Basecaller updates are more frequent now, with new models and speed improvements. Tonight, I watched the London Calling 2024 session on basecaller advancements. Sam Davis, Principal Scientist, Machine Learning with Oxford Nanopore Technologies (ONT), spoke about the recently released models and chemistries. The V5 models offer 1.5Q accuracy improvement when compared to previous versions. Davis shared data on bacterial assemblies for several versions, emphasizing the improvement in accuracy. Improved bacterial isolate accuracy and variant calling accuracy were observed. For RNA, the V5 model (SUP) has a 4.3Q improvement. Davis described the training and validation pipeline ONT has implemented, with signal and reference comparisons, training data sets, and model analysis. The model architecture for SUP has changed to a transformer model. The transformer encoder blocks can work in parallel. Davis summarized that the basecaller advancements include DNA and RNA increase in accuracy, basecaller transformer architecture, and everything is available in Dorado!
