Comparing General and Domain-Specific Pre-trained Language Models for Natural Product Classification

The classification of natural products (NPs) remains a central challenge in cheminformatics, due to their structural complexity and the limited availability of curated data. In this study, we explore whether general-purpose language models—pre-trained on non-chemical corpora—can achieve performance comparable to models explicitly trained on molecular data. We evaluate six models, including ChemBERTa-zinc-base-v1, ChemBERTa-77M-MLM, LLaMA, GPT-2, MPNet, and MiniLM, across two established datasets (REFMET and NPClassifier), using both unsupervised clustering (K-Means) and supervised classification (logistic regression, MLP). In addition, we assess model robustness under input degradation, introducing systematic SMILES occlusion and additive embedding noise. Our results show that general-purpose models can match or even surpass domain-specific alternatives, not only in clean classification scenarios but also under perturbation. Notably, general-purpose models demonstrate strong resilience and generalization, underscoring the transferability of domain-agnostic architectures to molecular tasks. These findings support the use of general-purpose language models as reliable tools for natural product classification, highlighting their potential to address complex challenges in cheminformatics through large-scale pre-trained architectures. Project page: https://github.com/bcorrad/MetaboLM25.