A quantitative study of cytotoxic compounds using graph based descriptors and machine learning.

Area of Science:

• Computational Chemistry
• Medicinal Chemistry
• Drug Discovery

Background:

• Cytotoxic drugs are vital for cancer treatment but have narrow therapeutic indices and significant side effects.
• Understanding physicochemical properties, such as Topological Polar Surface Area (Top_PSA), is essential for predicting drug absorption, distribution, and permeability.
• Top_PSA is a key indicator for membrane transport, passive diffusion, and blood-brain barrier penetration.

Purpose of the Study:

• To evaluate graph-theoretical and molecular descriptors as predictors of RDKit/Mordred-calculated Top_PSA values for diverse cytotoxic agents.
• To compare different data preprocessing and feature selection strategies for Quantitative Structure-Activity Relationship (QSAR) modeling.
• To identify the most effective modeling approaches for accurate Top_PSA prediction.

Main Methods:

• Calculated 58 molecular descriptors for 156 structure-diverse cytotoxic agents.
• Applied five preprocessing schemes: direct fitting, PCA, robust scaling, outlier identification/elimination, and VIF-based feature selection.
• Utilized linear, LASSO, and ridge regression models with k-fold cross-validation.
• Analyzed LASSO coefficients to identify significant contributing factors to Top_PSA.

Main Results:

• Robust scaling combined with LASSO regression yielded the best predictive performance, demonstrating effectiveness in handling heteroscedasticity and multicollinearity.
• Principal Component Analysis (PCA) offered similar predictive accuracy but reduced interpretability.
• VIF-based pruning was less effective than other methods.
• Significant factors influencing Top_PSA included heteroatom content, hydrogen-bonding capacity, and electronegativity-weighted indices.

Conclusions:

• Robust preprocessing and sparsity-inducing regularization (LASSO) are highly recommended for QSAR studies involving numerous descriptors.
• Clear computational pipelines are vital for reliable Top_PSA predictions.
• The study provides practical guidance on preprocessing, feature selection, and model selection in QSAR workflows.