MSformer: A Meta-Structure Based Interpretable Framework for Representation Learning of Natural Products

Natural products (NPs) are a treasure trove of drug discovery, yet their structural complexity and extreme data scarcity critically hinder AI-driven exploration. To address this challenge, we present MSformer, a transformer-based architecture that bridges this gap by leveraging molecule fragments to systematically encode NP chemical space. These fragments were generated by a mass spectrometry-inspired fragmentation algorithm, termed meta-structures. Unlike chemical models pretrained on comprehensive molecule databases, MSformer is totally pretrained on very limited NP data set by deconstructing 400,000 NPs into 234 million meta-structures. This design enables MSformer to capture the structural richness and drug-like relevance of NPs. Evaluated on 14 tasks across MoleculeNet and the Therapeutics Data Commons data sets, MSformer outperforms state-of-the-art models, demonstrating superior generalizability in property prediction. The abundant meta-structures enable MSformer hierarchical interpretability that reveals task-specific structural determinants and successfully deconstructing approved drugs into bioactive fragments. By integrating domain knowledge with deep learning, MSformer establishes a transformative paradigm for NP-based drug discovery, offering a scalable framework to navigate nature's underexplored chemical repertoire and accelerate the identification of bioactive candidates.