Self-supervised deep learning of gene-gene interactions for improved gene expression recovery

Qingyue Wei, Md Tauhidul Islam, Yuyin Zhou, Lei Xing

Introduction

Single-cell RNA sequencing (scRNA-seq) has transformed our ability to study cellular heterogeneity at an unprecedented resolution. However, technical limitations often lead to "dropout" events, where expressed genes are incorrectly recorded as zero counts due to low transcript capture efficiency. This data sparsity obscures biological signals, complicating cell type identification and trajectory inference. Existing imputation methods often rely on rigid statistical assumptions or fail to capture the complex, long-range gene-gene interactions essential for accurate recovery. To address this, we present TCER (Transform-and-Conquer Expression Recovery), a framework that leverages biological context through spatial mapping and self-supervised learning.

Methodology: The TCER Framework

TCER shifts the paradigm from processing 1D gene lists to analyzing 2D "GenoMaps." The framework consists of two primary stages: GenoMap construction and recovery via a specialized deep learning architecture.

GenoMap Construction

Using the Gromov-Wasserstein discrepancy, TCER repositions genes into a 2D grid where biologically interacting genes are placed in close physical proximity. This transformation converts abstract genomic data into a structured image format, enabling the use of advanced computer vision techniques to identify patterns of co-expression.

Visual representation of GenoMaps, converting 1D gene expression into 2D structured images.

ER-Net Architecture

The core of TCER is the ER-Net (Expression Recovery Network), a specialized encoder-decoder CNN featuring:

  • Deformable Fusion Attention (DFA): Unlike standard CNNs with fixed square filters, DFA allows the network to adaptively reshape its receptive field. This helps the model capture complex, non-linear gene interactions that are spatially distant on the GenoMap.
  • Dual-Attention Mechanism: The model utilizes both channel-wise and pixel-wise attention to prioritize the most informative genes and spatial features.
  • Self-Supervised Learning: The model is trained via dropout simulation, learning to reconstruct masked values and internalizing the underlying "grammar" of gene expression without requiring external ground-truth labels.

Experimental Results

TCER was validated against state-of-the-art methods across diverse simulated and real-world datasets, including mammalian brain, mouse intestinal epithelium, and zebrafish development data.

Improved Clustering and Trajectory Recovery

Visualizations using t-SNE and UMAP demonstrate that TCER-imputed data yields significantly cleaner and more distinct cell clusters compared to methods like scVI and MAGIC. Quantitatively, TCER consistently outperformed competitors across multiple metrics, including Accuracy, Adjusted Rand Index (ARI), and Normalized Mutual Information (NMI).

Comparison of TCER against other methods using t-SNE/UMAP visualizations and quantitative clustering metrics.

Robustness to Sparsity

TCER maintains high performance even at extreme dropout levels. For instance, in a mammalian brain dataset with only 0.4% capture efficiency, TCER achieved a Pearson correlation of 0.7974, while traditional deep learning models like scScope and dca failed to recover meaningful biological signals.

Discussion and Future Work

TCER demonstrates that treating genomic data as a spatial problem effectively leverages biological interaction context for data recovery. While the current grid-based mapping is highly effective, future iterations may explore non-Euclidean graph-based mappings to capture even more nuanced regulatory networks. Additionally, optimizing the computational cost of GenoMap construction will be key for scaling the framework to massive datasets with hundreds of thousands of cells. This "transform-and-conquer" strategy holds significant promise for other sparse biological modalities, such as proteomics and metabolomics, where interaction networks are equally vital.

Summary updated

Abstract

Single-cell RNA sequencing (scRNA-seq) has emerged as a powerful tool to gain biological insights at the cellular level. However, due to technical limitations of the existing sequencing technologies, low gene expression values are often omitted, leading to inaccurate gene counts. Existing methods, including advanced deep learning techniques, struggle to reliably impute gene expressions due to a lack of mechanisms that explicitly consider the underlying biological knowledge of the system. In reality, it has long been recognized that gene–gene interactions may serve as reflective indicators of underlying biology processes, presenting discriminative signatures of the cells. A genomic data analysis framework that is capable of leveraging the underlying gene–gene interactions is thus highly desirable and could allow for more reliable identification of distinctive patterns of the genomic data through extraction and integration of intricate biological characteristics of the genomic data. Here we tackle the problem in two steps to exploit the gene–gene interactions of the system. We first reposition the genes into a 2D grid such that their spatial configuration reflects their interactive relationships. To alleviate the need for labeled ground truth gene expression datasets, a self-supervised 2D convolutional neural network is employed to extract the contextual features of the interactions from the spatially configured genes and impute the omitted values. Extensive experiments with both simulated and experimental scRNA-seq datasets are carried out to demonstrate the superior performance of the proposed strategy against the existing imputation methods.
BibTeX
@article{wei2024self,
  author = {Qingyue Wei and Md Tauhidul Islam and Yuyin Zhou and Lei Xing},
  title = {{Self-supervised deep learning of gene-gene interactions for improved gene expression recovery}},
  journal = {Briefings in Bioinformatics},
  year = {2024},
  doi = {10.1093/bib/bbae031},
}
Figure