TorchProteinLibrary
TorchProteinLibrary copied to clipboard
PyTorch library of layers acting on protein representations
Very nice work! But is it possible to install TorchProteinLibrary on a CPU only (without cuda) environment?
I reconstructed the backbone for the protein 6MSP from the PDB, using real dihedral angles fed to Angles2Backbone. I calculated planar angles for both the original backbone versus for the...
In https://lupoglaz.github.io/TorchProteinLibrary/fullatommodel.html#Angles2Coords, the line `angles.data[:,3,:] = np.pi` should read `angles.data[:,2,:] = np.pi`.
Current unit tests are manual in most cases. We should automate it further and use torch.autograd.grad_check to check gradients instead of homemade code. Precision checks are also needed.
We need to refactor Angles2Coords, namely cConformation and cRigidBody classes, such that: 1. Chemical parameters are defined in one place and can be easily modified 2. Make Amino-acid construction code...
We need benchmarks for speed and memory usage for all the layers. Right now it's hard to predict run time of a new algorithm.
In principle we need Angles2Coords and other FullAtomModel layers working on GPU. The only complication is that we use recursive depth-first graph traversal which is not suited for vector processors....
The ultimate test for the library is the replication of a published paper. We have to replicate this paper and make learned models defaults as was done with ResNet's ,...
TorchProteinLibrary 0.2 Nucleic Acid Extension
Hello, I encountered a compilation error while trying to install TorchProteinLibrary. Below are the error details and a screenshot. Environment: Operating System: Linux Python version: 3.8 PyTorch version: 1.7.0 `gcc...