Abstract
Nanopore sequencing enables the simultaneous detection of genetic sequences and DNA modifications, yet the development of accurate, open-source computational models for these tasks, particularly for non-ONT platforms, remains challenging. To address this, we developed Bream, an open-source deep learning framework that integrates a convolutional neural network with a reverse long short-term memory network for base calling and a bidirectional LSTM with an attention mechanism for methylation detection. We trained and evaluated Bream on datasets from A. thaliana, O. sativa, and D. melanogaster generated using a novel nanopore sequencing platform (Qitan Technology's QCell-384) featuring engineered helicase and nanopore proteins. The framework achieved base-calling accuracies between 89.38% and 91.83%, comparable to ONT's R9.4 platform, and demonstrated high-performance methylation detection, with an AUC-ROC of 0.98 on a D. melanogaster dataset. Furthermore, its estimates of whole-genome CpG methylation frequency showed strong agreement (Pearson's r ≥ 0.96) with bisulfite sequencing data across species. These results demonstrate Bream as a powerful, transparent, and adaptable tool that facilitates simultaneous base calling and methylation detection on emerging nanopore sequencing platforms, thereby advancing open innovation in the field.