What drives performance in molecular MPNNs? An operator-level factorial benchmark

Researchers present a factorial benchmark decomposing 2D molecular message-passing neural networks into 84 distinct configurations to identify which operator components drive molecular property prediction performance. The study finds that message construction methods significantly outweigh update complexity in determining model effectiveness, with concatenation-based mixing showing superior performance in differentiating molecular structures.

This research addresses a critical gap in understanding what makes molecular neural networks effective by systematically isolating and testing individual components rather than treating models as black boxes. The factorial approach—examining message-seed initialization, node-edge fusion, and update operators independently—provides empirical evidence that not all architectural choices contribute equally to performance. The finding that message construction dominates performance while update complexity shows no statistically significant effect challenges conventional assumptions in model design and suggests researchers have been optimizing the wrong aspects of these systems.

The work builds on growing recognition that molecular graph neural networks need more interpretable design methodologies. Prior approaches often focused on increasing model complexity or combining components without understanding their individual contributions. This benchmark demonstrates that chemically-informed design choices, particularly how information enters the message-passing pipeline, matter far more than computational sophistication. The competitive performance on eight of ten MoleculeNet datasets validates the practical utility of these insights.

For the broader AI and computational chemistry community, this research democratizes molecular MPNN design by replacing expensive architecture search with targeted, informed decisions. The mechanistic analysis of concatenation-based mixing's ability to better preserve chemical distinctions and resist oversmoothing provides actionable guidance for practitioners. This approach could accelerate development cycles for drug discovery and materials science applications relying on molecular property prediction. The methodology itself—factorial benchmarking with statistical rigor—establishes a replicable framework for understanding deep learning component contributions across other domains.

• →Message-seed initialization and node-edge fusion have statistically significant effects on MPNN performance, while update operators show no meaningful impact
• →Concatenation-based mixing outperforms Hadamard gating in differentiating molecular heteroatoms and preventing oversmoothing
• →Representative configurations selected through this analysis achieve competitive or best performance on 80% of benchmark datasets
• →Factorial decomposition enables targeted architectural improvements rather than black-box hyperparameter optimization
• →The framework provides empirical design heuristics for molecular neural network development in drug discovery and materials science