De Novo Design and Bioactivity Prediction of Mitotic Kinesin Eg5 Inhibitors Using MPNN and LSTM-Based Transfer Learning

Abstract

Breast cancer, the most commonly diagnosed disease worldwide, has been linked to the overexpression of the kinesin Eg5 protein, a spindle motor protein crucial for the assembly and maintenance of the bipolar spindle during mitosis. This makes Eg5 an attractive therapeutic target for tumor treatment. To address the urgent need for effective treatments for this lifethreatening illness, we utilized generative AI to design novel and potential inhibitors of this protein. In this study, a generative LSTM model was pretrained on SMILES data from ChEMBL and subsequently fine-tuned using SMILES of compounds with reported activity against the Eg5 protein. The fine-tuned model generated valid compounds, which were screened using a machine learning model, drug-likeness filters, molecular docking, and molecular dynamics (MD) simulations conducted over 200 ns. Five novel compounds with better binding affinities to Eg5 compared to the co-crystallized ligand were identified. The top compound, Compound 103 (a bioisostere of the co-crystallized ligand), demonstrated a significantly improved binding free energy (-82.68 kcal/mol) compared to the co-crystallized ligand (-76.98 kcal/mol), as determined by MM-GBSA calculations. ADMET predictions and MD simulations further confirmed that the top compounds interacted effectively with the target protein and exhibited drug-like properties. This study shows the potential of generative AI to explore our vast chemical space and find promising drug candidates. However, further in vitro and in vivo studies are needed to confirm the predicted biological effects of the top compounds.