Abstract
<jats:p>The prediction of molecular properties using Artificial Intelligence (AI) is a transformative approach yet the effectiveness of such models depends heavily on the quality of the molecular representation. Deep learning architectures such as Graph Neural Networks (GNNs), combined with Self-Supervised Learning (SSL), have achieved considerable success, but they remain limited in their ability to capture long-range interactions and to integrate fundamental chemical information, such as atomic period, group, and orbital hybridization data. To address these limitations, this study proposes the Knowledge-Infused Transformer Encoder (KITE) model. KITE leverages the strengths of the Transformer architecture, which is well suited to processing long-range dependencies, and integrates fundamental chemical information directly into its input features. The methodology focuses on constructing the KITE model around the Bond-Edge Attention Model (BEAM) architecture to represent information from chemical bonds. The principal design 1 change lies in the atomic features: one-hot encoding is replaced by atom identification based on periodic-table period and group, while the hybridization state is represented by a five-component vector derived from Valence Shell Electron Pair Repulsion (VSEPR) theory. The model was pre-trained using a Knowledge-Guided Pre-training Strategy (KPS) on 965,812 molecules from ChEMBL29 and subsequently fine-tuned on eleven benchmark datasets, nine drawn from MoleculeNet together with the Estrogen and MetStab datasets. Performance was evaluated using the Area Under the Receiver Operating Characteristic Curve (ROC-AUC) and Root-Mean-Square Error (RMSE) metrics, together with a data contamination analysis based on canonical SMILES. On the eight classification tasks, KITE achieved an average ROC-AUC of 0.854, compared with 0.843 for Knowledge-guided Pre-training of Graph Transformer (KPGT), despite using a pre-training dataset less than half the size. On the three regression tasks, KITE achieved an average RMSE of 1.020, compared with 1.175 for KPGT. The overall exact-molecule contamination rate was 24.9% for KITE and 86.9% for KPGT. These results show that chemically informed atom and bond features can support competitive molecular property prediction with a comparatively small pre-training corpus. Further work should evaluate richer chemical features, alternative fine-tuning strategies, and practical virtual-screening applications with experimental validation.</jats:p>