Machine learning-based dual-target prediction models for EGFR/ROS1 and PDGFR/c-KIT inhibitors

The EGFR, ROS1, PDGFR, and c-KIT are crucial targets for anti-cancer therapy. However, no models for predicting the bioactivity of dual-target inhibitors against the EGFR/ROS1 and PDGFR/c-KIT targets have been reported. The present study aims to develop such models. Method: Predictive models were developed using machine learning algorithms, including Adaptive Boosting (AdaBoost) and Random Forest (RF). Results: The RF model achieved AUCs of 0.99 for PDGFR/C-KIT and 0.97 for EGFR/ROS1 targets, respectively. Overall, the RF and KNN models were identified as better for predicting the bioactivity of these dual-target inhibitors. We also found that the importance of fluorine in the inhibitors decreased in the following order: BRAF > EGFR/ROS1 > PDGFR/c-KIT. In molecular docking, compound LPC-002 showed high binding affinity with S scores of -9.84 for c-KIT and -7.14 for PDGFR. Similarly, compound LER-003 had S scores of -6.79 for EGFR and -7.52 for ROS1. The study is first one to develop dual-target machine learning (ML)models for EGFR/ROS1 and PDGFR/c-KIT using substructure-based descriptors. Our findings reveal a key shift in elemental dependence: the role of fluorine weakens, while that of chlorine strengthens, when moving from BRAF single-target to PDGFR/c-KIT dual-target inhibitors. Our study indicates the structure 'm’ is a critical substructure for enhancing bioactivity of PDGFR/C-KIT inhibitors. For these inhibitors, the optimal number of 5- and 4-rings in the structure is 0 to 2. In addition, molecular docking identified that the O = C groups are involved in forming hydrogen bonds, which may explain the importance of this group in EGFR/ROS1 inhibitors.