ignatovmg
commented
5 years ago One more idea for a little different CNN architecture:
Make two separate CNNs for MHC and peptide. So in the end youre gonna have two grids - one for MHC, one for the peptide. Then instead of fully connected layer, compute the dot product of the output grids (for each of the channel), which will yield a vector, which you can reweight (fully connected layer) to get a prediction.
This way, it will be a little more physics inspired.
Here is the paper link https://www.biorxiv.org/content/biorxiv/early/2019/08/23/738690.full.pdf
cubazis
We are using CNN to score original non-minimized peptide samples created by BRIKARD. Training is here and the models are here with ResNet34 and RegModel1 being equally okay. The pipeline is the following:
Many samples of a peptide in MHC binding site are created.
There are converted to multichannel 3D images.
Images are scored using CNN and best N are selected.
Best N are minimized using Rosetta (so far) and reranked
Best conformation is selected.
Summary:
The task is to improve the training routine and CNN architecture. First, we can try several things:
Right now I'm learning heavy atom RMSD values directly - this may not the the most optimal option. Instead, we can try to split peptide poses into 4 classes of prediction quality (e.g. high, medium, acceptable, bad) and instead of regression learn ranking (torch.nn.MarginRankingLoss).
For making grids I invented my own channels by assigning several properties to each charmm atom type. We should try to use embedding from pytorch instead. Namely each atom type will have its own channel and let pytorch mix them the way it finds best https://pytorch.org/docs/stable/nn.html#torch.nn.Embedding
Further play with CNN architecture.