Chemical Medicine Classification Through Chemical Properties Analysis

At the present time, only large multinational pharmaceutical companies have the financial ability to research new drugs. Thus, reducing the research and development costs of new drugs is an important subject. Through the in-depth mining of existing drug data, this paper aims to classify unknown drugs and provide assistance for drug screening during the development process. This will reduce the costs of original drug research and promote the transformation of China’s pharmaceutical industry. In this paper, we first collected a drug data set using a Web crawler. Based on this data set, we derived a formula for calculating the similarity between drugs and identified the parameters of the similarity calculation formula from a subset of the data. We used the
k
-nearest neighbor classifier to categorize the drug data based on the similarity of medicines. The results show that the proposed drug classification model can achieve 77.7% accuracy, which is far better than the classification performance of a decision tree and a random forest with only one decision tree, similar to that of a random forest with 10 decision trees, and worse than that of a random forest with 500 decision trees. Although the classification method proposed in this paper is reasonable and the experimental results are in line with expectations, the proposed technique could be improved to manage problems, such as overfitting. Because this classification method is based on chemical similarity and depends entirely on the available training data (which are limited), such fitting problems are inevitable. To solve this problem, more data are needed and the existing sampling method should be improved. One possible approach is to combine this algorithm with ensemble learning techniques to avoid the phenomenon of overfitting.