It's a generative-adversarial-network-based model called PLocGAN, which could generate protein fluorescence images with quantitative fraction annotation to alleviate the insufficiency of the quantitative fraction of protein expression.
There are two data sources. One is the real dataset, which can be accessed by 'https://murphylab.cbd.cmu.edu/software/2010_PNAS_Unmixing/', and the other comes from the subcellular section in the Human Protein Atlas (HPA, https://proteinatlas.org). Run the codes step by step to get the single-cell images.
The 'model.py' and 'network.py' are the base model without contrastive learning module, and the 'model_cl.py' and 'network_cl.py' have the contrastive learning module. Run .\train.py to train a generative model. And the model will be applied to the unmixing model (part3).
The model is based on Bestfitting, can be obtained by 'https://github.com/CellProfiling/HPA-competition-solutions/tree/master/bestfitting'. Run .\run\train_gan.py to get a quantitative prediction model that introduces PLocGAN. Run.\run\train.py to get a baseline if you want to compare to the unmixing model with PLocGAN.