XGBPred-ACSM: A Hybrid Descriptor-Driven XGBoost Framework for Anticancer Small Molecule Prediction

Area of Science:

• Computational oncology
• Drug discovery and design
• Machine learning applications in medicine

Background:

• Cancer poses a significant global health challenge due to limitations in current therapeutic specificity and toxicity.
• Developing more effective anticancer drugs requires advanced predictive modeling approaches.
• Machine learning offers a powerful tool for enhancing the efficiency of anticancer drug discovery.

Purpose of the Study:

• To develop and validate an advanced machine learning-based architecture for predicting anticancer small molecules.
• To systematically process molecular representations and benchmark various machine learning algorithms.
• To identify the most influential molecular determinants for anticancer activity.

Main Methods:

• Utilized 3600 compounds with experimentally validated IC50 values.
• Derived molecular representations using 2D physicochemical descriptors, structural fingerprints, and hybrid sets (Mordred, PaDEL).
• Trained and evaluated six machine learning algorithms (RF, XGB, GB, ET, AdaBoost, LightGBM) using 10-fold cross-validation and SHAP-based feature analysis.

Main Results:

• The XGB-Hybrid architecture demonstrated superior performance with an AUC of 0.88 and 79.11% accuracy on an independent test set.
• SHAP analysis provided mechanistic insights by quantifying feature contributions and identifying key molecular determinants.
• The developed framework is technically rigorous, interpretable, and high-performing for early-stage anticancer small molecule identification.

Conclusions:

• Machine learning, particularly the XGB-Hybrid framework, significantly advances computational oncology and rational drug design.
• This approach accelerates the identification of potential anticancer small molecules.
• The study highlights the transformative impact of machine learning in modern drug discovery pipelines.