ABSTRACT Graph neural networks (GNNs) have emerged as a powerful tool for molecular property prediction by representing chemical structures as graphs, yet they frequently struggle with oversmoothing and the inefficient capture of high‐order feature interactions. This paper proposes the wide and cross graph isomorphism network (WCGIN), an architecture designed to address these limitations through the integration of explicit cross‐feature interaction layers and a soft mixture‐of‐experts (MoE) parallel scaling paradigm. By incorporating matrix‐based cross‐layers into the graph isomorphism network (GIN) aggregation process, the model explicitly forms nonlinear feature relationships that might otherwise be lost during neighborhood processing. Simultaneously, the parallel MoE structure facilitates scaling by allowing multiple expert GNN modules to learn distinct representations, which are then aggregated via a gating network to mitigate the representational collapse typical of deep sequential stacking. The performance of WCGIN was evaluated across seven molecular benchmarks covering physical chemistry and biology, including four regression and three classification tasks. Experimental results indicate that adding even a single cross‐layer boosts performance in most cases, and the parallel scaling approach significantly outperforms standard stacked configurations for models with fewer than 10 layers. Furthermore, the study demonstrates that specialized mechanisms like load‐balancing loss are not required for these compact models to achieve robust predictive gains. These findings suggest that combining explicit feature interactions with parallel processing offers an efficient and effective pathway for scaling GNNs in chemical property prediction.
Racki et al. (Tue,) studied this question.