XOmiVAE: an interpretable deep learning model for cancer classification using high-dimensional omics data

Eloise Withnell, Xiaoyu Zhang, Kai Sun, Yike Guo

TL;DR

This paper introduces XOmiVAE, a novel deep learning model that not only classifies cancer types from complex omics data with high accuracy but also explains why it makes those predictions. By revealing the contribution of individual genes and latent features to classification and clustering, XOmiVAE aims to bridge the gap between powerful deep learning models and the need for biological interpretability in cancer research.

What Problem Does This Paper Address?

High-dimensional omics data (like gene expression) is incredibly rich but presents a significant challenge for analysis due to its sheer volume – often having far more features than samples. This "curse of dimensionality" can lead to overfitting and hinder practical applications. While deep learning models excel at finding complex patterns in such data, they are notoriously "black boxes." This lack of explainability is a major hurdle in biomedical research, as it undermines trust and limits the ability to discover new biological insights. Existing methods for interpreting deep learning models have limitations, especially when dealing with the non-linear relationships inherent in omics data and the latent spaces learned by models like Variational Autoencoders (VAEs).

Methods

The core of XOmiVAE is a Variational Autoencoder (VAE) architecture, a standard approach for dimensionality reduction and feature learning. What's novel here is the integration of an explainability layer. Specifically, the authors leverage Deep SHAP, an activation-level-based interpretation method, to dissect the model's decision-making process.

Here's a breakdown of the key methodological contributions:

  • Explainable VAE for Omics: XOmiVAE builds upon existing VAE-based omics analysis models (like OmiVAE) but adds a crucial interpretability component.
  • Deep SHAP Integration: Deep SHAP is used to quantify the contribution of each input gene and each latent dimension to the model's predictions. This goes beyond simpler methods like connection weight analysis, which can be misleading.
  • Multi-level Explanations: XOmiVAE provides explanations for both:
  • Supervised Classification: Identifying which genes and latent dimensions are most important for predicting a specific cancer type.
  • Unsupervised Clustering: Explaining what drives the separation of different sample groups (clusters) learned by the VAE, even without predefined labels.
  • Customized SHAP Application: The Deep SHAP explainer was adapted to handle both latent vectors and classification outputs, and the authors emphasize the importance of selecting appropriate reference samples for meaningful explanations.

Datasets

The primary dataset used in this study is the Cancer Genome Atlas (TCGA) pan-cancer dataset. This dataset contains gene expression profiles for 33 different tumor types, along with normal tissue samples. The authors selected 9,081 samples for their experiments. The data was pre-processed by removing genes with zero expression or missing values, and gene expression values (FPKM) were normalized to a 0-1 range. Phenotype data (age, gender, primary site, disease stage) was also utilized.

Key Insights and Ideas

  • Bridging the Black Box: The most significant insight is the successful demonstration that deep learning models for omics data can be made interpretable, providing actionable biological insights rather than just predictions.
  • Gene-Specific Contributions: XOmiVAE can pinpoint specific genes that are critical for classifying a particular cancer type, offering a more focused view than traditional differential gene expression analysis.
  • Latent Dimension Meaning: The model can reveal the biological relevance of the abstract latent dimensions learned by the VAE, showing how these reduced representations capture biological processes.
  • Unsupervised Discovery: The ability to explain why the VAE clusters samples together is a powerful tool for discovering novel subtypes or understanding the biological basis of existing ones, especially when labeled data is scarce.
  • Reference Sample Importance: The choice of reference samples for SHAP analysis significantly impacts the explanations, highlighting the need for careful consideration of biological context.

Key Results

  • High Classification Accuracy: XOmiVAE achieved high accuracy in cancer classification, for example, 99.6% for Breast invasive Carcinoma (BRCA) versus normal breast tissue.
  • Biomedical Validation: The genes identified by XOmiVAE as important for classification were validated against existing biomedical knowledge (GeneCard) and found to be highly relevant. For BRCA, top genes like SCGB2A2 (Mammaglobin A) are known breast cancer markers.
  • Overlap with DEGs: There was a notable overlap (48 out of 100 genes for BRCA) between XOmiVAE's top genes and differentially expressed genes (DEGs) identified by traditional methods, but XOmiVAE also identified genes not strictly DEGs, suggesting it captures a broader set of discriminative features.
  • Latent Dimension Specificity: The model identified specific latent dimensions that were crucial for differentiating tumor subtypes (e.g., KICH, KIRC, KIRP), and these dimensions were linked to relevant biological pathways.
  • Gender Discrimination: A specific latent dimension (dimension 78) was found to strongly separate male and female samples, with XIST (a key gene in X chromosome inactivation) being a top contributor.
  • Performance Improvement: Using the top 20 genes identified by XOmiVAE for classification significantly improved model performance compared to using randomly selected genes.

Limitations and Future Work

  • Reference Sample Selection: The paper acknowledges that the choice of reference samples for SHAP analysis is crucial and can influence the explanations. While they suggest using biologically relevant references (e.g., normal tissue for cancer classification, opposite gender for gender analysis), this remains an area requiring careful consideration.
  • Model Saturation: While Deep SHAP addresses some issues of model saturation, the paper notes that individual latent dimension ablations had limited impact, suggesting that combined effects are more pronounced.
  • Computational Cost: Deep SHAP can be computationally intensive, though the authors do not explicitly detail this as a limitation in terms of performance.
  • Future Work: The authors suggest that XOmiVAE has great potential for discovering novel biomedical knowledge and can be applied to other omics data types and biological problems. Further exploration of optimal reference sample selection strategies and extending the model to other complex biological questions are implied.

Summary updated

Abstract

The lack of explainability is one of the most prominent disadvantages of deep learning applications in omics. This ‘black box’ problem can undermine the credibility and limit the practical implementation of biomedical deep learning models. Here we present XOmiVAE, a variational autoencoder (VAE)-based interpretable deep learning model for cancer classification using high-dimensional omics data. XOmiVAE is capable of revealing the contribution of each gene and latent dimension for each classification prediction and the correlation between each gene and each latent dimension. It is also demonstrated that XOmiVAE can explain not only the supervised classification but also the unsupervised clustering results from the deep learning network. To the best of our knowledge, XOmiVAE is one of the first activation level-based interpretable deep learning models explaining novel clusters generated by VAE. The explainable results generated by XOmiVAE were validated by both the performance of downstream tasks and the biomedical knowledge. In our experiments, XOmiVAE explanations of deep learning-based cancer classification and clustering aligned with current domain knowledge including biological annotation and academic literature, which shows great potential for novel biomedical knowledge discovery from deep learning models.
BibTeX
@article{withnell2021xomivae,
  author = {Eloise Withnell and Xiaoyu Zhang and Kai Sun and Yike Guo},
  title = {{XOmiVAE: an interpretable deep learning model for cancer classification using high-dimensional omics data}},
  journal = {Briefings in Bioinformatics},
  year = {2021},
  url = {https://openreview.net/forum?id=MyOu8vvY9w1},
}

Loading related content...

Figure