Gia-Bao Truong

Molecular property prediction has become essential in accelerating advancements in drug discovery and materials science. Graph Neural Networks have recently demonstrated remarkable success in molecular representation learning; however, their broader adoption is impeded by two significant challenges: (1) data scarcity and constrained model generalization due to the expensive and timeconsuming task of acquiring labeled data, and (2) inadequate initial node and edge features that fail to incorporate comprehensive chemical domain knowledge, notably orbital information. To address these limitations, we introduce a Knowledge-Guided Graph (KGG) framework employing self-supervised learning to pre-train models using orbital-level features in order to mitigate reliance on extensive labeled datasets. In addition, we propose novel representations for atomic hybridization and bond types that explicitly consider orbital engagement. Our pre-training strategy is cost-efficient, utilizing approximately 250,000 molecules from the ZINC15 dataset, in contrast to contemporary approaches that typically require between two and ten million molecules, consequently reducing the risk of potential data contamination. Extensive evaluations on diverse downstream molecular property datasets demonstrate that our method significantly outperforms state-of-the-art baselines. Complementary analyses, including t-SNE visualizations and comparisons with traditional molecular fingerprints, further validate the effectiveness and robustness of our proposed KGG approach.

In the development of anticancer medications, vascular endothelial growth factor receptor 2 (VEGFR-2), which belongs to the protein tyrosine kinase family, emerges as one of the most significant targets of interest. The ongoing Food and Drug Administration (FDA) approval of novel therapeutic medicines toward VEGFR-2 emphasizes the urgent need to discover sophisticated molecular structures that are capable of reliably limiting VEGFR-2 activity. Recognizing the huge potential of deep-learning-based molecular model advancements, we focused our study on exploring the chemical space to find small molecules potentially inhibiting VEGFR-2. To achieve this goal, we utilized the junction tree variational autoencoder in combination with two optimization approaches on the latent space: the local Bayesian optimization on the initial data set and the gradient ascent on nine FDA-approved drugs targeting VEGFR-2. The optimization results yielded a set of 493 uncharted small molecules. Quantitative structure–activity relationship (QSAR) models and molecular docking were used to assess the generated molecules for their inhibitory potential using their predicted pIC50 and binding affinity. The QSAR model constructed on RDK7 fingerprints using the CatBoost algorithm achieved remarkable coefficients of determination (R2) of 0.792 ± 0.075 and 0.859 with respect to internal and external validation. Molecular docking was implemented using the 4ASD complex with optimistic retrospective control results (the ROC-AUC value was 0.710 and the binding activity threshold was −7.90 kcal/mol). Newly generated molecules possessing acceptable results corresponding to both assessments were shortlisted and checked for interactions with the protein at the binding site on important residues, including Cys919, Asp1046, and Glu885.