1 Master of Data Science, the University of Western Australia, Perth, Australia. Email: ladsanika01@gmail.com
2 Department of Computer Engineering, Thakur College of Engineering and Technology, Mumbai, India. Email: anantsingh1302@gmail.com
3 Department of Information Technology, Thakur College of Engineering and Technology, Mumbai, India. Email: shifakhan.work@gmail.com
4 Department of Artificial Intelligence & Machine Learning, Thakur College of Engineering and Technology, Mumbai, India. Email: afzalk280306@gmail.com
5 Master of Data Analytics Engineering, Northeastern University, Boston, MA, US. Email: patelkush9819@gmail.com
Drug delivery prediction has become an important problem in pharmaceutical informatics because formulation success depends on several tightly connected properties, including solubility, permeability, absorption, distribution behavior and toxicity. These properties rarely follow simple linear patterns and their behavior often changes across structurally different compounds which makes robust prediction difficult under real-world generalization settings. Artificial intelligence has emerged as a practical response to this difficulty because benchmark resources such as MoleculeNet and the Therapeutics Data Commons now provide curated molecular datasets, standardized evaluation tasks and scaffold-aware splitting schemes for realistic comparison. At the same time, the growing popularity of deep learning and newer liquid neural models has created a need for careful comparative work that separates genuine scientific benefit from architectural novelty. This paper presents a comparative research framework for machine learning, deep learning and Liquid Neural Networks in drug-delivery-relevant prediction tasks using real public datasets. The study is centered on benchmark datasets including ESOL, Lipophilicity, Caco2, HIA, BBB and AqSol which together cover regression and classification endpoints directly related to pharmaceutical transport and ADMET behavior. The proposed methodology is designed to stand out in four ways. First, it prioritizes scaffold-split generalization rather than relying only on random splits. Second, it requires repeated-run statistical testing to avoid reporting unstable gains. Third, it evaluates calibration and error behavior across chemical classes instead of reporting only a single mean score. Fourth, it includes rigorous feature-set ablations and a transparent discussion of when liquid architectures are scientifically justified. The main conclusion of the paper is straightforward. A strong drug-delivery prediction study should not ask only which model scores highest on average; it should ask which model generalizes best to unseen scaffolds, remains calibrated and is scientifically appropriate for the underlying prediction problem.
Keywords: Drug delivery prediction, ADMET modeling, Scaffold-split generalization, Machine learning benchmarking, Deep learning for drug discovery, Liquid Neural Networks, Molecular property prediction, Pharmaceutical AI
How to cite this article: Lad SS, Singh AM, Khan SS, Khan AS, Patel KP. Comparative Analysis of Machine Learning, Deep Learning and Liquid Neural Networks for Drug Delivery Prediction. Int J Drug Deliv Technol. 2026;16(5): 119-126. DOI: 10.25258/ijddt.16.5.12
Source of support: Nil.
Conflict of interest: None