Epitope Generation for Peptide-based Cancer Vaccine using Goal-directed Wasserstein Generative Adversarial Network with Gradient Penalty

Abstract We introduce a novel immunogenicity goal-directed peptide sequence generator in a Wasserstein generative adversarial network (GAN) with gradient penalty. The convolutional neural network-based generator is guided by a critic incorporated with an immunogenicity predictor from DeepImmuno-CNN to generate immunogenic epitopes for bladder cancer vaccines. The convolutional neural network-based critic network is trained to guide the training of the generator with the peptide sequences received from the generator or the training data containing the bladder cancer epitope sequences that are bound with the human leukocyte antigen, HLA-A*0201. We trained the critic with the generated peptide sequences with the immunogenicity scores lower than the average immunogenicity score from a batch of samples. The results show our generator can produce more immunogenic peptides and can produce peptides that are similar to the epitopes shown in bladder cancer cells. Our generator provides more epitope candidates to facilitate the design of peptide-based cancer vaccines.