NANO-AI has faced a useful test for a formulation model: four compounds absent from its training data—caffeine, nifurtimox, naringin and itraconazole. The study, published on 27 October, developed the model using 925 formulations containing nine other compounds.[1]
Molecular properties, ingredients and processing variables fed two stages: predicting whether tiny polymer carriers for medicines would form, then estimating their size and how much drug they held.[1]
Researchers made 15 formulations for each unfamiliar compound. Fifty-four of 60 formed particles. That 90% is a preparation success rate, not overall prediction accuracy. The mean absolute error in predicted particle size was 50.1 nanometres. Larger particles were harder to predict.[1]
The compounds differed in how far their chemical properties lay from the training examples. Unseen identities therefore offered varying degrees of unfamiliarity. A separate repeatability test used ten isoniazid formulations across three operator-and-laboratory settings.[1]
The experiments test transfer between compounds and experimental settings. Clinical performance remains a separate question.[1]
Sources & further reading
- A machine learning framework for polymeric nanoparticle design with zero-shot generalization across active pharmaceutical ingredients Drug Delivery and Translational Research · 27 Oct 2026 · Accessed 28 Oct 2026