CoDiffGRN: Rethinking Gene Regulatory Network Inference via the BEELINE-KGC Benchmark and Co-evolutionary Discrete Diffusion

Traditional gene regulatory network (GRN) inference methods often fail to generalize to unseen genes and rely on global classification metrics that do not reflect the practical needs of biologists. This paper addresses the misalignment between current computational benchmarks and the real-world requirement to identify a small, high-confidence set of regulatory interactions for experimental validation.

To bridge this gap, the authors introduce BEELINE-KGC, a benchmark that reformulates GRN inference as an inductive, ranking-centric knowledge graph completion (KGC) problem. This benchmark uses an inductive gene-holdout split to test generalization to unseen genes and employs ranking-based metrics like Hits@K and Mean Reciprocal Rank (MRR) to evaluate top-K predictions.

To solve this, the authors propose CoDiffGRN, the first co-evolutionary discrete diffusion framework for GRNs. Key innovations include:

• Cell-cluster-based Discretization: Transforms continuous scRNA-seq data into discrete activation patterns to prevent feature collapse on unseen genes.
• Joint Discrete Diffusion: A co-evolutionary process that models gene expression and regulatory edges simultaneously, using a state-dependent transition matrix to ensure that edge transitions are conditioned on the activity states of the endpoint genes.
• TF-ALL Subgraph Sampling (TASS): A data augmentation strategy that preserves the integrity of transcription factor (TF)-target relationships, enabling scalable training under data scarcity.

Experimental results on the BEELINE-KGC benchmark demonstrate that CoDiffGRN significantly outperforms existing state-of-the-art methods in novel regulatory discovery. The joint modeling of gene states and regulatory edges, combined with the TASS training strategy, allows the model to maintain high performance even when generalizing to genes not seen during training. Ablation studies confirm that the co-evolutionary diffusion process and the discretization strategy are critical for achieving these gains.