Multi-representation machine learning approaches for screening aurora kinases A & B inhibitors and insights from...

The Aurora kinase (AURK) family enzymes are almost identical to the serine or threonine kinase (STK), and they are essential for various cell functions, including cell cycle to cell division. Aurora kinases A (AURKA) and B (AURKB) play a vital role, and their overexpression is linked to the aetiology of different carcinomas; thus, they are attractive targets for cancer treatment. However, the developed therapies for cancer treatment are suffering from many severe drawbacks, including non-selectivity, resistance, and lethal side effects. Numerous clinical and preclinical studies aim to identify selective AURKA & AURKB inhibitors. In this study, a systematic filtration procedure was conducted with ARUKA (4268) and AURKB (3782) inhibitors, and a series of machine learning models with descriptors, fingerprints, and atom-level features was developed. The models were analysed with different statistical parameters and the best model was identified in each set of descriptors. Bayesian statistics was applied to the same training/test set to identify high-performing substructures recursively, thereby improving the model's explainability. The applicability of the developed model was evaluated by screening with a specific AURKA & AURKB clinical trial compounds. Structural alerts help identify the specific scaffolds within each set of inhibitors that may be responsible for kinase activity. The outcome of this study helps in designing more selective AURKA and AURKB inhibitors. The source code and dataset are available at https://github.com/NagaLab-MSN/AURK.