Preforms De novo helical protein design using machine learning and PyRosetta to generate a novel synthetic protein structure.

Install PyRosetta as the website describes.

Use the following command (in GNU/Linux) to update your system and install all necessary programs and python modules for this script to run successfully:

Get this script:

RamaNet (after Gopalasamudram Ramachandran) is a script that uses Machine Learning (a GAN network) and PyRosetta to preform De Novo Helical Protein Design (from the beginning) i.e. generate and design a synthetic helical protein structure completely computationally. There is no input for this script, it autonomously generates a topology (random every time) then designs a sequence that fits the designed topology using the RosettaDesign protocol. It then submits the structure's FASTA sequence to the Robetta server to generate and download the custom fragments files in preparation for an Abinitio folding simulation. The Abinitio protocol script can be found here. Finally it calculates the RMSD for each fragment's position on the designed structure and plots an (RMSD vs Position) graph to indicate how good the Abinitio folding simulation might go (ideally you want an average RMSD < 2Å).

The script will generate one structure. It is advised to run this script and generate multiple structures and see which one has the lowest average RMSD fragments value. But mind you, if you generate too many structures this might overwhelm the Robetta server by submitting and requesting too many fragment files, please be considerate and run this script once to generate one structure at a time only.

Here is a video that explains this whole process. For quick structure generation skip to the last steps (steps 3 and 4).

You do not need to generate the Machine Learning datasets, they are already provided and can be downloaded here:

Dataset name Description Helix PS dataset Dataset of only helical structures' Φ/Ψ angles

But if you want to replicate our work use the following command to generate the Machine Learning dataset from the Protein Databank Database (computation time ~168 hours and requires more than 128GB of free disk space):

python3 RamaNet.py --dataset or python3 RamaNet.py -d

The default parameters for the Database.py script is isolating proteins between 80 and 150 amino acids, that have more helices and strands than loops (a rigid structure), and with an Rg value of less than 15Å (compact structure). The script results in a dataset with the first column as the training example number, then the PDB ID of the file (and chain letter), then the angles Φ/Ψ for each amino acid (PS dataset). 0.0 indicates a position with no amino acids, not all protein structures have the same length, but the entire dataset does have the same length and shape because the empty spaces are filled with zeros. If errors occur, that is fine, some protein files will cause errors (they will be deleted/ignored), but the script should continue all the way to the end and result in a dataset file.

The dataset generation protocol is as follows:

Download the PDB database

Extract files

Remove non-protein structures

Remove structures less than or larger than a specified amino acid length

Remove structure with broken chains

Remove structures that have loops that are larger than a specific length

Renumber structures' amino acids

Remove structures that are below a specified Radius of Gyration value

########## --- HUMAN EYE FILTERING --- ##########

Clean every structure in the database

Make a list of all paths (if next step is performed in a high performance computer HPC)

Generate HPC submission file (PBS job scheduler)

Relax each structure multiple times (on a HPC or a local computer), this is to augment the examples

Get each residue's Φ/Ψ angles (and constraints if required - you have to comment in and out the relevent functions at the end of the script)

The most difficult step is the Human Eye Filtering step which requires a person to filter out all the unwanted structures manually before moving onto cleaning up each structure and augmenting the data which eventually results in a .csv file. Unwanted structures are non-compact structures, structures with more loops than helices and sheets, weird looking structures. Also, this is the step to separate structures and collect the ones with traits that you need; we decided to separate the dataset into structures with only helices before augmenting the dataset. This was because it is easier to construct helical proteins topology and we wanted to proof our concept before moving into more complicated datasets. The separation was done manually.

It is best to contact me if you want to generate your own database and I will walk you through the protocol, it is not difficult, but works on individual basis.

You do not need to train the neural neural network because it is already trained and the weights file is available here:

Weights name Description Helix PS Weights Neural network weights generated from the Helix PS dataset

If you wish, you can use the following command to train the neural network on the dataset (whether you use our dataset or generate your own):

python3 RamaNet.py --train or python3 RamaNet.py -t

Use the following command to generate a novel protein structure, generate fragments from the Robetta server, download these fragment files, and analyse these fragments:

python3 RamaNet.py --fragments USERNAME or python3 RamaNet.py -f USERNAME

USERNAME is the username at the Robetta server for fragment generation.

Use the following command to only generate a novel protein structure without generating any fragments:

Make sure you have the weights directory available, either from your training or downloaded from step 2 (provided by us), and that it is in the same directory as the RamaNet.py script. The directory must be named weights.

This option (computation time ~24 hours) will result in 1 file (if no fragments are requested), or 7 files (if fragments are requested):

The topology file, which is basically just the structure of the backbone constructed using a sequence of Valines (backbone.pdb)

The final designed structure file - RosettaDesign (structure.pdb)

Abinitio input files (structure.fasta, frags.200.3mers, frags.200.9mers, pre.psipred.ss2)

Fragment quality plot (plot_frag.pdf)

When using these scripts kindly reference the following: