Fig. 3: Predictive performance and characteristics of models found by GenomeNet-Architect and various baselines on the viral classification task.
From: Optimized model architectures for deep learning on genomic data
The inference time to classify 10,000 one hot encoded samples on a GPU and the class-balanced accuracy are shown. The size of the circles indicates the number of model parameters. We have not included Seeker8 in both graphs and PPR-Meta11 for long sequences because their performance was too low. a At the read-level (150 nt), the best-performing model selected by GenomeNet-Architect (CNN-GAP-6h) reduces the read-level misclassification rate by 19% relative to the best-performing deep learning baseline - Fiannaca10, despite having 83% fewer parameters and 67% faster inference time. b At the contig-level (10,000 nt), models found by GenomeNet-Architect perform on par or better than the best-performing baseline, although all these models perform very well in terms of accuracy. However, much faster (CNN-GAP-2h) and much smaller (CNN-RNN-6h) models are found, all with balanced accuracy close to the best baseline ( ~ 98.6%).
