HaiLong Wu, ZhiMo Wu, XiaoQiong Liu
Reliable evaluation of gene regulatory network (GRN) inference requires strict separation between training and held-out regulatory edges, particularly for negative edges in sparse networks. We present Mamba-GRN, a compact Mamba-inspired gene representation and edge-decoding framework evaluated under a corrected no-overlap protocol in which validation and test negatives are excluded from the training-negative pool. The implemented encoder combines expression-derived features and learnable gene-identity embeddings with residual blocks composed of layer normalization, linear expansion, depthwise one-dimensional convolution, GELU activation, and linear projection; it does not implement a selective-scan state-space recurrence. Across seven non-tiny GSD datasets and three random seeds, the full model achieved mean AUROC 0.6225, AUPRC 0.4754, and Precision@P 0.4444, compared with 0.5708/0.4022/0.4444 for GENIE3 and 0.5671/0.4446/0.3810 for GRNBoost2. Paired mean improvements over the mature tree-based baselines were positive, but Holm-adjusted Wilcoxon tests did not reach the 0.05 threshold; the revised analysis therefore reports effect estimates, bootstrap confidence intervals, and win rates without claiming universal statistical superiority. Sensitivity analyses showed broadly stable performance across 1:1, 2:1, and 5:1 training-negative ratios, while larger representation dimensions improved mean performance at increased parameter cost. In an independent K562 Perturb-seq benchmark with 2284 aligned genes and 20,795 perturbation-response associations, source-matched hard-negative evaluation yielded AUROC 0.7582 ± 0.0071 and AUPRC 0.6114 ± 0.0124. The pretrained frozen backbone provided only a modest, seed-dependent advantage over a randomly initialized frozen backbone. These results support Mamba-GRN as a controlled framework for held-out edge recovery, while limiting the claims to the evaluated networks, candidate-edge setting, and functional perturbation-response associations.