Cross-residual knowledge graph learning for robust multi-trait gene-trait prioritization in rice

Prioritizing genes associated with agronomic traits remains challenging because relevant evidence is distributed across heterogeneous biological resources and curated gene-trait associations are incomplete.

We formulate multi-trait crop gene prioritization as gene-trait link prediction on a rice-centered heterogeneous knowledge graph, where each candidate pair represents a graph-supported ranking hypothesis rather than an independent biological validation. We introduce a cross-residual knowledge-graph learning framework that keeps structural and relation-aware propagation in separate branches while allowing layer-wise residual exchange between them. This design lets broad topology and typed biological context refine each other during message passing, and residual filtering or transformation controls weak cross-branch corrections.

Under random and cold-gene evaluation protocols, Cross + JK variants provide consistent information gain within strict dual-branch architectural controls, remain near the top across the broader benchmark, and exhibit small seed-to-seed standard deviations on random-split ranking metrics. The default cold-gene split shows that feature-only MLP is a strong prior-driven baseline, whereas negative-ratio sensitivity shows that MLP becomes unstable when the sampled candidate distribution shifts and Cross + JK variants provide robust information gain in more uncertain candidate-ranking regimes. Broader comparisons with feature-only, heterogeneous graph, graph-Transformer, and KG-embedding baselines show that feature priors and alternative graph models remain strong in some regimes.

The results support cross-residual learning as a robust, condition-dependent strategy for integrating multi-source biological evidence in rice gene-trait prioritization.