血液脳関門透過性を予測するための包括的な機械学習フレームワーク：アリコロニー最適化による強化

The blood-brain barrier (BBB) severely restricts drug delivery to the central nervous system (CNS). Early, accurate in silico prediction of BBB permeability is a critical challenge in drug discovery. This study aimed to develop, validate, and optimize a robust machine learning framework for this task, with a specific focus on enhancing a Deep Neural Network (DNN) model using a bio-inspired Ant Colony Optimization (ACO) algorithm.

A dataset of 1957 compounds was curated and characterized by 212 molecular descriptors. The dataset was partitioned using a scaffold-based split. A suite of baseline models (Naive Bayes, k-NN, SVM, Random Forest, XGBoost, and a baseline DNN) was benchmarked. An ACO algorithm was then implemented to perform a systematic hyperparameter search for the DNN. The performance of all models was evaluated on an independent test set.

The Random Forest model provided a strong baseline performance (AUC = 0.913). The ACO algorithm identified a superior DNN architecture. The final ACO-Optimized DNN achieved a state-of-the-art performance on the independent test set, with an AUC-ROC of 0.921, outperforming all baseline classifiers. An error analysis of misclassified compounds provided insights into the model's limitations.

The successful performance gain highlights the advantage of metaheuristic optimization over traditional hyperparameter search approaches. The model demonstrated stability and strong generalization.

The ACO-Optimized DNN provides a highly accurate tool for virtual screening in CNS drug discovery, and the methodology serves as a generalizable template for complex model optimization.