TL;DR
This paper introduces a novel "Conditional Hierarchical Bayesian Tucker Decomposition" (HBT) to uncover hidden patterns in complex genetic data, specifically focusing on identifying shared risk factors for common cancers and autism spectrum disorder (ASD). By integrating gene and pathway information into a hierarchical topic model, HBT provides more coherent and interpretable groupings of genetic variants and biological pathways compared to existing methods.
What Problem Does This Paper Address?
Analyzing genetic data for diseases like cancer and ASD is challenging due to its high dimensionality and sparsity. Traditional methods often struggle to identify complex relationships between genes and pathways that might contribute to disease risk. Existing approaches like Latent Dirichlet Allocation (LDA) are primarily designed for text data and don't naturally handle the multi-modal and hierarchical nature of biological information. The authors aim to bridge this gap by developing a method that can effectively group patients based on their genetic profiles, revealing underlying disease mechanisms and potential risk factors.
Methods
The core contribution is the Conditional Hierarchical Bayesian Tucker Decomposition (HBT). This method extends standard Bayesian Tucker decomposition by incorporating hierarchical topic models. Here's a breakdown:
- Bayesian Tucker Decomposition: A tensor factorization technique that models multi-dimensional data (tensors) by decomposing them into a core tensor and factor matrices. This allows for the discovery of latent factors across different modes (e.g., patients, genes, pathways).
- Hierarchical Topic Models: These models, inspired by nested Chinese Restaurant Processes (nCRP) and Pachinko Allocation Machines (PAM), allow for topics to be organized in a hierarchy. This means that broader categories of genes or pathways can lead to more specific sub-categories, mirroring biological organization.
- Conditional Aspect: The model focuses on the conditional probability of genetic variants and pathways given a patient, making it directly applicable to disease association studies.
- Novelty: The paper introduces a new formulation for conditional Bayesian Tucker decomposition, integrates hierarchical topic models into this framework, and develops a collapsed Gibbs sampler for efficient computation. They also propose a feature reduction method for sparse counting data.
Datasets
The authors evaluated their models on two primary datasets:
- Cancer Data: Derived from The Cancer Genome Atlas (TCGA), this dataset includes patients with breast, lung, prostate, and colorectal cancers. It contains genetic variants, genes, and pathways, with a hierarchical structure reflecting biological organization.
- Autism Spectrum Disorder (ASD) Data: Obtained from the National Database for Autism Research, this dataset includes paired siblings (one with ASD, one without) and their genetic variants. Pathway information was also incorporated.
A third dataset, Reuters, was used for natural language processing experiments to demonstrate the model's versatility.
Key Insights and Ideas
- Hierarchical Structure is Key: Biological data, like genes and pathways, naturally has a hierarchical structure. HBT effectively leverages this, leading to more interpretable and biologically relevant groupings than non-hierarchical models.
- Coherence Matters: The study emphasizes "coherence" as a key metric for evaluating topic models. HBT models consistently generated more coherent topics (meaning the genes/pathways within a topic are more related) than baseline models, suggesting better biological interpretability.
- Beyond Individual Genes: Focusing on groups of genes and pathways, rather than individual markers, can reveal more nuanced insights into disease mechanisms and risk factors.
- Model Flexibility: The HBT framework allows for different dependency structures between modes (e.g., genes and pathways), offering flexibility for various biological scenarios.
Key Results
- Improved Coherence: HBT models consistently outperformed baseline models (hLDA, CP tree, CP TensorLy) on cancer and ASD data across multiple coherence measures (PMI and UMass). For example, on cancer data, an unweighted HBT model improved UMass pathways coherence by 34.52% over the best baseline.
- Clinical Relevance: Topics discovered in the ASD model corresponded to known risk factors like immune function, neurotransmission, and lipid homeostasis, suggesting clinical relevance.
- Scalability: The paper demonstrates that HBT models scale better to multiple modes than comparable baselines, with faster training times for complex datasets.
- Classification Performance: While HBT models did not improve classification accuracy over raw genetic variant data, they performed competitively among decomposition models, with CP TensorLy showing the best performance among baselines.
Limitations and Future Work
- Strict Partition Property: The authors prove that for hierarchical topic models with multiple modes (p=2), the strict partition property (where topic labels and order don't matter) cannot strictly hold. This led them to focus on the "loose partition property."
- Computational Complexity: While HBT shows good scalability, training complex models can still be computationally intensive. The paper mentions exploring variations for performance improvement.
- Generalization to p > 3: The paper discusses generalizations to more than two feature modes, suggesting future work in exploring more complex dependency structures and model variations.
- Model Selection: The choice of the best model depends on the specific coherence measure and application, indicating that further research into optimal model selection strategies is warranted.
Summary updated