The idea is to create a pipeline that can design a novel protein containing a knot. We run in a loop combination of designing a protein with diffusion followed by guiding the designed structure towards being knotted using a knot classifier model.
Scripts: Flipping pipeline
We are now using foldingdiff model to design the protein backbone.
Datasets for training the knot classifier: https://huggingface.co/datasets/EvaKlimentova/foldingdiff_knots
Models are trained on the above datasets. They use as inputs the same format as is used in foldingdiff (6-tuple of angles). I tried two model architectures: a simple feedforward neural network and a CNN.
Scripts: Feed forward neural network, CNN
Models: models folder
To better understand what is happening with the structure, we have a script that visualizes the structure in 3D and makes a video of the structure change. See the video folder.