XA4C: eXplainable representation learning via Autoencoders revealing Critical genes

Qing Li, Qing Li 0076, 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" – genes that significantly contribute to learned representations of gene expression data. Unlike traditional methods, XA4C provides a more nuanced understanding of gene function by revealing genes that capture complex interactions, offering new insights into disease mechanisms.

What Problem Does This Paper Address?

Understanding the complex interplay of genes in biological systems, particularly in diseases like cancer, is a major challenge. Traditional methods for identifying important genes, such as those that are differentially expressed (DiffEx) or highly connected in co-expression networks (Hub genes), often capture only marginal effects or simple correlations. While representation learning (RL) techniques, like autoencoders, can uncover complex, non-linear relationships in noisy data, the learned "latent variables" are notoriously difficult to interpret. This lack of interpretability hinders the ability to pinpoint specific genes for functional follow-up. The paper addresses this gap by developing a method that bridges the interpretability gap in RL, allowing for the identification of functionally relevant genes that might be missed by conventional approaches.

Methods

The core of XA4C is an eXplainable Autoencoder (XA4C). It combines two key components:

  • Autoencoders (AEs): These are unsupervised neural networks used for representation learning. XA4C employs AEs to learn compressed, low-dimensional representations (latent variables) of gene expression data. The paper uses both whole-transcriptome AEs and pathway-specific AEs.
  • Explainable Artificial Intelligence (XAI) - SHAP: To interpret the AE's learned representations, XA4C utilizes the SHapley Additive exPlanations (SHAP) framework. Specifically, it uses TreeSHAP with an XGBoost Regressor model. SHAP quantifies the contribution of each individual gene to each latent variable.

XA4C then defines a "Critical index" for each gene by averaging the absolute SHAP values across all latent variables and samples, weighted by the performance of the XGBoost model for each latent variable. Genes with the highest Critical indexes (top 1% by default) are identified as "Critical genes."

Datasets

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

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

The data was downloaded from the TCGA data portal and processed to retain genes with median expression levels higher than 1.0, resulting in approximately 15,000 genes per cancer type.

Key Insights and Ideas

  • "Critical genes" offer a new perspective: XA4C introduces the concept of "Critical genes," which are distinct from traditional DiffEx and Hub genes. They represent genes that are crucial for understanding the complex, non-linear relationships captured by autoencoders.
  • Beyond marginal effects and simple correlations: Critical genes capture more than just individual gene expression changes or pairwise correlations, suggesting they are involved in more intricate biological processes.
  • High relevance to disease: Critical genes show a significantly higher enrichment in disease gene databases (like DisGeNET) compared to DiffEx or Hub genes, indicating their strong functional relevance to disease pathology.
  • Novel biological insights: The low overlap between Critical genes and traditional gene sets suggests that XA4C uncovers novel biological information that was previously hidden.
  • Dynamic interaction patterns: Critical genes can highlight distinct interaction patterns between tumor and normal tissues within specific pathways, offering a deeper understanding of disease mechanisms.

Key Results

  • XA4C identifies novel genes: The study found that Critical genes have poor overlap with DiffEx and Hub genes across all six TCGA cancer types.
  • Superior disease gene enrichment: Critical genes demonstrated significantly higher enrichment in the DisGeNET database compared to DiffEx and Hub genes, indicating their greater relevance to disease. This was quantified by a higher success rate in identifying validated genes.
  • Pathway enrichment: XA4C identified pathways crucial in multiple cancers, including oxidative phosphorylation (OXPHOS) and glutathione metabolism.
  • Distinct interaction patterns: In the Lysine degradation pathway, XA4C identified Critical genes that were neither Hub nor DiffEx genes. These Critical genes showed dramatically different co-expression patterns between tumor and normal tissues, as evidenced by Pearson correlation analysis and Kolmogorov-Smirnov tests.
  • Performance comparison: XA4C outperformed other methods in terms of F1-score and accuracy when validated against DisGeNET, largely due to its superior precision.

Limitations and Future Work

  • Fixed AE architectures: The study used fixed autoencoder architectures. The authors suggest that incorporating Bayesian Hyperparameter Optimization could lead to more extensive exploration of optimal architectures.
  • Conventional AEs: Only conventional AEs were used. Future work could extend this to other AE variants like Variational AEs for causal inference or Graph AEs for learning causal structures.
  • TreeSHAP efficiency: While effective, TreeSHAP can be less efficient for deep neural networks. The authors suggest exploring other SHAP explainers like DeepSHAP for potentially better performance.
  • Sample-wise heterogeneity: The current Critical index is an average across all samples. Future work aims to uncover individualized Critical genes to support precision medicine by analyzing sample-specific heterogeneity.
  • Transcriptome data only: XA4C currently supports only transcriptome data. A natural extension would be to incorporate other omics data, such as proteomics and metabolomics.

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 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