Fig. 2: CNN-DM analysis.
From: Nanopore-based massively parallel sensing for peptide profiling and protein identification
a A 3D filtered scatter plot from the data points of Pep1: SDVTNQLVDFQWK (blue), Pep2: LGMAVSSDTCRSLK (orange), and Pep3: YPYVAVMLK (green) falling within mean ± one standard deviation of their τoff, I/I0, and STD, comprising 20,000 peptide reads for each peptide. b (top left) Examples of trend distribution map by overlaying all 3D filtered peptide reads of Pep1 (blue), Pep2 (orange), and Pep3 (green). (right) Algorithmic flow for transforming 3D-filtered superposition maps into density matrices (DM). (bottom left) The heatmap of the DMs corresponding to the peptides presented on top. c (left) Illustration of the CNN architecture and the training process. During the training phase, 70% and 10% of the data were used for training and validation, respectively. Then, the CNN performance was tested with the 20% test set during the inference phase. (right) The process of DM reconfirmation relies on a similarity assessment between the peptide-sensing trace and its designated DM. As illustrated for two reads initially classified as Pep1 by CNN: the red one, showing high similarity to the Pep1 DM reference, is retained; in contrast, the blue one, originally belonging to Pep2 but misclassified as Pep1, is filtered out due to low similarity. d For Pep1, Pep2 and Pep3, the classification accuracy of CNN alone was 97.5%, whereas the CNN-DM improved the accuracy to 99.2%.
