Table 1.
Dataset by condition and treatment.
Fig 1.
Schematic of the experimental design.
HMC3 microglia cells were transfected with Lifeact-mCherry treated with CBD or vehicle followed by LPS, GP120, or Aβ42. Confocal imaging within CellROX labeled cells was used for machine learning as well as analysis. A deep learning model was trained using z-stack slices to predict the CBD and the specific treatment condition.
Fig 2.
Components of the cell image analysis.
The Lifeact-mCherry signal (red) was used to outline individual cells and to measure their surface area (μm2). The CellROX signal (green) was used to measure average pixel fluorescence intensity; maximal pixel fluorescence intensity; the integrated density relative to the surface area; the raw integrated density (sum of pixel fluorescence intensity).
Fig 3.
The effect of CBD in LPS treated microglia.
A. Representative images of the ROS signal in LPS and CBD + LPS treated HMC3 cells. B. Average ROS signal in LPS and CBD + LPS treated cells. (*: p < 0.05, **: p < 0.01, ***: p < 0.001).
Fig 4.
The effect of CBD in Aβ42 treated microglia.
A. Representative images of ROS in Aβ2 and CBD + Aβ42 treated HMC3 cells. B. Average ROS signal in Aβ42 and CBD + Aβ42 treated cells. (*: p < 0.05, **: p < 0.01, ***: p < 0.001).
Fig 5.
The effect of CBD in GP120 treated microglia.
A. Representative images of ROS in GP120 and CBD + GP120 treated HMC3 cells. B. Average ROS signal in GP120 and CBD + GP120 treated cells. (*: p < 0.05, **: p < 0.01, ***: p < 0.001).
Fig 6.
Image processing and the development of the deep learning models.
A, raw z-stack slice image of the ROS signal (left) is normalized (middle) and then transformed using histogram equalization (right). B, the architecture of a convolutional neural network (CNN) used to train the model by passing images through two convolutional layers, three linear layers, and one Softmax layer. Models 1-3 (black) were developed to predict the presence (+) or absence (−) of CBD in the experiment, Model 4 (red) was developed to distinguish between LPS, Aβ42, or GP120.
Fig 7.
Performance analysis of models 1-3.
Top panels: Accuracy curves display the model progression over ten epochs and measure the percentage of test cases correctly predicted. Middle panels: Confusion matrices show the number of true negatives (upper left), false positives (upper right), false negatives (lower left), and true positives (bottom right). Bottom panels: Cross-entropy training and validation loss curves show that all models are not overfitted.
Fig 8.
A, Accuracy curves display the model progression over ten epochs and measure the percentage of test cases correctly predicted. B, Confusion matrices show the number of true negatives (upper left), false positives (upper right), false negatives (lower left), and true positives (bottom right). C, Cross-entropy training and validation loss curves show that all models are not overfitted.
Fig 9.
Saliency maps for CBD treated cells.
A comparison of ROS fluorescent images (top) and saliency maps (bottom) between CBD treated cells that were stimulated with LPS (A), GP120 (B), and Aβ42 (C).