Neural network–based identification of nanobarcode proteins.

(A) Schematic of the neural network used for identification of nanobarcodes from pixel-wise fluorescence information. Brightness values across all emission channels are fed to the network as input, which, in turn, has been trained to predict the probability of this information pertaining to a specific nanobarcode, or a blank pixel. The trained network can readily be applied to full micrographs as well as stacks of images to produce false color outputs illustrating spatial distribution of proteins (further details in S18 Fig). (B) Example images of HEK293 cells transfected with specific nanobarcodes. To account for all possible emission features (including bleed-through), we acquired 11 frames for each area, consisting of the following: 405 nm excitation, with emission windows in blue, green, red, deep red; 488 nm excitation, with emission windows in green, red, deep red; 561 nm excitation, with emission windows in red and deep red; 633 nm excitation, with an emission window in deep red; brightfield. The panels in the left column show an overlay of the 4 brightest frames: 405 nm excitation, blue emission (in cyan); 488 nm excitation, green emission (in green); 561 nm excitation, red emission (in red); 633 nm excitation, deep red emission (in magenta). False color neural network output images are shown in the right column of (A). (C) Prediction accuracy of the neural network over a hold-out test dataset. For each protein, bars represent the precision (top), recall (middle), and F1-score (bottom). (D) False positive and false negative protein identifications (as percentage of all false predictions). For further details about the experimental procedures, imaging settings and neural network analysis, see the Methods section. For practical implementation purposes, we concentrated here on a subset of the labeled proteins, which were also used for the Nrxn/Nlgn experiments in Fig 4. Scale bars: 20 μm. The data underlying this Figure are available as file “Fig 2_CD.xlsx” from http://dx.doi.org/10.17169/refubium-40101.