MoleculeX-Predictive: An augmentation of transformers with explainable AI and knowledge layers for efficient drug property prediction

We present MoleculeX-Predictive, a knowledge-augmented transformer framework for explainable molecular property prediction. The system fine-tunes ChemBERTa on 16,847 ChEMBL compounds, concatenating learned SMILES embeddings with five RDKit physicochemical descriptors through a two-layer MLP to jointly predict pIC50, logP, and atom count. Under scaffold-based cross-validation the model achieves R 2 = 0 . 918 ± 0 . 012 for pIC50 (MAE = 0.152), R 2 = 0 . 867 ± 0 . 015 for logP, and R 2 = 0 . 991 ± 0 . 003 for atom count, outperforming a standard ChemBERTa baseline by 10.5% on pIC50 and a Random Forest trained on the same descriptor set by 40.4%. To interpret these predictions we integrate three complementary XAI methods—SHAP, LIME, and Integrated Gradients—and show that their top-feature rankings correlate at r = 0 . 68 –0.73, with 78% three-way agreement on the most influential descriptors. Across 200 test molecules the explanations identify atoms belonging to known pharmacophore motifs in 85% of cases. These results demonstrate that augmenting a chemical-language model with domain knowledge can improve both predictive accuracy and the chemical interpretability of the resulting explanations. • Knowledge-augmented ChemBERTa achieving R 2 = 0 . 918 for pIC50 on scaffold splits, outperforming standard ChemBERTa by 10.5%. • Multi-modal XAI framework (SHAP, LIME, Integrated Gradients) with 78% three-way agreement and 85% pharmacophore identification rate. • Ablation study isolating contributions of transformer embeddings, RDKit descriptors, and knowledge-fusion layer.