Uni-Macro: A Flexible 3D Transformer for Multicomponent Molecular Systems with Insights into Polymer Deep Learning Paradigms

Most polymer Transformers are built on monomer-level representations and have shown promise for predicting polymer properties. However, many existing polymer Transformers rely either on string-only inputs or incorporate limited pairwise structural information, such as interatomic distances or graph distances. Here, we introduce Uni-Macro, a flexible architecture that integrates computation modules rich in atomic embeddings and pairwise structural representations, achieving competitive performance across diverse polymer benchmarks without relying on large-scale pretraining. In addition, most Transformer architectures with pairwise structural features in both small-molecule and polymer modeling are not well-suited for disconnected multimolecule systems. To address this limitation, we propose a simple yet efficient grouping strategy that enables effective handling of multicomponent inputs while preserving structural patterns. Leveraging this strong framework, we conduct an ablation study, and the results show that while atomic encodings and structural representations are individually beneficial, performance gains from monomer-centric modeling show diminishing returns. Consistent with this observation, evidence from pretraining and model capacities indicates that neither large-scale pretraining nor expanded model capacity yields systematic improvements under end-to-end training. Together, these results establish Uni-Macro as a competitive and versatile 3D Transformer for polymer property prediction, and the results of diminishing returns under our framework indicate polymer deep learning as a challenging and scientifically rich frontier calling for higher-order physical effects.