以多任務學習預測化學物於活體外實驗模式之肺吸收

Abstract

肺吸收不管作為藥物傳遞或是潛在的毒物曝露都是很重要的途徑。而為了精簡藥物或化學物評估流程並順應減少動物實驗的趨勢，吸入性生物藥物學研究除了傳統的動物實驗之外，有越來越多不同的實驗模式系統被發展出來，包括體外實驗 (in vitro)、活體外實驗 (ex vivo) 和電腦模擬實驗 (in silico)。以細胞為實驗模式的體外實驗，其中Calu-3和Caco-2細胞穿透模型已常在肺穿透或肺吸收相關研究中被使用。活體外實驗模式則是以大鼠分離灌流肺 (isolated perfused lung, IPL) 模型研究為主。本研究以蒐集的IPL肺吸收速率常數 (absorption rate constant, ka) 資料、Caco-2細胞穿透係數 (apparent permeability coefficient, Papp) 資料和Calu-3細胞穿透係數資料來建構機器學習模型，利用三種不同實驗模式訓練資料建立多任務極限樹 (ExtraTrees) 模型。由實驗結果得知，與傳統單任務學習模型相比，訊練過程中加入了Caco-2和Calu-3相關任務的訓練資料，在遷移學習過程中獲得的資訊確實可以進一步幫助IPL吸收速率常數的預測。模型在特徵選擇中找出了7個重要的物化特性敘述子。獨立測試結果顯示模型表現良好，預測值和觀察值的皮爾森相關係數 (Pearson correlation coefficient) 達到0.84，為高度相關。此外，此機器學習模型也與公開文獻發表過現有的偏最小平方法 (partial least squares, PLS) 統計迴歸模型相比較並有更好的表現、易用性以及泛化能力。最後，本研究也使用此模型預測美國FDA核准之吸入性藥物進行個案研究，預測結果顯示用於局部治療藥物的長效型β2致效劑吸收速率較短效型慢；用在系統性全身麻醉的氣體麻醉劑吸收速率明顯較快，和臨床用途相符，進一步驗證此模型之準確性，第二個個案研究以本模型預測呼吸致敏物吸收速率，預測結果顯示呼吸致敏物於組織停留時間較久，此結果差異提供了快速篩選致敏物的可能性，後續可進一步驗證呼吸致敏物於組織穿透的差異性。本研究之模型架構將可隨著日後更多資料的產生與整合，可望作為藥物開發前期篩選吸入型藥物或評估潛在毒性化學物肺吸收的預測工具。

Pulmonary absorption is an important route for both drug delivery and toxic chemical exposure. Pulmonary inhalation biopharmaceutical research can be conducted with different model systems, including in vitro, ex vivo and in silico models. To streamline the drug or chemicals assessment process and to reduce animal experiments, various model systems thus have been developed. For example, Calu-3 and Caco-2 permeability cell models have been used in pulmonary permeability or absorption studies. Isolated perfused lung (IPL) ex vivo models which is usually conducted in rats can be used for measurements of pulmonary absorption. In this study, we developed a machine learning model that employed a tree-based multitask learning approach to predict IPL absorption rate constant (kaIPL) of various chemicals, combining the knowledge obtained from relevant tasks of Caco-2 and Calu-3 permeability. The results showed that the multitask learning architecture improved the prediction accuracy compared with single task learning. After feature selection, seven physicochemical descriptors were used to construct the final prediction model. The results show good performance of multitask learning with a high correlation between predictions and observations of logkaIPL (r = 0.84) in the independent test dataset. In addition, results show that our model has better performance, generalization ability and usability when compared with the existing partial least squares (PLS) model. This study also utilized this model to predict the absorption rates of FDA approved inhalation drugs and respiratory sensitizers for two case studies. The results show that the absorption rate of long-acting β2 agonists are slower than that of short-acting agonists, while the absorption rates of inhalational anesthetics are obviously faster, which are consistent with the clinical use and desirable property in general anesthesia. In the second case study, the prediction results show that there is an overall longer absorption time for sensitizers. The difference of the distributions can make the model be a potential screening tool to distinguish the relatively high-risk respiratory sensitizers. These results further verify the accuracy of the model and explore potential applications. With more available data and integration, the proposed model architecture may serve as a valuable tool to screen drug or chemical absorption of lung in the early stage drug development and toxic chemical assessment.