Self-Supervision Before Screening for Learning Natural Product Representations from Unlabelled Molecular and Spectral Collections

Download PDF

Natural-product discovery faces a fundamental asymmetry: chemical information is abundant, yet biological labels are sparse, heterogeneous, and task-specific. Molecular structures, provenance data, three-dimensional conformations, and experimental spectra thus offer substantial reusable knowledge before any bioactivity endpoint is assigned. This article develops a computational framework that treats such unlabelled data as a pretraining substrate rather than as material awaiting conventional screening. The framework distinguishes learning signals from molecular graphs, sequences, geometries, and spectra, and it proposes chemically informed pretext objectives that preserve distinctions relevant for later pharmacological reasoning while avoiding artifacts from dataset biases. Experimental spectra are treated as a complementary chemical view because they encode composition and structure under measurement-specific conditions, and cross-modal learning is positioned to relate partially shared rather than interchangeable representations. The proposed structure separates pretraining, chemically meaningful representation testing, rare-label fine-tuning, and transfer evaluation across scaffold, target, and experimental domains. Chemical meaningfulness is not reduced to benchmark accuracy; representations are instead examined for structural sensitivity, coherent chemical neighbourhoods, modality correspondence, resistance to shortcuts, and behaviour outside the training distribution. This framework remains methodological rather than a validated universal architecture. Its practical utility depends on data provenance, spectral quality, conformational assumptions, label definitions, chemical-space coverage, and prospective testing on genuinely unseen natural products.