About the Training Data Generation Pipeline

Introduction

To prepare simulation data for eventual use in training a transferable coarse grained (CG) force field, these data must first be mapped to the CG resolution and then processed to incorporate prior energy terms.

This is done as follows:

  1. First, in mlcg-tk-gen_input_data the atomistic simulations are loaded and mapped to the lower resolution, followed by the construction of the neighbor lists specific to the chosen prior energy model.

  2. Next, in the case that a prior model has not already been generated, this can be done by accumulating the statistics of the saved CG data and fitting these statistics to a set of predefined energy functions.

  3. Finally, the outputs are fed to mlcg-tk-produce_delta_forces along with the prior model so that delta forces can be calculated.

The result is a set of CG coordinates, embeddings, and delta forces which can be used to train a neural network as well as the neighbor lists associated with each molecule which can be fitted to produce new prior models, if needed.

In the details below, input files are provided as an example.

Example: 1LY2 toy dataset

To exemplify the usage of this code, we provide an example on how to load a tiny dataset of All-Atom (AA) simulations of a trpcage-variant, 1L2Y.

Warning

This example is just to demonstrate the usage of the package to transport from AA simulations to the input for training an MLCG model. This dataset does not contain enough data, nor is it in the right distribution, so that it would create a good CG model.

The dataset is provided under the folder ./demo_raw_data/. It contains a readme with details of the simulation.

0) Create a directory to output all the intermediate files

The commands will output a lot of data and its better to save it in a separate directory to ensure hygenic file management.

mkdir ./demo_processed_data