Mitigating Class Imbalance in Small Molecule Binary Classification Using Graph Neural Networks and Prototype Loss Optimization

Efficient and accurate small molecule classification methods can significantly improve the efficiency of scientific research and industrial applications, but in real scenarios, many datasets suffer from a serious imbalance in the ratio of positive and negative samples. For example, in the drug screening process, the number of active compounds (positive samples) is much smaller than inactive compounds (negative samples), the imbalance of samples would affect the classification model in the training process, then the model tends to predict the majority of classes. So traditional classification methods and common deep learning models are often difficult to learn robust decision boundaries, resulting in degraded classification performance. To solve this problem, we propose a binary classification model based on graph neural networks, which effectively mitigates the class imbalance problem by constructing a more discriminative model structure and optimizing the inter-class distinction using a prototype loss function. The experimental results show that the method has significant improvement in the recognition ability of minority class samples.