A causal reinforcement learning framework for reliable gene regulatory network inference

Gene regulatory network inference is crucial for revealing the mechanisms of cellular functions and the causal mechanisms of disease occurrence. However, traditional correlation-based methods struggle to identify causal directions, and deep learning methods lack structural interpretability. Moreover, neither of them can efficiently adapt to the search of the solution space for large-scale gene regulatory networks, making it difficult to accurately capture the real regulatory patterns between genes. This study proposes a gene regulatory network inference method based on causal reinforcement learning to address the above problems. Experiments on the DREAM5 dataset and scRNA-seq datasets of six cell types from BEELINE demonstrate that the proposed method outperforms existing mainstream baseline models in inference accuracy, structural sparsity, and biological plausibility. Further validation via GNNExplainer interpretability analysis and KEGG pathway enrichment shows that the constructed gene regulatory network clarifies causal relationships in gene regulation, enhances structural interpretability, and aligns closely with real biological networks and known biological pathways in both topological and functional aspects. The proposed causal reinforcement learning-based framework realizes efficient global optimization of large-scale gene network structures, and the inferred regulatory networks possess high causal rationality and biological interpretability. These results confirm the feasibility and effectiveness of the method for real-world applications in systems biology and computational biology, providing a scalable approach for causal structure modeling of high-dimensional gene expression data.