Transfer Learning Without Chemical Amnesia for Predicting Natural Product Bioactivity Across Rare Scaffolds, New Targets, and Data-Sparse Therapeutic Domains

Download PDF

Natural products remain a valuable but challenging source of bioactive compounds due to their structural diversity, stereochemical complexity, and uneven representation in public datasets, which often diverge from the distributions used to train contemporary machine-learning models. Transfer learning provides a promising solution for data-sparse prediction tasks, yet adaptation to new targets can inadvertently suppress information that is critical for chemically rare scaffolds, a phenomenon termed chemical amnesia that conventional accuracy metrics may fail to detect. This conceptual framework systematically examines the determinants of transfer reliability, including source-domain support, representation retention, biological-context conditioning, adaptation intensity, few-shot support composition, and domain-shift evaluation. The proposed Chemical-Memory Transfer Framework argues that successful transfer should be assessed not only by improvements in target-task performance but also by whether chemically meaningful information remains accessible throughout the adaptation process. The framework challenges the assumption that more extensive pretraining, broader multitask learning, or parameter-efficient fine-tuning is inherently beneficial, emphasizing instead that rare-scaffold prediction requires explicit evaluation of what knowledge is transferred, what is forgotten, and under which chemical and biological shifts. Prospective validation using independently sourced natural-product datasets and novel target contexts remains essential for establishing the practical utility of this approach.