Efficient hyperparameter optimization by using Bayesian optimization for drug-target interaction prediction
Efficient hyperparameter optimization by using Bayesian optimization for drug-target interaction prediction
复制标题
DOI：
10.1109/iccabs.2017.8114299
复制
发表时间：
2017-10
期刊：
影响因子：
--
通讯作者：
Tomohiro Ban;M. Ohue;Y. Akiyama
中科院分区：
文献类型：
--
作者：
Tomohiro Ban;M. Ohue;Y. Akiyama
贝叶斯优化技术能够在保持预测模型的准确性的同时，为包括许多超参数的复杂预测模型提供短的搜索时间。在这里，我们将贝叶斯优化技术应用于药物-靶点相互作用（DTI）预测问题，作为计算药物发现的方法。我们的目标是邻域正则化逻辑矩阵分解（NRLMF）（Liu等人，2016），这是一种最先进的DTI预测方法，并使用高斯过程互信息（GP-MI）加速参数搜索。四个一般的基准数据集的实验结果表明，我们的GP-MI为基础的方法获得了一个8.94倍的计算时间平均减少和几乎相同的预测精度时，测量曲线下面积（AUC）的所有数据集相比，网格参数搜索，这是通常用于DTI预测。此外，如果允许稍微的精度降低（AUC约为0.002），则可以获得18倍或更多的计算速度的增加。我们的研究结果首次表明，贝叶斯优化的DTI预测问题有效。通过加速耗时的参数搜索，即使待预测的候选药物和靶蛋白的数量增加，也可以使用最先进的模型。我们的方法的源代码可以在https://github.com/akiyamalab/BO-DTI上获得。
A Bayesian optimization technique enables a short search time for a complex prediction model that includes many hyperparameters while maintaining the accuracy of the prediction model. Here, we apply a Bayesian optimization technique to the drug-target interaction (DTI) prediction problem as a method for computational drug discovery. We target neighborhood regularized logistic matrix factorization (NRLMF) (Liu et al., 2016), which is a state-of-the-art DTI prediction method, and accelerated parameter searches with the Gaussian process mutual information (GP-MI). Experimental results with four general benchmark datasets show that our GP-MI-based method obtained an 8.94-fold decrease in the computational time on average and almost the same prediction accuracy when measured with area under the curve (AUC) for all datasets compared to those of a grid parameter search, which was generally used in DTI predictions. Moreover, if a slight accuracy reduction (approximately 0.002 for AUC) is allowed, an increase in the calculation speed of 18 times or more can be obtained. Our results show for the first time that Bayesian optimization works effectively for the DTI prediction problem. By accelerating the time-consuming parameter search, the most advanced models can be used even if the number of drug candidates and target proteins to be predicted increase. Our method's source code is available at https://github.com/akiyamalab/BO-DTI.