Harnessing Machine Learning for the Virtual Screening of Natural Compounds as Both EGFR and HER2 Inhibitors in Colorectal Cancer: A Novel Therapeutic Approach.

Colorectal cancer (CRC) is a type of cancer that affects the colon and rectum, with overexpression of epidermal growth factor receptor (EGFR) and human epidermal growth factor receptor 2 (HER2) observed in up to 85% of colorectal cancer cases. Although CRC treatment has progressed with the introduction of targeted drugs, current approaches have primarily focused on a single blockage of EGFR or HER2 to combat colon cancer. However, monotherapies that target either the EGFR or HER2 receptor frequently have low efficacy due to mutations in downstream effectors such as Kirsten Rat Sarcoma 2 Viral Oncogene Homologue (KRAS) and the activation of compensatory signaling pathways that support tumor survival and proliferation. Hence, the discovery and development of a therapy with the capability to combat CRC by simultaneously inhibiting both EGFR and HER2 remain avenues for further exploration. This study introduces a novel machine learning (ML)-based stacking ensemble framework for rapidly and accurately identifying dual EGFR and HER2 inhibitors using SMILES notation. A benchmark data set comprising active and inactive compounds against EGFR and HER2 was curated from the ChEMBL database. Based on this data set, 40 baseline models were developed and optimized using a comprehensive set of well-known molecular descriptors and ML algorithms (Figure 1). These models generated probabilistic features integrated via a logistic regression model as a final estimate to construct the final stacking ensemble model. The model was applied to bioactive compounds identified through LC-MS/MS profiling of Ceratonia siliqua extract as well as to a subset of natural products from the LOTUS database for virtual screening. The cytotoxic potential of Ceratonia siliqua was experimentally validated using the MTT assay against HCT116 colorectal cancer cells and noncancerous Vero cells, where the extract exhibited an IC50 value of 13.32 ± 1.09 μg/mL against HCT116 cells, indicating its significant anticancer potential. Additionally, molecular docking and in silico ADMET studies were conducted on the top three compounds from both the LOTUS database and the predicted candidate from the LC-MS/MS data set identified by the stacking model, alongside four FDA-approved anticancer drugs for comparative analysis. Among these, LTS0131923 demonstrated the highest binding affinity against HER2 (PDB ID: 7MN5), with a binding energy of -11.2 kcal/mol and an inhibition constant of 0.00626 μM, outperforming Tucatinib, a standard CRC treatment. This study reveals the potential of ML-driven approaches to accelerate the discovery of dual-target inhibitors for CRC therapy and highlights Ceratonia siliqua L. as a promising source of bioactive compounds for cancer treatment.