基於局部極值的圖卷積與列表損失之藥物與標靶的互 動預測

Abstract

預測藥物-靶標親和力 (DTA) 是藥物發現和設計的重要部分之一。 研究人員提出了預測 DTA 的計算方法，以規避更昂貴的體內和體外測試。 最新的方法採用深度網絡架構來獲取藥物分子和蛋白質一維序列的特徵。 在本論文中，我們驗證了二維藥物表示比一維表示包含更多信息，有助於更好地預測 DTA。 具體來說，藥物化合物可表示為圖，目標蛋白質表示為序列訊息。 我們開發了一種新的基於圖神經網絡的預測模型，稱為 LE-DTA，其性能優於目前文獻所提出的方法。 LE-DTA 應用局部極值卷積進行有效的特徵提取。 它側重於節點嵌入圖的局部和全局極值。 我們探討了所提出模型在三個不同基準數據集（即 Davis、KIBA 和 BindingDB）上的性能。 我們提出的模型改進了已知模型的一致性指數 (CI) 和均方誤差 (MSE)。 實驗結果顯示，我們所提出的 LE-DTA 在 Davis、KIBA 和 BindingDB 數據集上分別實現了 0.898、0.902、0.855 的 CI 與 0.210、0.120 和 0.464 的 MSE。 這些結果在 Davis 資料集上得出與已知模型相當的結果，但在 KIBA 資料集中，CI 提高了 1.12%， MSE 降低了 7.7%。 最後，在 BindingDB 資料集上，CI 比已知模型提高 0.35%，MSE 降低了 3.33%。 我們的模型顯示出令人滿意的預測準確性，並顯著提高了藥物發現過程的效率。

One of the essential parts of drug discovery and design is the prediction of drug-target affinity (DTA). Researchers have proposed computational approaches for predicting DTA to circumvent the more expensive in vivo and in vitro tests. More recent approaches employed deep network architectures to obtain the features from the drug molecules and protein 1D sequences. In this work, we demonstrated that 2D drug representation contains more information than 1D representation and helps predict DTA better. Specifically, the drug compounds are represented as graphs to extract this information. We developed a new graph-based prediction model, termed LE-DTA, that performed better than existing benchmark models. LE-DTA utilizes local extrema convolutions for effective feature extraction. It focuses on the local and global extrema of graphs for node embedding. We investigated the performances of the proposed model on three different benchmark datasets, i.e., Davis, KIBA, and BindingDB. Our proposed models have improved the Concordance Index (CI) and Mean Square Error (MSE) over existing benchmarks. Experiment results showed that the proposed LE-DTA achieved a CI of 0.898, 0.902, 0.855 and an MSE of 0.210, 0.120, and 0.464 on the Davis, KIBA, and BindingDB datasets, respectively. These results are in the range of the existing benchmarks for Davis, while it shows a 1.12% improvement in CI with a 7.7% reduction in MSE for KIBA. Finally, on BindingDB, the CI is 0.35% better than the baseline models, with an MSE reduction of 3.33%. Our models show satisfactory prediction accuracies and improve the efficiency of the drug discovery process.