Integrated Machine Learning and Molecular Modelling Strategies to Probe Anti-inflammatory Response of New Chemical Entities

Abstract

5-lipoxygenase activating protein (FLAP) inhibitors have proven to attenuate 5-LO pathway activity and high leukotriene (LT) production that lead to chronic inflammation which is the root cause behind number of pathological conditions and even cancer. Mainly, FLAP inhibitors have been studied by structure activity relationship (SAR) studies presenting structurally diverse compounds but neither of them could reach the market because of poor pharmacokinetics and toxicity. Therefore, FLAP inhibitors are highly desirable in order to relieve patients suffering from inflammatory disorders and looking forward to receive anti-LT therapy. Because of structural diversity of data, there is not a single universal predictive model available that can vaticinate the anti-inflammatory properties of inhibitors targeting FLAP. To accomplish this goal, we utilized advance machine learning techniques and classical modelling strategies for prediction of important anti-inflammatory properties of FLAP inhibitors. For this purpose, neural network model with classification accuracy of 96% has been developed by using 2D descriptors of structurally diverse compounds active against FLAP followed by validation with two test sets. As in biological systems, molecules interact in 3D manner, therefore molecular docking studies were performed to probe the molecular basis of interaction of diverse anti inflammatory compounds with FLAP. Furthermore, common scaffold clustering based on 8 structurally diverse classes of compounds revealed the distinct binding pattern of each class. By using optimal poses from common scaffolds, we developed a 3D predictive GRIND model that displayed R2 and q2 values of 0.70 and 0.66 respectively. Moreover, GRIND indicated the importance of two hydrogen bond donor, two hydrophobic, and two shape-based features for inhibition of FLAP. Additionally, projection of the FLAP binding pocket onto the identified features complements the presence of corresponding amino acids including Lys116, Arg117 as hydrogen bond acceptors while Val61 and Ala63 as hydrophobic. At the end, GRIND model was also subjected to validation by same two test sets. In a nutshell, we were able to develop a predictive neural network model using 2D descriptors and a GRIND model using 3D molecular fields followed by subsequent external validation. In future, these models can be used for the selection of potential FLAP inhibitors by screening of compounds based on 2D and 3D properties.