Extended Data Fig. 1: Overall architecture of self-supervised cellular content exploration module.
From: MiLoPYP: self-supervised molecular pattern mining and particle localization in situ
a) Preprocessing step which involves initial coordinate calculation based on Difference of Gaussians (DoG) and patch extraction from corresponding locations in 3D tomogram and 2D tilt series. 2D patches obtained by averaging tilt series projections centered on each position in 3D and corresponding sub-tomogram projections in the Z-direction are concatenated and serve as input to train the model. A non-maximum suppression (NMS) module is used to identify the final candidate positions. b) MiLoPYP’s contrastive representation learning network follows a simple Siamese ‘SimSiam’ architecture. The input and its augmented version are used as inputs and the network learns to maximize the similarity between the pairs. The network is composed of an ResNet-based encoder, a multi-layer perceptron (MLP) projector and a MLP predictor. Stop-gradient is used to avoid model collapse.
