Pangenome-Informed Language Models for Privacy-Preserving Synthetic Genome Sequence Generation 1. The paper introduces two novel tokenization schemes for synthetic genome generation using pangenome graphs: Pangenome-based Node Tokenization (PNT) and Pangenome-based k-mer Tokenization (PKMT). 2. These tokenization methods aim to enhance DNA sequence modeling by capturing longer sequence contexts while ensuring privacy through differential privacy (DP) techniques. 3. Pangenome graphs provide critical sequence alignment information, boosting model performance in generating synthetic DNA sequences compared to traditional tokenization schemes. 4. The study emphasizes the importance of privacy in genomic data, addressing the challenges of sharing sensitive data while maintaining high utility in synthetic datasets. 5. Differential Privacy (DP) is applied during training, introducing noise to protect individual genome sequences, balancing between privacy and model accuracy. 6. The experiments demonstrate the effectiveness of PNT and PKMT, showing improvements in training time, sequence alignment, and overall model scalability for genomic tasks. 7. Future work will focus on improving DP training performance and further exploring the application of pangenome-informed tokenization schemes in various genomic research areas. @f_charton @erikgarrison 📜Paper: doi.org/10.1101/2024.09.18.6…