應用G2DE分群法於分子嵌合能量回歸模型之研究

Abstract

研究蛋白質(protein)與配體(ligand)交互作用在基礎生物化學領域中相當重要。在這其中，一個很大的問題就是估計受體(receptor)與配體之間結合能(binding affinity)的評分函數。但如今分子嵌合模擬評分函數的部分，仍有很多的進步空間。在虛擬藥物篩選中，一個良好的評分函數，相當於在過程中扮演守門員的角色。 本篇論文利用G2DE分群法，使用六個特徵的特性，將851筆複合體的資料集分為若干群組。這六個特徵為凡德瓦力(Van der Waals force)、靜電力(Electrostatic interaction)、氫鍵(Hydrogen bond)、退溶(Desolvation)、配體可扭轉之自由度(number of torsion bonds of a ligand)以及AutoDock 3特有的特徵水分子與極性原子結合時其氫鍵之平均估計能量值(Ehbond)。 將主群組與例外者群組分開後，我們在例外者群組內進行分析討論，我們發現含有MHC_I功能區塊的複合體在預測嵌合能量值上偏差較大，單純去掉12條含有MHC_I的複合體之後，即可以讓RMSE(root-mean-squared-error)從2.12下降至2.046。 我們也針對主群組建立一回歸模型，可以讓資料集的RMSE降到2.006，這也是眾多評分函數努力的目標，而R2有超過0.49，換算成相關係數(correlation coefficient)則是超過0.7，這也是相當不錯的結果。 如結果所示，新的評分函數配合例外者群組的分析，可以提供未來生化分析時更多的線索。

Research on protein-ligand interactions is a crucial part in basic biochemistry field. In this field, one of the important issues is to estimate the binding affinity between receptors and ligands. However, there is still much room for improvement in design of scoring function. In virtual screening, a good scoring function is like a strict goal keeper. Our studies applying G2DE by using 6 features, which are Van der Waals force, Electrostatic interaction, Hydrogen bond, Desolvation, number of torsion bonds of a ligand and Ehbond, divided the 851 protein-ligand complexes dataset into several groups. After the dataset was separated into outliers group and main group, we discovered that there are 12 complexes contains MHC_I domain was far away from the actual binding energy. By eliminating the 12, the RMSE of the predicting binding energy of the dataset is dropped from 2.12 to 2.046. We also construct an empirical scoring function according to the main group. The RMSE of the predicting binding energy of the main group RMSE is 2.006, and the R2 is 0.49. Our paper shows the new scoring function and the outlier detection method by using G2DE, which can provide further clues in biochemistry analysis.