STaR: Unified Spatiotemporal-Causal Graph Reasoning for Molecular Property Prediction
Jiaxuan Li, patoiciar, 550275784, 1765081253 · 2025
We propose STaR (Spatiotemporal Topology-aware Reasoning), a unified and theory-grounded framework for spatiotemporal-causal graph inference that integrates spatial structures, temporal evolution patterns, and causal semantics within a cohesive architecture. STaR combines a multimodal Transformer that unifies molecular graphs and SMILES sequences through cross-modal attention, a temporal graph integrator with multi-head positional attention to capture dynamic reaction behaviors, an adaptive diffusion kernel that enables high-order interaction modeling via a learnable propagation mechanism to mitigate over-smoothing and enhance expressive message passing, and a causal intervention module grounded in structural causal models for counterfactual reasoning and fine-grained interpretability. Experiments on the ZINC benchmark demonstrate that STaR achieves a 9.2% reduction in mean absolute error compared to state-of-the-art graph models, with only a 7% runtime overhead on both NVIDIA RTX 4060 and A100 GPUs. Furthermore, zero-shot transfer to a scholarly citation forecasting task involving 34 million temporal edges yields a 13.5% improvement in nDCG@20 over leading temporal graph baselines. These results highlight STaR as a universal, interpretable, and scalable platform for spatiotemporal reasoning, with broad applicability across molecular modeling, scientific discovery, and temporal knowledge graphs. To facilitate reproducibility, all code, pretrained models, and datasets will be made publicly available as part of a unified release upon acceptance.