ProteinClusterTools 1.0.3

Pipeline for clustering protein sequences in large scale datasets, using homology or pLM representation.

Protein Cluster Tools

By John Chen and Barnabas Gall. GitHub

A pipeline for analyzing protein families as clusters of related sequences.

Usage

See details on installation below.

This pipeline provides different approaches for grouping (clustering) protein sequences in sequence space, and provides ways to visualize these clusterings in an interactive manner.

The work flow broadly follows 3 steps:

1. Starting with a fasta file of protein sequences, create protein clusterings based on homology or vector representation based methods. See the run pipeline notebook.

2. Once the cluster definitions are made, create interactive visualizations of the sequence space, with user defined annotations. The main methods featured in this pipeline are:

   a. A hierarchical cluster plot that captures cluster separations broadly at different levels of clustering. Works for all methods of clustering in 1). See the hierarchical cluster plot notebook.
   b. A tree structure that shows detailed pairwise separations for all sequences (from indivual sequences to fully clustered). This is specifically for the vector representation based hierarchical clustering. See the tree structure notebook.
   c. Plotting the vector representation of each sequence individually after a UMAP dimensionality reduction. See the UMAP plot notebook.

3. When the user has explored sequence space, they can select desired representative sequences from target clusters using either an HMM or vector based approach. See representative selection, or end of tree structure in the case of dealing with targetting any node in a tree structure.

Installation

It is recommended to use a package manager like miniconda to install the required environment. If the user prefers a more modular install, see Details on dependencies below.

Note on operating system: The pipeline is in Python, but a number of dependencies require a Linux environment. If running this on Windows, it is recommended to use Windows Subsystem for Linux (WSL). Alternatively, windows users can SSH into a Linux computer.

Note on hardware: Part of the pipeline requires a GPU for using protein language models and vector representations. It is recommended to install on a computer that has a GPU (in theory CPU will also work but will be slower). Otherwise, the homology based method can be run on just CPU. The env.yaml for installation assumes the user has an Nvidia GPU that uses CUDA, if using an AMD GPU the user can try installing pytorch manually by pip (see below).

Note for conda: When using conda, it is recommended to install Mamba into the base environment for faster dependency resolution. Then use the mamba command in place of the conda command (they are mostly interchangeable).

conda install mamba

The command to install all requirements using one command is as follows:

conda env create -f env.yaml

This will create a new environment called "proteinclustertools". To run code from the command line, be sure to activate this environment first. For Jupyter notebooks (such as the examples in analyses), select this as the kernel.

conda activate proteinclustertools

Details on dependencies

The pipeline makes use of various software. The benefit of installing using conda is that the virtual environment can also handle non-Python software, helping avoid version conflicts (such as for CUDA). In other cases, installing with can be easier with conda (same 'install' command), and avoids needing different details for each separate tool.

Python dependencies

For most Python packages, if the user does not wish to use conda, they can instead use pip. The following packages are listed in the 'req.txt' file that can be used with pip.

pandas
matplotlib
biopython
seaborn
scikit-learn
scipy
fastcluster
Box2d>=2.3.10
tqdm
pyyaml
ipykernel
bokeh
umap-learn
fair-esm
proteinclustertools

The following command can then be used to install.

pip install -r req.txt

PyTorch

Part of the package requires PyTorch. By itself, PyTorch should still run on CPU (slower and may take quite a bit of RAM, requiring lower tokens per batch). But if the user wishes to leverage their GPU, try installing using the pip commands from the PyTorch webtool. Pay attension to the version, which has to match the GPU CUDA or ROCm driver.

Nvidia (CUDA) example:

pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

AMD (ROCm):

pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm6.0

The user also has to install CUDA or ROCm separately.

Other command line tools

There are 3 command line tools used in this package that would need to be installed separately. See the following links for their install instructions.

• MMSeqs - for all-by-all pairwise alignments in the homology based method
• MAFFT - for creating multiple sequence alignments during representative selection
• HMMER - for creating HMM profiles during representative selection