Fusion (ANF) effectively integrate heterogeneous multi-omic data to produce a robust, noise-resistant patient affinity network for cancer clustering?
• RQ2 How does ANF compare to existing methods like SNF in terms of computational efficiency and clustering accuracy on real-world cancer datasets?
• RQ3 To what extent can ANF-based representations support few-shot learning in cancer patient classification, and how much labeled data is required to achieve high accuracy?
• RQ4 Can the patient representations learned via ANF generalize across different cancer types and be effectively fine-tuned for transfer learning in semi-supervised settings?
• RQ5 Does the fused affinity network generated by ANF align with known biological subtypes and support reliable discovery of novel cancer subtypes?

Key findings

• ANF significantly reduces computational complexity compared to SNF while achieving comparable or better clustering performance on a harmonized cancer dataset of 2193 patients.
• The fused affinity network generated by ANF closely matches true disease subtype labels and demonstrates strong consistency with eigengap analysis, indicating reliable cluster structure.
• In semi-supervised learning, the model achieved over 90% test accuracy using less than 1% of the training data, demonstrating strong few-shot learning capability.
• Fine-tuning the second-to-last layer of the neural network yielded slightly better performance than fine-tuning only the final layer, especially when using three omic views (gene, miRNA, methylation).
• The ANF-based representation was effective even when trained on only two of three kidney cancer subtypes, successfully capturing the underlying three-cluster structure during fine-tuning.
• The method demonstrated robustness to noise, as performance remained strong when using noisy examples derived from unsupervised ANF and spectral clustering for training data augmentation.