In-silico identification of a Doxorubicin alternative with reduced cardiotoxicity informed by LLM-assisted modeling

Anthracyclines such as doxorubicin are potent anticancer agents but suffer from cardiotoxicity and poor clearance. We developed a hybrid pipeline integrating docking (2375 runs), molecular dynamics, and a large language model (LLM) to predict ADME/ADMET properties. Ligand-2 (CHEMBL1514139) showed strong affinity for UGT1A9 (-11.4 kcal/mol) and minimal interaction with topoisomerase IIβ. MD simulations confirmed stable binding. LLM-based regression predicted favorable ADME properties, including R² = 0.840 for clearance and R² > 0.89 for BBB and absorption. hERG II inhibition was classified using a random forest (RF) model, achieving 92 % accuracy and an R² score of 91 %. Model interpretability was further achieved using Shapley Additive Explanations (SHAP), which identified key molecular descriptors influencing pharmacokinetic and toxicity predictions. Ligand-2 demonstrated comparable or improved pharmacokinetics versus doxorubicin, supporting its potential as a safer anthracycline analog.