Short MDs

Code

Introduction

The structural.

Training process
Figure 1: Structural part of.

Dataset

.

A Dataset sample
BDataset sample
CDataset sample
DDataset sample
Figure 2: Protein dataset generation steps.

Internal representation of a protein structure

Iterative SE(3) part of the model has internal representation of features as fields. However, final result should be converted to a set of rigid bodies. To connect the two representations, we have to represent rotation matrix and translation vector of a rigid bodies in a common reference frame as vector fields.









Training process
Figure 3: Rigid bodies representation of a protein.

Loss function

Model

Essentially the model for this dataset stays the same as for the previous one (Figure 4). THe only change is that now we have to pass another vector field between two-layer SE(3) transformer blocks, that represents rotations. Additionally, as an input, we have to pass initial protein conformation: in our case it is just an extended chain. However, in the DeepMind implementation the initial conformation can come from a template structure or from the previous result.

Training process
Figure 4: Protein structure prediction model.