BiMSGP: A Bidirectional Mamba-Based Model for Scalable and Accurate Genomic Prediction in Plants

Background/Objectives: Accurate prediction of complex traits from high-dimensional genomic markers remains difficult in plant breeding. Linear models have limited capacity for non-linear interactions, and many deep learning models scale poorly on long marker sequences. This study introduces BiMSGP, a bidirectional Mamba-based model for scalable genomic prediction. Methods: BiMSGP extends a selective state space model with forward and backward processing so that each marker can integrate sequence context from both directions while retaining linear computational complexity; this is a modeling device for long-range marker dependencies, not a demonstration of biological epistasis. Regression was evaluated on Wheat599, Wheat2000, and SoyBase, and classification on SoyBase, using 10-fold cross-validation with shared fold assignments and no inner validation split; the held-out fold was used for checkpoint selection, learning-rate adjustment, and reporting. Ablation studies compared bidirectional and unidirectional stacks and different depths. Exploratory paired tests of fold-wise PCC differences were used to distinguish numerical rank from a paired difference under this protocol. Results: Under the reported evaluation protocol, BiMSGP attained the numerically highest mean Pearson correlation coefficient (PCC) in all four Wheat599 environments, six of eight Wheat2000 traits, and five of seven quantitative SoyBase traits. Several of these numerical ranks did not have exploratory pt<0.05 against the strongest baseline; GBLUP had a higher mean PCC for SoyBase oil (pt=0.025). Shallow bidirectional designs had higher mean PCC than deeper and unidirectional variants on Wheat599 Environment 1. Training cost was higher than that of lightweight baselines on larger datasets, but inference remained in the millisecond range. Conclusions: Under this protocol, bidirectional selective state space modeling yielded competitive mean scores with a higher training cost than lightweight baselines. The scores are exploratory and may be optimistic; they are not unbiased estimates of generalization. Evaluation that separates model selection from assessment, and tests in independent populations, years, or environments, remains future work.