Ian Holmes: Hybrid HMM/neural network decoding of individual nanopore reads

Ian Holmes from the University of California, Berkeley provided an informative overview of how dynamic programming techniques are key to many nanopore basecalling and analysis. He started by showing a tweet from the London Calling conference which implied that hidden Markov models are old technology, however Ian countered that they are intimately linked with recurrent neural networks (RNN) and there is still an active role for them in basecalling. Ian described how his team are using automata theory (understanding and simplifying how machines compute) to bring together different algorithms such to deliver enhanced results.

He elaborated on how his team at UC Berkeley are developing Machine Boss, a tool for combining modular state machines (i.e. a way of building up a more complex machine from relatively simple components). One challenge of this combining modular state machines though is the huge increase in the number of states. Ian revealed that by keeping only the most likely paths, they are able to drastically reduce the size of the automata. He showed data revealing that Machine Boss, when used to combine the output of protein sequence and base calling, delivered a 12 bit improvement in signal. He also described another new tool from his group, called PoreOver. This new basecaller incorporates a nanopore-basecalling RNN and can also accept the output of other similar RNNs. The tool has been trained on large volumes of publically available data. Ian shared data revealing that the use of PoreOver on 1D² reads resulted in a 5% increase in read accuracy. Both Machine Boss and PoreOver are available on GitHub.