Abstract
Inferring Gene Regulatory Networks (GRNs) from single-cell RNA sequencing (scRNA-seq) data is a fundamental challenge in computational biology, primarily due to the high levels of noise and "dropout" events inherent in the technology. This paper introduces 3DCEMA (3D Co-Expression Matrix Analysis), a novel deep learning framework that employs 3D Convolutional Neural Networks (CNNs) to infer regulatory relationships. By shifting from gene pairs to gene triplets and representing their joint distributions as 3D voxel grids, 3DCEMA effectively filters out technical artifacts. The model demonstrates state-of-the-art accuracy and exceptional stability across various simulated and real-world datasets, even under extreme sparsity.
Introduction
Understanding the complex "circuitry" of how transcription factors regulate genes is essential for decoding cellular behavior. While single-cell RNA sequencing (scRNA-seq) provides high-resolution data, it is notoriously messy. Two major hurdles—zero-inflation (dropouts) and a low signal-to-noise ratio—often overwhelm traditional regulatory inference methods. Furthermore, many existing tools rely on "pseudotime" estimations, which can propagate errors throughout the network inference process. 3DCEMA addresses these issues by leveraging the spatial pattern-recognition capabilities of 3D CNNs.
Methods: The 3D Co-Expression Framework
3DCEMA shifts the perspective of network inference from gene pairs to gene triplets $(x, y, z)$. For each triplet, the model constructs a $16 \times 16 \times 16$ voxel grid where each voxel represents the joint distribution of the three genes across all cells.
The "Reference" Gene Concept
By introducing a third gene ($z$) as a comparison factor for the pair $(x, y)$, the model focuses on the asymmetry of regulation. Because technical noise and dropouts typically affect all genes within a single cell similarly, comparing the expression of $y$ and $z$ relative to $x$ helps cancel out common noise. To maintain computational efficiency and avoid an $O(g^3)$ complexity, the authors introduced an "AVG" gene—a synthetic reference representing the average expression of all genes—reducing complexity to $O(g^2c)$.
Visualizing Regulatory Patterns
The core innovation of 3DCEMA lies in treating gene regulation as a spatial classification problem. The model uses a ResNet-10 architecture to recognize patterns in the 3D "cloud" of gene expression.
As illustrated in the figure above, different regulatory types correspond to specific spatial distributions. The authors observed that as the expression of the regulator ($x$) increases, the cluster of high-intensity voxels moves in predictable directions, resembling a "clock face" of regulation. This relative movement is significantly more robust to data sparsity than absolute correlation metrics.
Key Results
3DCEMA was evaluated on both simulated (BEELINE framework) and real-world scRNA-seq datasets (mESC and mHSC-E).
- Superior Stability: The model maintained high performance even at 70% dropout rates. In simulated GSD datasets, it achieved an AUPRC of approximately 0.8, nearly doubling the performance of several traditional methods.
- Accuracy on Real Data: 3DCEMA consistently outperformed state-of-the-art methods like CNNC, PIDC, and GRISLI on mouse embryonic and hematopoietic stem cell data.
- Transferability: Models trained on simulated data generalized remarkably well to real biological datasets, suggesting that the visual patterns of gene regulation are universal enough for deep learning to identify across different species and tissues.
Conclusion and Future Work
3DCEMA demonstrates that 3D CNNs can effectively overcome the noise limitations of scRNA-seq data by analyzing the joint distribution of gene triplets. While the method is computationally more intensive than 1D or 2D approaches, its robustness to dropouts makes it a powerful tool for discovering hidden biological relationships. Future work will focus on refining edge weights to determine the strength of regulation and improving the identification of causal directionality in symmetric co-expression patterns.
Summary updated