利用機器學習比較分子指紋和描述符對於粒線體毒性之可解釋性預測

Abstract

毒性預測是藥物開發上的一個重要關鍵，有助於科學家更精確的分析化合物潛在毒性，且能省去在體外生物實驗花費過多時間，近期的研究指出粒線體在藥物的毒性上有著至關重要的影響，粒線體是細胞內的能量中心，也是自由基產生的地點，而多種疾病的發生被認為與粒線體受損有相關性，一些藥物也因為其對粒線體的毒性而退出市場，如曲格列酮等。因此在研究哪種藥物對於粒線體可能會在產生毒性方面，藥物開發上已變成一個關鍵的試驗終點，傳統上透過臨床試驗來取得化合物對於粒線體的毒性，但臨床試驗通常需要投入大量時間和龐大的資源方能完成，並且亦有倫理道德等問題。在體外試驗上則可能有細胞異質性的潛在問題，因此我們決定採用定量結構-性質關係（QSAR）在這項研究中，來利用機器學習預測未知化合物的毒性，也避免了以上幾種方式帶來的潛在問題。我們的資料來源於Tox21，PubChem以及許多文獻，使用相同的標準之粒線體膜電位試驗資料，並且利用分別利用了化合物的分子指紋以及描述符來分析毒性對化合物的化學物理性質之關係，也分析毒性和結構的關聯，在結構分析上使用了Tanimoto演算法和PCA分析毒性分佈，在可解釋性分析中，我們對分子指紋的化合物結構警示進行分析，來判斷可能造成毒性的子結構，也利用機器學習對分子描述符進行解釋性分析，由於粒線體膜電位試驗之複雜性導致於資料的不平衡，我們也透過多種方法改變陰性陽性資料比例來增加模型之表現，包含比較了過擬和方法、重組化合物以及產生異構物等方法，我們利用不同的機器和深度學習模型來評估粒線體毒性，包含了SVM，隨機森林模型等等，最後透過模型之特徵重要性找出了影響毒性分析之性質，以及探索了粒線體毒性之結構警示，這都有助於科學家在藥物開發上更好的開發的對於人體更安全的藥物。

Toxicity prediction constitutes a pivotal phase in drug development, enabling scientists to rapidly and accurately assess the potential toxicity of compounds. Recent research highlights the pivotal role of mitochondria in toxicity, as they serve as cellular energy centers and sites of free radical generation. Damaged mitochondria are associated with various diseases, and certain pharmaceuticals have been withdrawn from the market owing to their adverse impact on mitochondrial function, such as rosiglitazone. Consequently, investigating the potential mitochondrial toxicity of drugs has become a crucial endpoint in drug development. Traditionally, assessing compound toxicity toward mitochondria involves time-consuming and resource-intensive clinical trials, fraught with ethical concerns. In vitro tests may present issues related to cellular heterogeneity. Thus, this study employs Quantitative Structure-Activity Relationship (QSAR) methodologies, utilizing statistical methods such as machine learning to forecast the unknown compounds’ toxicity, thereby circumventing potential limitations posed by other approaches. Data for this research originates from Tox21[1], PubChem, and various literature sources, utilizing standardized mitochondrial membrane potential assay data. Molecular fingerprints and descriptors are utilized to analyze the relationship between toxicity and structural or physicochemical properties. The structural analysis incorporates the Tanimoto algorithm and Principal Component Analysis (PCA) to scrutinize toxicity distribution, while interpretability analysis explores compound structure alerts within molecular fingerprints to identify potential substructures causing toxicity. Machine learning is also employed for interpretable analysis of molecular descriptors. Given the complexity of mitochondrial membrane potential assays leading to imbalanced data, various algorithms are employed to adjust the ratio of negative to positive data, encompassing methods like oversampling and compound recombination. This study evaluates mitochondrial toxicity using six distinct machine and deep learning models, including SVM and random forest models. Ultimately, by assessing the feature importance of these models, key properties influencing toxicity analysis are identified, along with the exploration of structural alerts associated with mitochondrial toxicity. These findings collectively aid scientists in developing safer drugs for human use in the realm of drug development.