Marcus Stoiber, a Senior Data Analyst at Oxford Nanopore Technologies (ONT), presented on “Nanopore methylation: a better way to mods” at the Nanopore Community Meeting 2021. Stoiber and team devised the Remora modified base detection framework. Remora separates methylation/modified base calling from the basecalling neural networks. This separation provides high accuracy for methylation calls and canonical basecalls as well as simpler training samples. With the kit 12 (R10.4) the accuracy was increased. With simpler training samples, Stoiber noted that this allowed high accuracy 5mC. Stoiber explained how remora works using neural network weights, convolution, and modified base probabilities as an output. The new approach will improve incorporation of new models. Now Remora models can be trained in hours on reasonable GPU resources. Stoiber explained how benchmarking was performed training on native human data. The improved accuracy of canonical and modified base calls seems to be related. Also, 5hmC training samples were tested. Using human data, they identified 5hmC in HG002 at about 2.7% of CpG calls, which corresponds with data. For future directions, Stoiver spoke about optimizing training and incorporating all-context and new modified base models into Remora and Guppy. This presentation was two years ago, and now there are several additional improvements and incorporation of updates into Guppy.
