可解釋性的圖神經網路之於預測 P-醣蛋白受體

Abstract

P-糖蛋白（P-gp），是ABC轉運蛋白家族的一員，存在於細胞膜上。它與各種外來物質結合，並主動將它們從細胞中排出，從而減少細胞對它們的吸收。與P-gp相互作用並隨後被排出的這些物質稱為P-gp受體。鑑於其在全身範圍內的廣泛分佈，包括血腦屏障，了解藥物是否屬於P-gp受體以及其穿越血腦屏障的能力對於藥物開發至關重要。 在我們的研究中，我們開發了一套強大的方法，利用各種圖神經網路（GNN）模型在預測P-gp受體上實現卓越的準確性，同時保持可解釋性。具體來說，我們探索了三種GNN架構：圖卷積神經網路（GCN）、AttentiveFP和基於AttentiveFP的集成模型。我們的重點是預測給定的藥物分子是否為P-gp受體。我們整理了一個包含1995個藥物分子的資料集，其中包括1202個P-gp受體和793個P-gp非受體，以9:1的比例分為訓練集和測試集。我們的方法優於傳統的機器學習模型，達到了優秀的0.848的ROC-AUC和0.815的準確度。利用積分梯度方法，我們提取了與P-gp受體相關的6個關鍵子結構，這與現有文獻發現一致。此外，我們的方法不僅提供出色的預測結果，還為藥物開發人員提供透明的見解，使其成為P-gp受體預測之外的藥物開發和優化的寶貴工具。

P-glycoprotein (P-gp), a member of the ABC transporter family, is located on cell membranes; it binds to various foreign substances and actively transports them out of cells, thereby reducing their absorption. These substances that interact with P-gp and are subsequently expelled are termed P-gp substrates. Given the extensive distribution of P-gp throughout the body, including in the blood‒brain barrier, determining whether a drug is a P-gp substrate and understanding its ability to cross the blood‒brain barrier are crucial steps in drug development. In our study, we developed a robust protocol leveraging various graph neural network (GNN) models to accurately predict P-gp substrates while maintaining interpretability. Specifically, we explored three GNN architectures: the graph convolutional neural network (GCN), AttentiveFP, and an ensemble model based on AttentiveFP. We focused on predicting whether a given drug molecule is a P-gp substrate. We curated a dataset comprising 1995 drug molecules, including 1202 P-gp substrates and 793 P-gp non-substrates, which was split into a training set and a testing set with a 9:1 ratio. Our approach outperformed traditional machine learning models, achieving an impressive receiver operating characteristic (ROC)-area under the curve (AUC) of 0.848 and an accuracy of 0.815. Using the integrated gradient method, we identified 6 critical substructures associated with P-gp substrates, consistent with previous studies findings. Our protocol achieved outstanding prediction results and can provide transparent insights for drug developers, making it a valuable tool for drug development and optimization beyond P-gp substrate prediction.