AutoGERN: single-cell RNA-seq gene regulatory network inference via explicit link modeling and adaptive architectures

Single-cell RNA sequencing (scRNA-seq) enables transcriptome-wide profiling at single-cell resolution, revealing heterogeneous regulatory programs and making gene regulatory network (GRN) inference both central and challenging. Recent graph neural network (GNN)-based approaches for GRN inference typically model edges only implicitly (for example, via concatenated node embeddings), which limits their ability to capture complex regulatory dependencies. In addition, distributional shifts across scRNA-seq datasets make a single fixed GNN architecture poorly suited for broad generalization.

We present AutoGERN, a GNN framework tailored for GRN inference from scRNA-seq data. AutoGERN explicitly models regulatory information in the message-passing space and learns expressive link (edge) embeddings, which a lightweight multilayer perceptron uses to score gene-gene regulatory associations. To enhance flexibility and representational power, AutoGERN employs dual message-passing spaces (within-layer and cross-layer) and integrates a robust AutoGNN-based architecture search to adapt the network design to differing dataset distributions. Extensive experiments on multiple real scRNA-seq datasets demonstrate that AutoGERN consistently achieves superior performance and robustness compared with state-of-the-art baselines.

The code and data of AutoGERN are available on GitHub at https://github.com/JChander/AutoGERN and on Zenodo at https://doi.org/10.5281/zenodo.18659807.