Variational Autoencoder for DNA Methylation Latent Space Embedding

The patent applies pre-training methods where models are first trained on large unlabeled datasets to learn general patterns and representations before being fine-tuned on smaller labeled datasets for specific tasks. This preliminary action allows the model to acquire general knowledge without requiring extensive labeled data for each specific task, thereby resolving the contradiction between model performance and training data requirements.

The patent employs transfer learning where knowledge learned from source tasks is transferred to target tasks. The model parameters learned from pre-training on large datasets are copied and adapted to new tasks with limited labeled data, allowing high performance to be achieved without proportionally increasing the amount of training data required for each new task.

The patent uses pre-training on large unlabeled biological datasets to establish robust initial models that capture general biological patterns. This preliminary action reduces the need for extensive annotation of task-specific labeled data, thereby maintaining high prediction accuracy while significantly reducing annotation time and costs.

The model performs self-supervised learning by utilizing the data structure and patterns inherent in the input data itself without requiring external annotations. The model learns to predict certain aspects of the data (such as masking certain tokens in language models or predicting missing values) using the data's own structure, eliminating the need for manual annotation while maintaining high accuracy.