基於分子圖與圖神經網絡於藥物與標靶之親和力預測

Abstract

藥物-靶標親和力（DTA），表示藥物與靶標之間的結合強度，對於藥物開發至關重要。準確預測DTA可以識別出適用於靶標蛋白的潛在藥物，加速了藥物開發過程。隨著深度學習的最新發展，基於深度學習的預測模型可以精確預測DTA。目前，大多數基於深度學習的DTA預測模型以1D蛋白質序列字符串作為模型輸入，這比圖所含的信息更少。然而，由於蛋白質複雜結構，蛋白質的接觸圖（contact map）難以獲得。鑒於蛋白質序列字符串提供的信息有限，我們提出了NG-DTA，將蛋白質序列轉化為分子子圖（molecular sub-graph），作為蛋白質的輸入並以圖神經網絡處理。實驗表明，在不同尖端深度學習的DTA預測模型中，NG-DTA在一致性指數（CI）和平均平方誤差（MSE）方面表現最佳（Davis數據集：CI為0.905，MSE為0.196；Kiba數據集：CI為0.904，MSE為0.120）。此外，我們應用此模型，對FDA批准的藥物與COVID-19和猴痘的重要蛋白質之間的DTA進行排名，以找出潛在藥物。我們的模型在準確性方面取得了令人滿意的結果，並改善藥物篩選過程。

Drug–target affinity (DTA), which indicates the binding strength between a drug and a target, is essential to drug development. An accurate prediction of DTA can identify the potential drugs for target proteins, speeding up the drug development process. With recent developments in deep learning, deep-learning-based prediction models can precisely predict DTA. Currently, most deep-learning-based DTA prediction models take 1D protein sequences string as model input, which is less informative than the graph representation. However, due to the difficulties of determining the protein structure process, the contact maps of proteins are not always available. In view of the limited information provided by the protein sequence string, we proposed NG-DTA which converted the protein sequence to molecular sub-graphs which are processed by the graph neural networks as input of the protein. Experiment shows that NG-DTA performs the best among different baseline deep-learning-based DTA prediction models in terms of concordance index (CI) and mean square error (MSE) (CI: 0.905, MSE: 0.196 for the Davis dataset; CI: 0.904, MSE: 0.120 for Kiba dataset). Furthermore, we deployed our model in ranking the DTA between the FDA-approved drugs and important COVID-19 and Monkeypox proteins to uncover potential drugs to combat these diseases. Our model has satisfactory accuracy and improves the drug screening process.