Effectiveness of machine learning at modeling the relationship between Hi-C data and copy number variation

Yuyang Wang, Yu Sun, Zeyu Liu, Bijia Chen, Hebing Chen, Chao Ren, Xuanwei Lin, Pengzhen Hu, Peiheng Jia, Xiang Xu, Kang Xu, Ximeng Liu, Hao Li, Xiaochen BoQuant. Biol 2024 · DOI

TL;DR

This paper introduces a Graph Convolutional Network (GCN) model that effectively predicts Copy Number Variations (CNVs) from Hi-C data, outperforming traditional linear models. By treating Hi-C interaction maps as graphs, the GCN captures complex spatial relationships, offering a more robust approach to understanding the link between 3D genome structure and genetic variations.

What Problem Does This Paper Address?

Copy Number Variations (CNVs) are significant genetic alterations that can lead to diseases. While traditionally studied using Whole Genome Sequencing (WGS), Hi-C data, which captures the 3D organization of the genome, holds potential for inferring CNVs. However, Hi-C data is notoriously sparse and noisy, making it challenging to directly extract CNV information. Existing methods often struggle with the complexity and scale of this data. This paper aims to bridge this gap by developing and evaluating machine learning models, specifically a GCN, to effectively leverage Hi-C data for accurate CNV prediction and to understand the underlying relationship between 3D chromatin structure and CNVs.

Methods

The researchers explored two main machine learning approaches:

  • Linear Transformation Model: This served as a baseline, attempting to find a direct linear relationship between 1D Hi-C read counts and CNVs. They found that while a linear relationship exists for individual chromosomes, it's not generalizable across the entire genome.
  • Graph Convolutional Network (GCN): This is the core of their proposed solution.
  • Graph Representation: Hi-C interaction matrices were converted into graphs, where genomic bins are nodes and interactions are weighted edges.
  • Node Embeddings: A node2vec model was used to learn feature representations (embeddings) for each node, capturing local and global graph structures.
  • GCN Architecture: A four-layer GCN with residual connections was employed to process these graph structures and node features, predicting CNV labels for each bin.
  • Perturbation Experiments: To assess the model's robustness and the importance of different data aspects, they applied various perturbations to the Hi-C data and CNV labels (e.g., label shuffling, row shuffling, value noise).

Datasets

The study utilized Hi-C datasets from three human cancer cell lines:

  • RPMI8226 (multiple myeloma)
  • U266 (multiple myeloma)
  • K562 (chronic myelogenous leukemia)

CNV labels were generated using Whole Genome Sequencing (WGS) data from the same cell lines, processed with Control-FREEC at a 40-kb resolution, serving as the ground truth.

Key Insights and Ideas

  • Linearity is Chromosome-Specific: The relationship between Hi-C read counts and CNVs is evident but highly specific to individual chromosomes, making a single linear model for the entire genome ineffective.
  • GCN Captures Spatial Structure: By representing Hi-C data as a graph, the GCN can effectively learn complex spatial relationships within the chromatin structure, which are crucial for accurate CNV inference.
  • Structural Features are Key: The GCN's ability to predict CNVs is driven by the underlying chromosomal structural features, not just the CNV labels themselves. This is evidenced by the model's robustness to noise and its sensitivity to label perturbations.
  • Transfer Learning Potential: The GCN model shows promise for transfer learning, allowing for effective CNV prediction in new cell lines with minimal labeled data, suggesting conserved structural patterns across different cancer types.

Key Results

  • Linear Model Performance: Achieved high prediction accuracies (e.g., 97.64% for K562) when trained individually per chromosome, but significantly lower accuracy (e.g., 48.79% for U226) when a single unified model was attempted without dimension reduction.
  • GCN Performance: The GCN-based model achieved an average prediction accuracy of 95.95% across the three cell lines using raw Hi-C data. This outperformed traditional methods like HiNT and HiCnv.
  • Robustness to Noise: The GCN model demonstrated robustness to Gaussian noise added to Hi-C data, with only a minor decrease in accuracy (3.56% on average).
  • Sensitivity to Label Perturbations: The model's accuracy dropped dramatically (to 25.09%) when CNV labels were reconstructed, highlighting that the model learns the structural basis of CNVs rather than simply memorizing labels.

Limitations and Future Work

  • Dimensionality of GCN: While the GCN can handle various dimensions, a lower dimension (e.g., 16) might impair performance, suggesting an optimal range (e.g., >32) for stable performance.
  • Inter-Chromosomal Differences: While the GCN can model relationships across chromosomes within a single cell line, transfer learning experiments showed that structural features associated with CNVs can differ significantly between cell lines, requiring fine-tuning.
  • Future Work: The authors suggest that further research into mining structural variations from chromosomal characteristics can lead to better results, especially in mitigating Hi-C sequencing noise. The transfer learning approach also opens avenues for developing more generalized CNV identification methods across diverse genomic contexts.

Summary updated

BibTeX
@article{wang2024effectiveness,
  author = {Yuyang Wang and Yu Sun and Zeyu Liu and Bijia Chen and Hebing Chen and Chao Ren and Xuanwei Lin and Pengzhen Hu and Peiheng Jia and Xiang Xu and Kang Xu and Ximeng Liu and Hao Li and Xiaochen Bo},
  title = {{Effectiveness of machine learning at modeling the relationship between Hi-C data and copy number variation}},
  journal = {Quant. Biol.},
  year = {2024},
  doi = {10.1002/qub2.52},
}
Figure