Predicting lymphatic transport potential using graph transformer based on limited historical data from in vivo studies.
Journal:
Journal of controlled release : official journal of the Controlled Release Society
Published Date:
Aug 10, 2025
Abstract
The lymphatic system hosts a large number of therapeutic targets that can be used to modulate a wide range of diseases including cancers, autoimmune and inflammatory disorders, infectious diseases and metabolic syndrome; however, drug access to the lymphatic system is often challenging. Over the past decades significant efforts have been made to promote drug transport to the lymphatics through medicinal chemistry approaches, and a number of promising progresses are emerging. Nevertheless, so far it remains difficult to clearly delineate the mechanism of lymphatic drug transport and to map the design criteria for lymphotropic drug molecules, and the attempts to synthesize lymph-directing drug candidates or drug derivatives are largely in an experience-driven, trial and error basis. Furthermore, complex experimental procedures required for the study of lymphatic drug transport have limited data accumulation in the field, and this in turn hampers mechanistic studies and understanding of drug design criteria. Our current study aims to 1) review and summarize published work that assessed lymphatic drug transport by both direct measurement (e.g. determination of drug concentrations in lymph fluid) or indirect measurement (e.g. imaging methods or by comparing the changes of pharmacokinetics profile in the absence and presence of lymphatic transport blocker); 2) to analyze lymphatic drug transport data of 185 drugs according to experimental models and conditions, followed by dataset regrouping according to the extent of lymphatic transport; 3) to establish different Artificial Intelligence (AI) models including Graph Convolutional Network (GCN), Graph Attention Network (GAT) and Graph Transformer (GT) to predict the potential of drug transport via the lymphatics following oral administration, during which process data augmentation approaches were employed to compensate for the limited data. The results demonstrated that our model can enhance data and lymphatic drug transport prediction by correlating in vivo data with the chemical structure of drugs (represented by Simplified Molecular Input Line Entry System, SMILES). Additionally, we analyzed the relationship between the extent of lymphatic transport and a number of physicochemical parameters (including LogP, LogD and molecular weight) of drugs with reported lymphatic absorption data. The results demonstrate that the capability of lymphatic transport does not appear to be determined by any single parameter alone.
