To detect cancer we turned genomes into images and put them through DCNNs (opens in new tab)

This is right, a convolutoonal network is the right choice for images, but images are not the right choice for “vectorising” the genomic data, because vectorising it in a 2D grid and then using a convolutional network on it (a network designed to exploit spatial hierarchy) is introducing unnecessary and arbitrary constraints on data locality. Definitely a souped up Honda Civic - love the metaphor

Traditional variant calling, that is bayesian methods, have gotten very good at detecting point mutations, that is single letters. But have hit a wall when it comes to more complex structural mutations, we wanted to build this for example to focus on driver/passenger mutations in tumor/normal samples. (check out the "Preface" post for our LSTM version, and the logic behind this)