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

TMLR Paper5367 Authors

Abstract: We introduce a novel goal-directed Wasserstein Generative Adversarial Network with Gradient Penalty (GD-WGAN-GP) for training a generator capable of producing peptide sequences with high predicted immunogenicity and strong binding affinity to the human leukocyte antigen HLA-A*0201, thereby eliciting cytotoxic T-cell immune responses. The proposed GD-WGAN-GP incorporates a critic network to guide the generator in producing peptides with a strong binding affinity similar to those in the training set and a reward network to steer the generator toward producing sequences with high predicted immunogenicity. To avoid the generator prioritizing the objective of the critic at the expense of immunogenicity, we introduce a scaling factor to balance the influence of the reward in the loss of the generator. To reduce peptide repetition, we integrate the reward into the loss of the generator using two approaches: a switching mechanism that excludes the reward term when duplicated peptides are present in a batch, and otherwise multiplies it by a $\gamma_{max}$ parameter to control the reward's contribution, and (2) a repetition penalty from ORGAN, where each reward is divided by the number of occurrences of its corresponding peptide within the batch. Experiments on bladder cancer epitope sequences demonstrate that GD-WGAN-GP with the switching mechanism enables a tunable trade-off between the number of unique peptides and the average immunogenicity score via varying $\gamma_{max}$. Furthermore, the generator trained by the GD-WGAN-GP with the ORGAN’s repetition penalty achieves an optimal balance of uniqueness and immunogenicity. Across multiple datasets, GD-WGAN-GP outperforms existing methods by effectively reducing peptide redundancy while preserving high immunogenicity scores and strong binding affinity. The Python codes are provided at: \url{https://github.com/AnnonymousForPapers/GP-WGAN-GP_with_switch_and_ORGAN_penalty}.

Submission Length: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Adam_Arany1

Submission Number: 5367