Graph neural network-guided identification and biological evaluation of potential AKT1 inhibitors for triple-negative...

Triple-negative breast cancer (TNBC) presents a significant therapeutic challenge due to its aggressive behavior and lack of targeted therapies. The PI3K/AKT/mTOR signaling pathway, particularly AKT1, is frequently dysregulated in TNBC, driving disease progression. Despite extensive research, many clinically evaluated AKT1 inhibitors have encountered challenges related to both efficacy and tolerability, highlighting the need for novel therapeutics. Here, we employed graph neural networks (GNNs) for molecular graph-based prediction of potential AKT1 inhibitors. Six GNN architectures, including attention-based (AttentiveFP, GATv2Conv, TransformerConv) and non-attention-based (GCNConv, GINConv, GraphSAGE) models were trained and benchmarked against traditional machine learning (ML) methods using random and scaffold-based data splits. To enhance predictive relevance and model generalizability, we integrated phenotypic screening data from breast cancer (BC) cell lines alongside AKT1 bioassay data to capture broader pathway effects. Screening the Maybridge chemical library, we identified 9 novel scaffold compounds through consensus hit selection, molecular docking, and novelty filtration. Enzymatic validation confirmed 4 early-stage AKT1 inhibitors with low-micromolar potency (IC₅₀ down to 2.5 μM). Explainable AI analyses using Integrated Gradients and Captum saliency maps highlighted key structural features driving AKT1 inhibition, providing interpretable structure-activity relationship (SAR) insights. Scaffold diversity analysis further confirmed that the validated hits occupy chemical space distinct from known AKT1 inhibitors. Overall, this study presents an interpretable AI-driven discovery framework that identifies novel AKT1 inhibitor scaffolds and provides a validated starting point for hit-to-lead optimization in TNBC drug discovery.