XA4C: eXplainable representation learning via Autoencoders revealing Critical genes

Qing Li, Yang Yu, Pathum Kossinna, Theodore Lun, Wenyuan Liao, Qingrun Zhang

TL;DR

This paper introduces XA4C, a novel method that uses explainable autoencoders to identify "Critical genes" from gene expression data. These Critical genes offer new biological insights beyond traditional methods and show strong relevance to disease pathology, particularly in cancer.

What Problem Does This Paper Address?

Analyzing gene expression data, especially from complex diseases like cancer, often involves machine learning (ML) models. Autoencoders (AEs), a type of representation learning (RL) model, are good at finding meaningful patterns in noisy data by learning "latent variables." However, these learned representations are typically black boxes, making it hard to pinpoint specific genes that are driving these patterns or are functionally important. Traditional methods like identifying differentially expressed (DiffEx) or co-expressed (DiffCoEx) genes, or "Hub" genes in networks, capture only marginal effects or simple correlations, potentially missing more complex, non-linear interactions. This lack of interpretability hinders the practical application of powerful RL models for identifying key biological players.

Methods

The core of this work is the XA4C (eXplainable Autoencoder for Critical genes) tool. It combines two main components:

  1. Explainable Autoencoders: XA4C uses autoencoders to learn compressed representations (latent variables) of gene expression data. The novelty here lies in making these AEs "explainable."
  2. SHAP for Gene Contribution: To understand which genes contribute most to these learned latent variables, XA4C employs eXplainable Artificial Intelligence (XAI) techniques, specifically SHAP (SHapley Additive exPlanations). SHAP values are used to quantify each gene's contribution to each latent variable.

XA4C then aggregates these SHAP values to compute a "Critical index" for each gene. Genes with the highest Critical indexes are identified as "Critical genes." The authors also developed both whole-transcriptome AEs and pathway-specific AEs.

Datasets

The study utilized gene expression data from The Cancer Genome Atlas (TCGA) for six different cancer types:

  • Breast invasive carcinoma (BRCA)
  • Colon adenocarcinoma (COAD)
  • Kidney renal clear cell carcinoma (KIRC)
  • Lung adenocarcinoma (LUAD)
  • Prostate adenocarcinoma (PRAD)
  • Thyroid carcinoma (THCA)

The raw count data was processed into Transcripts Per Million (TPM) matrices.

Key Insights and Ideas

  • "Critical Genes" as a New Paradigm: The introduction of "Critical genes" provides a novel way to interpret RL models by directly linking learned representations back to individual genes. This offers a new perspective beyond traditional gene prioritization methods.
  • Beyond Marginal Effects: Critical genes capture complex, non-linear gene interactions that are missed by DiffEx (marginal effects) and DiffCoEx/Hub genes (correlations).
  • Disease Relevance: Critical genes show a significantly higher enrichment in disease gene databases like DisGeNET, suggesting they play crucial functional roles in disease pathology.
  • Novelty in Cancer Pathways: Critical genes often have low overlap with DiffEx and Hub genes, indicating they highlight previously unrecognized biological mechanisms.
  • Distinct Interaction Patterns: Critical genes can reveal distinct co-expression network patterns between tumor and normal tissues, even for genes not identified by traditional methods.

Key Results

  • High Reconstruction Accuracy: Whole-transcriptome AEs with a high compression ratio (32 latent variables) achieved good reconstruction accuracy (R² values from 0.42-0.69), comparable to AEs with many more latent variables.
  • Low Overlap with Traditional Methods: Critical genes showed poor overlap with Hub genes (identified by WGCNA) and DiffEx genes across all six cancers.
  • Superior Disease Gene Enrichment: Critical genes demonstrated significantly higher enrichment in the DisGeNET database compared to Hub and DiffEx genes.
  • Pathway Insights: Pathway enrichment analysis revealed that Critical genes are associated with key cancer-related pathways like oxidative phosphorylation and glutathione metabolism.
  • Example in Lysine Degradation: In the Lysine degradation pathway, Critical genes were identified that displayed distinct interaction patterns in tumor vs. normal tissues, highlighting their role in disease pathogenesis.

