基於複數多維結構特徵提取的雌激素受體活性預測器-多維度學習

Abstract

隨著人類工業及科技發展，許多新興的化學物質隨著工業製程，或是其為產品本身被排放至自然環境中。其中某些新興化學物質將會干擾體內荷爾蒙運作，因此被稱作是內分泌干擾素。目前已有多項報告顯示人類所排放的化學物質將會對環境中的生物造成生理功能上的負面影響，另外，某些人類於工業製程中使用的化學物質也具有一定的腺體毒性。然而，此類新興物質具有大量類似異構物，實驗上鑑定大量化學物質需耗費大量時程，因此提出一個能預測化學物質與受體活性的模型將有助於人類鑑定及分析內分泌干擾素。本研究利用CERAPP所提供的資料集，建構出一種以化學物質之One-hot編碼與化學結構圖，兩種不同維度的結構資訊做為特徵提取對象，並嘗試預測其與雌激素受體的結合活性，稱作Multi-D learning。本研究報導了透過兩種複合數據訓練的深度學習網路比兩者個別訓練的網路表現得更好，並進行了案例研究，並且得出了幾例十分有利的結果。綜上所述，我們訓練了一個多輸入的類神經網路稱作Multi-D learning，並且可以用於其他研究人員篩選會與雌激素受體結合的化學物質。Because of the development of industry and technology, a lot of synthesized emerging chemicals are emitted into the environment. Some of them can interfere with functions of body’s hormones, called endocrine disrupter chemicals (EDCs). A lot of studies report that chemicals emitted by human will cause negative effects on physiological functions to creatures in the environment. Moreover, some of chemicals that are used in processes of industrial are also poisonous to glands. However, there are many similar modified molecules of those emerging chemicals. It will take a huge amounts of time to identify and research all of EDCs. According that, building an automatic research system to help scientists quickly study mechanism of EDCs is important. Here, an artificial neural network, called Multi-D learning, was built for predicting the binding score between chemicals and estrogen receptor by CERAPP dataset. Data from CERAPP were transformed into two kinds of data, one hot encoding and 2D chemical structure picture. They were used to train the model individually or together. The research reported that the model trained by multi data show better performance on the prediction accuracy. To sum up, a multi input artificial neuron network was created, called Multi-D learning. It will be used to help other researcher filter estrogen receptor binding chemicals.