Computational Prediction of Dye Membrane Permeability Using Molecular Fingerprint Features

Abstract:
Membrane permeability is a critical property for biomedical dyes in live-cell imaging and diagnostics, yet its experimental evaluation is time-consuming and resource-intensive. Here, we present DyeMemPred, a computational framework for predicting dye permeability based on FP4 molecular fingerprints. A curated dataset of 196 dyes with experimentally validated permeability annotations was constructed through literature mining and partitioned into training and independent test sets. Through integrated feature selection using Information Gain, Lasso, Random Forest, and Recursive Feature Elimination, four key fingerprint bits were identified as strong predictors. Using these features, six machine learning models were developed and evaluated, with the Decision Tree and XGBoost classifiers achieving optimal performance (AUC > 0.93, accuracy > 89%). Structural interpretation revealed that these bits correspond to substructures known to promote membrane transport. DyeMemPred offers an efficient and interpretable tool for virtual screening of permeable dyes. All data and tool code are available at: https://github.com/Chengxinlin125/DyeMempred.git