Tonight, I watched Harika Urel from the Helmholtz AI Institute and Technical University of Munich present a five-minute talk at London Calling 2023. The intriguing title of the recording is “Squiggle analysis for metagenomics viability inference.” Urel is a Ph.D. student in Dr. Lara Urban’s lab. I have watched several talks by Urban in the last couple of years. Urel’s project is focused on squiggle analysis for metagenomic viability analyses, and they began by describing metagenomics and studying microbial communities through sequencing approaches. However, one disadvantage is that metagenomics cannot distinguish living and non-viable organisms. Urel believes that raw Oxford Nanopore Technologies (ONT) “squiggle” data may be able to distinguish viable and non-viable organisms. They generated data by culturing E. coli without nutrients for forty-two days. Urel centrifuged the samples and obtained floating DNA from the supernatant, assuming it was mostly from dead microbes. Urel then created a “one-dimensional residual network” model to classify signals. Urel shared several metrics. While the recall was high (0.77), specificity was low (0.33), presumably because there is a mix of DNA from both living and dead E. coli. The computational framework using neural networks will be improved upon with more data, hopefully generating better models and specificity. Urel is working on hyper parameter optimization and incorporation of explainable AI techniques too. I will try to follow Urban and Urel’s work to learn about releases of this model, which we could use in the metagenomics course, for example.
