New Benchmark and Model Advance Gene Regulatory Network Inference

Inferring gene regulatory networks (GRNs) from single-cell transcriptomic data is crucial for biological research, but existing methods and benchmarks often fail to align with practical needs. Current evaluations rely on transductive splits and global classification metrics, which don't adequately assess a model's ability to discover novel interactions involving previously unseen genes. To bridge this gap, researchers have reformulated GRN inference as an inductive, ranking-centric graph completion problem. They introduce BEELINE-KGC, a new benchmark that incorporates an inductive gene-holdout split and uses knowledge graph completion metrics to better evaluate top-ranked predictions, reflecting how biologists seek high-confidence interactions for validation. Building on this, the paper proposes CoDiffGRN, the first co-evolutionary discrete diffusion framework. CoDiffGRN jointly models biologically coherent discretized gene expression states and regulatory interactions, leading to robust inductive generalization and improved discovery of novel regulatory elements. It also includes TF-ALL Subgraph Sampling (TASS) for scalable training, demonstrating state-of-the-art performance on the new benchmark.