AlphaRING

AlphaRING is a package designed to predict the pathogenicity of any given missense variant. It is a customised implementation of AlphaFold2, which models a monomeric wild-type protein and a missense variant counterpart. Using RING4, AlphaRING captures their non-covalent bonds and analyses the differences in bond formation between the wild-type and variant proteins to predict the pathogenicity of the missense variant.

Note

AlphaRING's source code refers to bonds as edges and residues as nodes.

An AlphaRING manuscript is currently under review by RECOMB 2025 and is available via bioRxiv.

AlphaRING benchmarking data will be made available soon...

Overview

Figure 1 Overview of the AlphaRING workflow

For any given missense variant, AlphaRING conducts the following workflow:

1. Accept FASTAs:

   In this step, AlphaRING accepts two FASTA files: one for the monomeric wild-type protein and one for the missense variant, which differs by a single residue.

2. Predict structures:

   In this step, AlphaFold2 is used to predict the structure of the wild-type and variant proteins. The best model of each is relaxed and used going forward.

3. Generate residue interaction networks (RINs)

   In this step, RING4 is used to generate a RIN of both the wild-type and variant models, capturing their non-covalent bonds.

4. Calculate residue weightings

   In this step, AlphaRING assigns each non-covalent bond a weighting of importance to protein stability. Weightings are calculated using novel bond-type specific formulas that take into account bond-specific energy and geometry as calculated by RING4. As the energy values provided by RING4 are fixed for a given bond-type, our formulas consider the bond's variables, distance and angle, to multiply energy by a value between 0–2. Both distance and angle can contribute a value between 0–1 towards the multiplier. More favourable distances and angles result in a larger multiplier. Therefore, a higher bond weighting indicates greater importance to protein stability.

   Our bond-type specific formulas come in three flavours. The first flavour is used when a shorter distance and smaller angle is favourable (π-cation and π-π stacking):

$$Bond_{weight} = energy \times \left( \left(1 - \left(\frac{distance}{distance_{max}}\right)\right) + \left(1 - \left(\frac{angle}{angle_{max}}\right)\right) \right)$$

        The second flavour is used when a shorter distance and larger angle is favourable (hydrogen):

$$Bond_{weight} = energy \times \left( \left(1 - \left(\frac{distance}{distance_{max}}\right)\right) + \left(\frac{angle}{angle_{max}}\right)\right)$$

        The third flavour is used when a shorter distance is favourable and angle is negligible (ionic and π-hydrogen):

$$Bond_{weight} = energy \times 2 \times \left(1 - \left(\frac{distance}{distance_{max}}\right)\right)$$

        Each residue's weight in the wild-type and variant proteins is calculated by summing the weight of its bonds.

5. Calculate fold change (FC):

   In this step, AlphaRING calculates the FC between the weight of the wild-type and variant residue at the position of substitution. Therefore, values further from 1 indicate a greater change in weighting.

6. Calculate AlphaRING score:

   In this step, the absolute log2 of the FC is taken to provide a final metric, the AlphaRING score. This results in a minimum AlphaRING score of 0. Therefore, higher AlphaRING scores indicate greater predicted pathogenicity.

Dependencies

Before installation, you will need a machine running Linux and a modern NVIDIA GPU. In addition, ensure you have the following dependencies:

• AlphaFold2 databases (databases for AlphaFold-Multimer are not required)

• RING4 (version v4.0-2-ge939f57)

• Git

• wget

• Miniconda

Installation

Firstly, git clone the alpharing repository into the same parent directory as RING4 and cd into it:

git clone --recurse-submodules https://github.com/loggy01/alpharing
cd ./alpharing

wget stereo_chemical_props.txt into the alphafold subdir:

wget -P alphafold/alphafold/common/ https://git.scicore.unibas.ch/schwede/openstructure/-/raw/7102c63615b64735c4941278d92b554ec94415f8/modules/mol/alg/src/stereo_chemical_props.txt

cp your RING4 directory into the alpharing directory:

cp -r ../<dir> ring

Note

Replace <dir> with the name of your RING4 directory.

conda create the alpharing environment:

conda create -n alpharing -c bioconda -c conda-forge hhsuite hmmer kalign2 openmm=8.0.0 pdbfixer python=3.10

Finally, conda activate the alpharing environment and pip install necessary packages:

conda activate alpharing
pip install absl-py==1.0.0 biopython==1.79 chex==0.1.86 dm-haiku==0.0.12 dm-tree==0.1.8 immutabledict==2.0.0 jax==0.4.25 ml-collections==0.1.0 numpy==1.24.3 pandas==2.0.3 plotly==5.15.0 scipy==1.11.1 tensorflow-cpu==2.16.1 jaxlib==0.4.25+cuda11.cudnn86 -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.html

Usage

conda activate the alpharing environment and run the script run_alpharing.py in a shell script as follows:

conda activate alpharing
<path to run_alpharing.py> \
  --fasta_paths=<path to wild-type FASTA>,<path to variant FASTA> \
  --max_template_date=<yyyy-mm-dd e.g. 2020-01-15> \
  --data_dir=<path to AlphaFold2 databases dir> \
  --output_dir=<path to dir to save all results> \
  --uniref90_database_path=<path to uniref90.fasta> \
  --mgnify_database_path=<path to mgy_clusters_2022_05.fa> \
  --template_mmcif_dir=<path to mmcif_files dir> \
  --obsolete_pdbs_path=<path to obsolete.dat> \
  --bfd_database_path=<path to bfd_metaclust_clu_complete_id30_c90_final_seq.sorted_opt> \
  --uniref30_database_path=<path to UniRef30_2021_03> \
  --pdb70_database_path=<path to pdb70> \
  --use_gpu_relax=<whether to relax with GPU i.e. True or False>

Note

Replace argument values with your own. AlphaFold2 databases named with dates may differ.

Warning

These are the minimum requirements to run AlphaRING. Specific users (particularly HPC users) may need to include more in their shell script.

Downstream analysis

For any given missense variant, AlphaRING will generate two subdirs within output_dir: one for and named after the wild-type FASTA, and one for and named after the variant FASTA.

In addition to the regular AlphaFold2 and RING4 output, each subdir will contain a version of the ringEdges and ringNodes files with a Weight column attached. The latter file type is plotted and saved as an interactive residue weightPlot HTML:

Figure 2 Example of a pair of wild-type and variant weightPlots

The HTML can be used to visually compare weights of a substituted residue between a wild-type and variant, or be used for exploratory analysis of unsubstituted residues. Additionally, the HTML file can be saved as a PNG.

Finally, the variant subdir will contain an alpharing_score.txt file, which only contains the AlphaRING score of the missense variant.