Abstract
Optimizing the prediction of anti-colorectal cancer agent activity is essential aspect in the identification and creation of medications. Machine learning (ML) techniques, which have gained widespread adoption in computational chemistry, offer a rapid and reliable approach for evaluating the relationship between molecular structures and bioactivity. In this paper, a comprehensive dataset of molecular descriptors and quantum chemical properties was compiled, encompassing general molecular properties, electronic and quantum characteristics, aromatic ring structure, halogen effects, functional groups, specific structural features, and molecular charge characteristics. This dataset enhances the adaptability of data-driven models and mitigates the risk of overfitting. Seven tree-based ML algorithms, including Gradient Boosting, Random Forest, Decision Tree, Light gradient boosting (LightGBM), Categorical boosting (CatBoost), Extreme gradient boosting (XGBoost), and Extra Trees, were utilized to forecast the bioactivity of candidate compounds against colon cancer cell lines. Key molecular predictors were analyzed, and interaction terms between predictors were incorporated to improve model accuracy. The study utilizes the Tree-Structured Parzen Estimator for fine-tuning hyperparameters to enhance model efficiency and predictive accuracy. Additionally, k-fold cross-validation is utilized to avoid overfitting and guarantee a strong model evaluation and adaptability. These approaches enhance the dependability and effectiveness of data-driven models. The findings revealed that all models exhibited exceptional performance, with Extra Trees emerging as the top-performing algorithm due to its swift optimization process and superior performance in F1-Score and Recall metrics. These results highlight the potential of ML-driven methods to significantly enhance the prediction of anti-colorectal cancer agent activity by optimizing predictor selection based on quantum chemical properties and molecular interactions. This research offers novel perspectives on leveraging ML for quantitative structure-activity relationship (QSAR) modeling in drug discovery. By addressing challenges such as scarce labeled data and data gaps, and conducting an in-depth analysis of multiple ML algorithms, our study provides vital insights for computational chemists and pharmaceutical researchers, aiding them in selecting the most suitable algorithms for QSAR-based drug design. Ultimately, this work contributes to the advancement of anti-colorectal cancer drug discovery, enabling more efficient and sustainable drug development practices.
