Cancer Risk Assessment Based on Human Immune Repertoire and Deep Learning

Abstract

This paper discusses two types of neural networks used to construct cancer risk indices - Convolutional Neural Networks (CNN) and Transformer. Three models were derived, two of which are based on CNNs, which have essentially the same network architecture but use different encoded forms of input; the other model is based on Transformer’s architecture. The design of these three models considers the different lengths of the input sequences and utilizes all the information of the training data as much as possible. All models achieved good performance on the test dataset. With these deep models, cancer risk indices based on human immune repertoires can be constructed. Applying the risk index to real breast cancer data clearly distinguishes cancer and non-cancer groups. Moreover, the cancer risk index based on the Transformer model has the best performance.