Machine learning-based repurposing of Asciminib analogues and FDA approved drugs against CML cell-lines

Background and Aims: Chronic myeloid leukemia (CML) is a type of blood cancer driven by BCR-ABL fusion protein. This abnormal protein results from chromosomal translocation leading to unregulated tyrosine kinase (TK) activity. Despite satisfying clinical outcomes, Asciminib, a later-generation TK inhibitor, faces challenges due to emerging allosteric resistance. This issue necessitates the query for alternative compounds. Within present contribution, various machine learning (ML) algorithms were utilized to evaluate the repurposing potential of Asciminib analogues and FDA approved drugs as growth inhibitors of CML-related cancerous cell lines. Methods: A dataset comprising ۲۳۳ small molecules with relevant cell-based (K۵۶۲, BV۱۷۳) IC۵۰ values was curated from ChEMBL database (https://www.ebi.ac.uk/chembl/). Morgan fingerprint, atom graphs, pharmacophore Graph (one-hot-encoded pharmacophore as node, bonds type as edge), junction tree graph (generated by converting bonds, rings and junction atoms to nodes) and functional group graph (predefined groups as nodes, and predefined bond as edge) were calculated for each molecule under study as a feature. AutoDock-GPU was used to calculate mean-binding free energies for each molecule in the allosteric site of wild and mutant BCR-ABL. We designed a graph convolutional neural network to learn from paired data and choose active one subsequently. Other promising ML algorithms such as XGBoost and Random Forest were also tested. Subsequently, the trained models were used to predict the repurposing potential of Asciminib like molecules and FDA drugs. Results: XGBoost was top-ranked regression-task model (R۲ = ۰.۷) whereas Graph convolutional neural network was the best classification-task model (۸۰% for test sensitivity and ۷۰% for ROC-AUC value). Top-ranked Asciminib like molecules and FDA drugs were found to be Venetoclax, CHEMBL۴۱۱۱۰۰۳, Simeprevir, Lurasidone and CHEMBL۳۹۷۳۵۱۸. It is worth noting that Venetoclax has already been studied for advanced-phase CML. Moreover, CHEMBL۴۱۱۱۰۰۳ and CHEMBL۳۹۷۳۵۱۸ were subjected to in-silico pharmacokinetics analysis in terms of bioavailability and metabolism. Conclusion: Overall ML models showed promising results only when fed with proper data and subsequently hypertuned. On the basis of outlined computational workflow, diversifying the candidate molecules under study could suggest more potent and mutation-resistant CML candidates. Further analysis of model parameters within larger

کلیدواژه ها:

Chronic myeloid leukemia ، Asciminib ، machine learning ، FDA approved drugs ، CML cell lines