基于降噪自编码神经网络的化合物毒性预测方面的研究

Research on toxicity prediction of compound based on denoising autoencoder neural network

常规毒理学实验方法周期长、耗资高,对现代药物研发和环境化合物安全性评估具有局限性,通过对化合物毒理性研究,提取1 047维分子指纹特征,提出去噪自编码神经网络无监督学习机制及对腐败特征的自联想学习特性提取隐含毒性化合物特征,实现化合物毒性预测和毒性化合物的活性预测。该方法在化合物毒性预测和活性预测中的预测精度分别为79.825%、80.85%,敏感性分别为79.62%、80.25%,特异性分别为80.03%、81.45%。实验结果表明,去噪自编码网络较浅层机器学习更适用于高通量化合物毒性预测,较传统自编码网络更具优越性。

Conventional toxicology experimental method has a long cycle and high cost,which is limited to the safety assessment of modern drug development and environmental compounds.Based on the study of the toxicity of compounds,this paper extracted 1 047 dimensional molecular fingerprint features,using the unsupervised learning mechanism of the denoising autoencoder neural network and the characteristics of the self association learning of the corrupt characters extracted the features of the hidden toxic compounds,achieved the prediction of chemical toxicity of compounds and activity of toxic compounds.The prediction accuracy of this method was 79.825%,80.85%,sensitivity was 79.62%,80.25%and specificity was 80.03%,81.45%respectively.The experimental results show that denoising autoencoder neural network is more suitable for high throughput toxicity prediction than shallow machine learning,and is superior to the traditional autoencoder neural network.