This repo contains the functions and methods for the paper
HyperMPNN ‒ A general strategy to design thermostable proteins learned from hyperthermophilic organisms
https://doi.org/10.1101/2024.11.26.625397
and is primarily intended for the reproduction of the results. Additionally the weights for the different training settings (added backbone noise) are also included in this repository under retrained_models and can be used directly with the original ProteinMPNN (GitLink). All credit for the architecture and training scripts goes to the authors of ProteinMPNN and we would like to take this opportunity to thank them for making everything open source and available to the general public.
Important
If you just want to use HyperMPNN you can download the model weights from here and use the original ProteinMPNN implementation while specifying the new model weights with --path_to_model_weights and --model_name . An example can be found here.
This repository contains several folders (sections named after them) for which we give a brief overview in the next few sections.
In the af_models folder are two files called ids_[ecoli|thermophilic].*, containing the UniProt ids for all models used in the analysis and training. The ids belong to the proteins after the sequence clustering and plddt filter (keep > 70) step. You can download them by using the provided script and store them in two separate folders. The example workflow could look like the following:
cd af_models
mkdir -p ecoli_alphaFold_models hyperthermo_alphaFold_models
bash download_alphafold_structures.sh ids_ecoli.after_filter.list ecoli_alphaFold_models
bash download_alphafold_structures.sh ids_thermophilic.after_filter.list hyperthermo_alphaFold_models
All downloaded structures were processed again, i.e. long N and C terminal regions with a poor pLDDT value of less than 70 were truncated. You use the function filter_pLDDT.py in the af_models folder to do this and the command could look like the following below. The 20 in the command belongs to the number of cores used for filtering. That depends a lot on your machine!
For the script you also need to install pyrosetta, pandas, numpy and tqdm.
mkdir -p ecoli_filtered hyperthermo_filtered
python3 ecoli_alphaFold_models ecoli_filtered 20
python3 hyperthermo_alphaFold_models hyperthermo_filtered 20
This folder contains a jupyter notebook called calc_structural_aa_frequencies.ipynb to perform the calculation of amino acid frequencies for wild type and designed structures. All data, which was collected for native and designed structures, is stored in the folder calc_frequencies/data in pickle files.
Based on the filtered and truncated structures, the frequencies of all amino acids are calculated for the E.coli and hyperthermophilic organisms. The amino acids are then divided into different groups (e.g. positively/negatively charged) and visualized. The code for this can be found in section 2 of the notebook and an example figure looks like the following:
After the training of ProteinMPNN with structures from hyperthermophilic organisms (named HyperMPNN), both E.coli and hyperthermophilic structures were designed with default ProteinMPNN and HyperMPNN in a one-shot sequence prediction approach. Afterwards the changed frequencies of amino acids were visualized in radar plot. An example is shown below. The code for this can be found in section 3 of the notebook and an example figure looks like the following
This folder contains a jupyter notebook called calc_structural_packing_metrics.ipynb to perform the calculation of packing metrics like contact order, charge distribution, salt bridges, radius of gyration and others. The calculation was performed for native and designed for collected structures of E.coli and hyperthermophilic organisms from previous steps. All data, which was collected for native and designed structures, is stored in the folder calc_metrics/data in pickle files.
The notebook is devided in three sections. The first calculates the metrics, the second visualizes the results and in the third section the salt bridges are calculated separately. The visualisation of e.g the contact order distribution is shown below.
In this folder all scripts for the design approach of the I53-50B.4PT1 pentamer for experimental validation of the new weights. The full structure was downloaded from the PDB 6P6F and the pentamer manually extracted. After renumbering the three scripts (prefix 1_ to 3_) were executed one after the other for the selection of the interface (01_command_selection.sh), the design using the new weights (02_command_design.sh) and the generation of a consensus sequence using the top scoring sequences based on ProteinMPNN's global score (3_command_consensus.sh).
This folder contains everything to retrain ProteinMPNN using new structures. At this point please refer to the README in the folder.