Cropping Graph Convolution Neural Network for Prediction of Compound Carcinogenicity

In drug discovery, the assessment of the compound carcinogenicity is an important task. Many compound tests have been conducted in the determination of carcinogenicity in rodents. However, compound tests on animals are expensive, and usually too slow to cope with the number of chemicals. In order to reduce the cost of compound tests, there is an urgent need for carcinogenicity prediction models based on chemical structure and properties. Aiming at the problem of compound carcinogenicity prediction, we propose the Cropping Graph Convolution Neural Network (CropGCN). CropGCN utilizes clustering convolution (ClusterConv) to focus on the aggregation of neighboring atom information, avoiding the interference of far-topological atom information to achieve better atom embeddings. In addition, Node Filter Attention (NFA) is designed in CropGCN to train the atom importance, and then filters the lower importance ranked atoms to obtain high-quality molecule embedding by pooling. Experimental results on multiple rodent carcinogenic compound datasets show that our proposed CropGCN achieves the best accuracy of compound carcinogenicity prediction.