Figure 7. NeuroPAL Software: An Algorithm for Semi-automated Neuronal Identification and an Algorithm to Generate Optimal-Coloring Solutions for Cell Identification.
(A–C) The algorithm used for semi-automated neuronal identification. (A) Raw images are filtered to remove non-neuronal fluorescence and neurons are detected in the filtered image. Detected neurons are identified by matching them to a statistical atlas of neuronal colors and positions (Table S3). (B and C) Semi-automated neuronal identification accuracy begins at 86% for the head and 94% for the tail. Manually identifying eight neurons raises the head accuracy above 90%. Overall accuracy is displayed as a black line. Accuracy for each ganglion is displayed as a dotted, colored line (see legend). Many of the neurons and ganglia have high identification accuracy and confidence. The ventral ganglion is a problematic area, likely due to the density and high positional variance therein.
(D and E) The algorithm used to generate optimal-coloring solutions for cell identification—for any collection of cells in any organism. We show simulations of two approximately optimal alternatives to NeuroPAL, one that permits as many reporters as NeuroPAL (D) and one that restricts the transgene to only 3 reporters (E). With the exception of the number of reporters, both alternatives were generated using parameters similar to NeuroPAL: three landmark fluorophores, where each fluorophore is distinguishable at three intensities (high, medium, and low). Reporters were chosen by the algorithm from those available in WormBase, a community-curated database of cell-specific reporter expression. Similar databases are available for other model organisms (e.g., fly, fish, and mouse). We evaluated the two NeuroPAL alternatives by computing the percentage of their color violations, defined as neighboring neuron pairs with indistinguishable colors. See Methods S1 and S2 for algorithmic details and validation.