Limitations and Future Work

The authors acknowledge several limitations and suggest future directions:

  • AE Architecture Optimization: While they tested several AE configurations, a more comprehensive approach like Bayesian Hyperparameter Optimization could further refine architectures.
  • Expanding AE Types: The study primarily used conventional AEs. Future work could explore Variational AEs for causal inference or Graph AEs for learning causal structures.
  • SHAP Explainer Efficiency: For neural network models, more advanced SHAP explainers like DeepSHAP might be more efficient than TreeSHAP.
  • Sample Heterogeneity: The current Critical index summarizes SHAP values across all samples, ignoring individual sample heterogeneity. Future work will focus on uncovering individualized Critical genes for precision medicine.
  • Multi-omics Integration: XA4C currently supports only transcriptome data. Integrating additional omics data, such as proteomics and metabolomics, is a natural extension.

Summary updated

Abstract

ABSTRACT Machine Learning models have been frequently used in transcriptome analyses. Particularly, Representation Learning (RL), e.g., autoencoders, are effective in learning critical representations in noisy data. However, learned representations, e.g., the “latent variables” in an autoencoder, are difficult to interpret, not to mention prioritizing essential genes for functional follow-up. In contrast, in traditional analyses, one may identify important genes such as Differentially Expressed (DiffEx), Differentially Co-Expressed (DiffCoEx), and Hub genes. Intuitively, the complex gene-gene interactions may be beyond the capture of marginal effects (DiffEx) or correlations (DiffCoEx and Hub), indicating the need of powerful RL models. However, the lack of interpretability and individual target genes is an obstacle for RL’s broad use in practice. To facilitate interpretable analysis and gene-identification using RL, we propose “Critical genes”, defined as genes that contribute highly to learned representations (e.g., latent variables in an autoencoder). As a proof-of-concept, supported by eXplainable Artificial Intelligence (XAI), we implemented eXplainable Autoencoder for Critical genes (XA4C) that quantifies each gene’s contribution to latent variables, based on which Critical genes are prioritized. Applying XA4C to gene expression data in six cancers showed that Critical genes capture essential pathways underlying cancers. Remarkably, Critical genes has little overlap with Hub or DiffEx genes, however, has a higher enrichment in a comprehensive disease gene database (DisGeNET), evidencing its potential to disclose massive unknown biology . As an example, we discovered five Critical genes sitting in the center of Lysine degradation (hsa00310) pathway, displaying distinct interaction patterns in tumor and normal tissues. In conclusion, XA4C facilitates explainable analysis using RL and Critical genes discovered by explainable RL empowers the study of complex interactions. Author Summary We propose a gene expression data analysis tool, XA4C, which builds an eXplainable Autoencoder to reveal Critical genes. XA4C disentangles the black box of the neural network of an autoencoder by providing each gene’s contribution to the latent variables in the autoencoder. Next, a gene’s ability to contribute to the latent variables is used to define the importance of this gene, based on which XA4C prioritizes “Critical genes”. Notably, we discovered that Critical genes enjoy two properties: (1) Their overlap with traditional differentially expressed genes and hub genes are poor, suggesting that they indeed brought novel insights into transcriptome data that cannot be captured by traditional analysis. (2) The enrichment of Critical genes in a comprehensive disease gene database (DisGeNET) is higher than differentially expressed or hub genes, evidencing their strong relevance to disease pathology. Therefore, we conclude that XA4C can reveal an additional landscape of gene expression data.
BibTeX
@article{li2023xa,
  author = {Qing Li and Yang Yu and Pathum Kossinna and Theodore Lun and Wenyuan Liao and Qingrun Zhang},
  title = {{XA4C: eXplainable representation learning via Autoencoders revealing Critical genes}},
  journal = {PLOS Computational Biology},
  year = {2023},
  doi = {10.1371/journal.pcbi.1011476},
}
Figure