Design of novel PI3Kα and PI3Kγ inhibitors for cancer treatment using pharmacophore, protein-ligand contacts, and machine learning methods.

Cancer is a complex disease characterised by the unregulated growth of abnormal cells. The intracellular signaling pathway, specifically phosphatidylinositol 3-kinase (PI3K)/AKT, is reported to be mutated in various cancers, including colorectal, gastric, and breast cancers. The pathway plays a crucial role in cancer cell survival and metastasis, making it an important therapeutic target for cancer treatment. Thus, targeting the key proteins of the PI3K signaling pathway, which are implicated in cancer, is necessary for the therapeutic intervention. In this endeavor, predictive machine learning (ML) models were employed to build PLIP and PRODIGY-derived molecular features-based classification and regression models on the 136 PI3Kα and PI3Kγ co-crystallised ligands from research collaboratory for structural bioinformatics (RCSB) protein data bank (PDB), along with RDKit-derived 1D and 2D molecular descriptors-based classification models. It was found that the four regression-based models (Linear regression, SMOreg, multilayer perceptron network (MLP), and Gaussian processes) were suitable for our dataset based on their higher predictive performance (Matthew's correlation coefficient of 0.9). Pharmacophore mapping, molecular docking-assisted structural analysis suggested certain criteria in the chemical compound, such as number of heavy atoms (> 25), number of rotatable bonds (> 4), molecular weight (> 400 Da), log P (> 2), to be favorable for better binding to the receptor. The role of non-bonding interactions measured with the number of atomic contacts within a 10.5 Å cutoff at the binding site of protein ligand complex, such as CC (> 2000), CO (> 800), CX (> 30), and the number of NN contacts (< 200), also favored the binding affinity of inhibitors.