Leopard-EM proposes a parallelizable, GPU-friendly, accessible package for 2DTM with a PyTorch backend. In addition to a 2DTM Python package, the authors propose and implement two new ideas: optimizing the 3D reference template scaling to have the highest correlation with the micrographs and a constrained search among position, angle, and defocus to significantly improve speed and accuracy of finding particles that tend to be near another particle. In this work, we replicated and extended the Leopard-EM framework to verify the authors’ claims of accuracy and ease of use. Using the EMPIAR-10998 dataset, we performed unconstrained 2DTM searches with the same 3D references (60S LSU, 40S SSU, and 20S proteasome). We reproduced the constrained search strategy and attempted to benchmark accuracy and runtime against the C++ cisTEM implementation. We further extended Leopard-EM by applying it to ATP synthase. Our experiments can be reproduced utilizing the code in this repository.
To reproduce the code, first create a virtual conda environment with the proper dependencies. Then, activate the virual environment.
conda env create -f environment.yml
conda activate myenvThe ground truth MIP graphs can be downloaded from EMPIAR-10998 datasets, and fall under yeastcell/mature60S. They are not in the repository because the mrc files take up a lot of space.