Data-driven characterization of molecular phenotypes across heterogenous sample collections

Juha Mehtonen, Petri Pölönen, Sergei Häyrynen, Jake Lin, Thomas Liuksiala, K Granberg, Olli Lohi, Ville Hautamäki, Matti NykterbioRxiv 2018 · DOI

TL;DR

This paper introduces a novel framework for integrating diverse genomic datasets to uncover molecular subtypes of diseases. By leveraging dimensionality reduction techniques like t-SNE and developing new metrics to assess data quality, the authors enable robust comparisons across studies and data types, leading to new insights into acute myeloid leukemia (AML).

What Problem Does This Paper Address?

The explosion of publicly available genomic data (like gene expression profiles) holds immense promise for understanding diseases and stratifying patients. However, a major hurdle is the heterogeneity of these datasets: they come from different labs, use different technologies (microarrays, RNA-seq), and have varying experimental protocols. This "technical variation" can obscure the underlying biological signals, making it difficult to compare results or combine data from multiple sources. The authors aim to bridge this gap by developing methods that can effectively integrate these heterogeneous datasets, allowing for a more comprehensive and reliable characterization of disease subtypes.

Methods

The core of the proposed methodology relies on the t-SNE (t-distributed Stochastic Neighbor Embedding) algorithm, a powerful dimensionality reduction technique that visualizes high-dimensional data in a low-dimensional space (typically 2D). The authors adapt t-SNE to handle heterogeneous data by:

  • Developing new metrics (NMIp and NMLe): These metrics quantify how well a clustering separates biological subtypes (phenotypes, NMIp) versus technical variations (data series, NMLe). This allows for objective evaluation and optimization of feature selection for dimensionality reduction.
  • Implementing a remapping algorithm: This allows new samples or even entire datasets to be projected onto an existing t-SNE map. This is crucial for integrating new data without recomputing the entire analysis, preserving the structure of the original map.
  • Adapting similarity metrics: For RNA-seq data, the Euclidean distance used in t-SNE was replaced with a correlation-based distance to better align it with microarray data.
  • Utilizing gene set enrichment analysis (GSVA): This technique is used to characterize the discovered clusters by identifying enriched biological pathways and functions, and to compare clusters across different datasets.
  • Integrating multi-omics data: The authors correlate cluster assignments with other data types like mutations, karyotypes, and methylation levels to provide a multi-faceted view of the identified subtypes.

Datasets

The primary dataset used for method development and testing is the Hemap dataset, comprising 9,544 gene expression profiles from various hematopoietic cancers, cell lines, and normal blood cells. This dataset was curated from the Gene Expression Omnibus (GEO).

Additional datasets used for validation and benchmarking include:

  • Left-out samples from the Hemap dataset: Used for remapping validation.
  • TCGA AML dataset: RNA-seq data from adult acute myeloid leukemia patients.
  • A separate microarray dataset (GSE49032): Used for remapping validation in acute lymphoblastic leukemia (ALL).
  • Ross ALL dataset: Microarray data for comparative analysis of ALL subtypes.

The paper does not introduce a new benchmark dataset but rather uses existing large-scale repositories to demonstrate the robustness of their approach.

Key Insights and Ideas

  • Technical variation is a major bottleneck: The paper strongly emphasizes that technical differences between datasets can easily mask biological signals, and robust methods are needed to overcome this.
  • t-SNE is a powerful tool for integration, but needs careful parameter tuning: While t-SNE is effective, the choice of features (genes) and preprocessing steps significantly impacts the results, especially in heterogeneous data.
  • New metrics are essential for quality control: NMIp and NMLe provide a quantitative way to assess the biological relevance and technical purity of dimensionality reduction results, guiding feature selection.
  • Remapping enables dynamic data integration: The ability to add new samples or datasets to an existing t-SNE map is a significant advancement, allowing for continuous updating and expansion of analyses.
  • Multi-omics data enriches subtype characterization: Integrating mutations, karyotypes, and methylation data provides a deeper understanding of the biological underpinnings of the identified molecular subtypes.

Key Results

  • The proposed framework successfully integrated multi-center and multi-platform data for hematological malignancies.
  • The remapping algorithm demonstrated robustness, correctly assigning new samples (microarray and RNA-seq) to their respective clusters with high accuracy (e.g., 88% for AML RNA-seq samples).
  • The analysis of AML revealed distinct molecular subtypes, including a cluster characterized by CEBPA mutations that showed significantly elevated SAM-dependent methyltransferase activity and global DNA methylation levels.
  • The new metrics (NMIp and NMLe) were shown to effectively guide feature selection, reducing technical biases and improving the separation of biological subtypes. For instance, selecting 15% of the most variable genes for t-SNE favorably reduced technical bias compared to using PCA preprocessing.

Limitations and Future Work

  • Representativeness of initial datasets: The authors note that the choice of the initial dataset for remapping or defining gene sets can influence results. The biological context and sample composition of the reference dataset are important considerations.
  • Generalizability to other data types: While the framework was extended to RNA-seq, its application to other single-cell technologies or more complex multi-omics integration might require further adaptation.
  • Future work: The authors suggest that their approach is highly relevant for single-cell contexts where batch effects are a significant challenge. They also highlight the potential for their framework to be applied to other biological systems and disease contexts beyond hematological malignancies.

Summary updated

Abstract

Abstract The existing large gene expression data repositories hold enormous potential to elucidate disease mechanisms, characterize changes in cellular pathways, and to stratify patients based on their molecular profile. To achieve this goal, integrative resources and tools are needed that allow comparison of results across datasets and data types. We propose an intuitive approach for data-driven stratifications of molecular profiles and benchmark our methodology using the dimensional reduction algorithm t-SNE with multi-center and multi-platform data representing hematological malignancies. Our approach enables assessing the contribution of biological versus technical variation to sample clustering, direct incorporation of additional datasets to the same low dimensional representation of molecular disease subtypes, comparison of sample groups between separate t-SNE representations, or maps, and characterization of the obtained clusters based on pathway databases and additional multi-omics data. In the example application, our approach revealed differential activity of SAM-dependent DNA methylation pathway in the acute myeloid leukemia patient cluster characterized with CEBPA mutations that accordingly was validated to have globally elevated DNA methylation levels.
BibTeX
@misc{mehtonen2018data,
  author = {Juha Mehtonen and Petri Pölönen and Sergei Häyrynen and Jake Lin and Thomas Liuksiala and K Granberg and Olli Lohi and Ville Hautamäki and Matti Nykter and Sergei Häyrynen},
  title = {{Data-driven characterization of molecular phenotypes across heterogenous sample collections}},
  year = {2018},
  doi = {10.1101/248096},
  howpublished = {bioRxiv},
}
Figure