Fig. 3

Sample training and test images for the proposed DRPnet from selected cryoEM micrographs. a A magnified image patch of a TPRV1 particle (scale is 50 Å). b Corresponding ground truth training label obtained by applying the distance transform to the binary particle mask, with blue and yellow indicating lower and higher distance values, respectively. c Smoothed 2D particle prediction map corresponding to the output of the fully convolutional regression network shown in Fig. 2b for a single particle (left) and its 3D visualization (right). d Sample cryoEM micrograph input into DRPnet (scale is 885 Å). Yellow box represents the particle shown in a. e Particle prediction map from C computed by DRPnet for the entire cryoEM micrograph (left), and its 3D visualization with circled area showing a magnified view of the local maxima (right). f, g Positive (f, blue circles) and negative (g, yellow circles) particles used to train the classification network shown in Fig. 2c. Positive samples represent true particles, negative samples represent false detections. These positive and negative training samples are selected in an unsupervised way using the prediction confidence values from the fully convolutional regression network depicted in Fig. 2b, with high or low confidence particles corresponding to positive or negative training samples, respectively (scale is 885 Å)