4D-QSAR分析TW01系列類似物之定量構效關係

Abstract

TW01類似物的作用機轉之一是為蛋白質酪氨酸激酶抑制劑(protein tyrosine kinase inhibitors)，本論文即是利用四維定量構效關係(4D-QSAR)法分析TW01類似物與四個人類癌症細胞株MDA-MB-231、PC-3、Hep3B和HUVEC的關係，建構TW01類似物結構與對癌症細胞株測得之-logIC50值的4D-QSAR數學模型。四組癌症細胞株模型訓練資料集總共合有42個TW01類似物，其中MDA-MB-231與PC-3訓練集各含有24個化合物，Hep3B訓練集含有8個化合物以及HUVEC訓練集含有13個化合物。 數個獨特的4D-QSAR模型從四個訓練集中訓練得來，其中MDA-MB-231模型的Q2值在0.716到0.795間，PC-3模型的Q2值在0.827到0.844間，Hep3B模型的Q2值在0.839到0.884間以及HUVEC模型的Q2值在0.742 到0.812間，四組模型統計意義皆有Q2值大於0.7的顯著性。此外Chauvenet準則亦用來識別在資料集中可能是離群值的化合物，MDA-MB-231與PC-3訓練集在幾個離群值的化合物移除後，訓練出了幾個更有統計意義的4D-QSAR模型，其中MDA-MB-231模型Q2值為0.824而PC-3模型Q2值在0.839到0.882間。而對於MDA-MB-231與PC-3訓練集，亦有用MIT-GA基因演算法套件來取代原本4D-QSAR法中訓練模型之WOLF套件，藉MIT-GA訓練模型後，獲得之最佳MDA-MB-231模型Q2值為0.807而PC-3模型Q2值為0.772。 以PC-3訓練集建構出之模型，亦用來預測未在訓練集裡的TW01類似物，在結果上，一個模型對有觀測值的分子做的預估，達到觀測與預估-logIC50值之間的誤差絕對值小於0.5。另外，亦有新的TW01類以物加入PC-3訓練集，在移除離群集後，建構出一個R2值為0.811而Q2值為0.734之具有顯著性的模型。 本論文建構出之4D-QSAR模型中普遍地指出TW01類似物附近重要的空間與疏水的藥效基團(pharmacophore)區域，這與McGregor所建構之蛋白質激酶上ATP結合區域的藥效基團地圖是相符合的，因此TW01類似物的標靶(target)可能是蛋白質激酶的ATP結合區域。這些建構在不同癌症細胞株的4D-QSAR模型能回饋給TW01團隊設計TW01類似物，另一方面也可以用來做電腦虛擬高通量篩選(virtually high throughput screening)過濾化合物，以期找到更有效的TW01類似物。

Receptor-independent 4D-quantitative structure-activity relationships analyses were carried out on TW01 analogues, the possible tyrosine kinase inhibitors, to construct 4D-QSAR models for four human cancer cell lines, MDA-MB-231, PC-3, Hep3B and HUVEC, for which the -logIC50 values were measured. Total of 42 TW01 analogues were included in the training sets. It is divided into four subsets, 24 compounds for MDA-MB-231 and PC-3, eight compounds for Hep3B and 13 compounds for HUVEC. Distinct 4D-QSAR models were identified for MDA-MB-231 (three models with Q2 value of 0.716 to 0.795), PC-3 (three models with Q2 value of 0.827 to 0.844), Hep3B (two models with Q2 value of 0.839 and 0.884) and HUVEC training sets (three models with Q2 value of 0.742 to 0.812). Four sets of models yielded statistical significance (Q2 value larger than 0.7). After outlier removed from training set with Chauvenet's criterion, more statistically significant models were available for MDA-MB-231 (one model with Q2 value of 0.824) and PC-3 training sets (four models with Q2 value of 0.839 to 0.882). Two methods of genetic algorithm (WOLF, and MIT-GA) for model construction were used and compared in 4D-QSAR analysis. MIT-GA package was alternatively applied for MDA-MB-231 as well as PC-3 training sets and one 4D-QSAR model was identified for either of both two training sets (Q2 value of 0.807 for MDA-MB-231 and 0.772 for PC-3). The models constructed using PC-3 training set were also utilized to predict TW01 analogues not in the training set. Consequently, a constructed model yielded absolute value of residuals between observed and predicted -logIC50 values less than 0.5 for the compounds already assayed. Furthermore, PC-3 training set was extended by adding more TW01 analogues and yielded a significant model with R2 value of 0.811 and Q2 value of 0.734 after outliers were removed from the extended PC-3 training set. In the optimized 4DQSAR models, steric and hydrophobic sites were generally embedded in these models. Compared our 4D-QSAR models with McGregor’s work on pharmacophore analysis with series of ATP binding site of tyrosince kinases, the possible targets of TW01 analogues seem to fit into the ATP binding sites of protein kinases. The distinct optimized 4DQSAR models for each cancer cell line are feedback to TW01 synthetic team for designing and can be used for virtual high throughput screening for more potent TW01 analogues.