Awesome

[[TOC]]

VESPA -- Variant Effect Score Prediction without Alignments

VESPA is a simple, yet powerful Single Amino Acid Variant (SAV) effect predictor based on embeddings of the Protein Language Model ProtT5.

The single-sequence-based SAV effect prediction is set up in a multistage pipeline that includes (1) generating ProtT5 embeddings, (2) extracting per-residue conservation predictions, (3) (optionally) extracting per-variant log-odds ratios, and (4) predicting the effect of all possible amino acid substitutions. Step (4) can be completed by either using VESPA with (2) and (3) as input, or by using the computationally more efficient method VESPA-light (VESPAl) with only step (2) as input for a small drop in prediction performance.

The specifics of VESPA and VESPAl can be found in our paper, Embeddings from protein language models predict conservation and variant effects (Marquet et al., 2021). The performance of VESPA when evaluated against SOTA methods can be seen below.

Also check out MutAlign, a tool to visually compare VESPA mutation effect predictions for highly similar sequences (e.g. the same protein in different organisms).

Precomputed VESPA and VESPAl predictions

Precomputed VESPA and VESPAl predictions are currently available for 39 DMS experiments here: Supplementary file 3 of (Marquet et al., 2021). Furthermore, VESPAl predictions are available for the human proteome (downloaded 22/01/17), and for the fly (drosophila melanogaster) proteome (downloaded 22/03/01).

Usage

The preferred method to install VESPA is via pip:

pip install vespa-effect

Input Files

Required: A single FASTA file containing all your wildtype sequences (note: this file can contain any number of sequences).

Optional: If you are only interested in a subset of possible mutations (specific mutations), you can add -m mutations.txt to the code line in Quickstart (note: per default all mutations are considered). Click here to head to the file format explanation of mutations.txt.

Quickstart

For simplicity of this guide, we will assume a folder containing all data: f.e., the FASTA file is placed at data/sequences.fasta and the (optional) mutations.txt at data/mutations.txt.

After installing the repository, you can run the following:

vespa data/sequences.fasta --prott5_weights_cache data/cache

By default, this runs VESPAl for all possible mutations. It will generate a new folder vespa_run_directory in your current working directory. Within this folder you will find two files containing input features (embeddings, conservation prediction) and an output folder with .csv files containing VESPAl predictions for each sequence with all possible mutations. More details can be found under output.

WARNING Creating embeddings requires a powerful GPU (we recommend at least 12GB of VRAM). The same applies to running VESPA. For more details, see Step 1 on generating embeddings and Step 3 on extracting log-odds ratios.

Optional:

• If you already have precomputed embeddings available, you can use those instead of generating new ones by adding --use_existing_embeddings followed by the file location. Then running VESPAl will not require a GPU.
• If you have a GPU available you can generate VESPA predictions by adding --vespa to the code line above. After running VESPA you will find another file in your folder vespa_run_directory containing log-odds ratios as additional input.
• --prott5_weights_cache followed by a folder path specifies the location of a caching folder (or preloaded ProtT5 weigths). Instead of downloading the weights for every run of VESPA and/or VESPAl, the weights will be downloaded into this folder or just reused if already present.
• In case you want to output a single [.h5 file](#.h5 files) in addition to the default .csv files with predictions, add --h5_output followed by the output file location. If you are only interested in an .h5 file, add --no_csv.

Below you can find information on how to run each step of VESPA and VESPAl individually. Running the vespa script will automatically the substeps below for you (with the optional specifications you of your choice). Use the substeps if you are interested in a particular intermediate result.

The substeps of VESPA and VESPAl

The following steps can be run individually and for any number of sequences contained in a FASTA file.

Step 1: Extracting ProtT5 embeddings

To run VESPA and/or VESPAl to obtain SAV predictions, you will need the ProtT5 embeddings of your sequences. If you have a GPU, you can use the included embedding script on your own machine to generate your protein embeddings:

vespa_emb data/sequences.fasta -o data/embeddings.h5  --prott5_weights_cache data/cache

As you can see, you need to specify an output file location (-o), and the FASTA file location of your input sequences.

Optional: --prott5_weights_cache specifies a caching directory to store ProtT5 weights. If not present, the line above will download the weights to your current working directory for every run of VESPA and/or VESPAl.

Step 2: Conservation Prediction

VESPA and VESPAl take per-residue conservation predictions as input. To generate them, run the following (in the VESPA folder):

vespa_conspred data/embeddings.h5 -o data/conspred.h5

The input for our conservation predictor are the ProtT5 embeddings. In your output file (-o), you will find 9-state conservation probabilities (per-residue) needed as input for the models. For more details on the conservation prediction please see the paper mentioned above.

Optional: In case you are interested in generating a file that contains the predicted conservation classes instead of the assigned class probabilities, add --output_classes above.

Step 3: Log-odds ratio of masked marginal probabilities

This step requires a GPU! The two models VESPA and VESPAl differ in their predictive performance but also in the required input. If you are only interested in running VESPAl, you can skip this step.

To generate the log-odds ratios of masked marginal probabilities required for VESPA, run (in the VESPA folder):

vespa_logodds data/sequences.fasta -o data/logodds.h5 

Optional:

• To run the log-odds script including the mutations file, add -m data/mutations.txt to the code line above.
• Per default the command above generates the .h5 file required for the subsequent steps. We provide two options to output a human readable version of the log-odds scores by adding --single_csv data/single_logodds.csv or --csv_dir data/csv_dir/ at the end of the code line above. The first option generates a single csv file with all sequences and SAVs. For large sets, we recommend to use the second option, which outputs multiple csv files separated by sequence ID of the given FASTA file. The format of the .csv files is described below.

Step 4: Run VESPA and/or VESPAl

Now you have all the data required to run VESPA and/or VESPAl. Per default, the vespa script will predict SAV effects for VESPAl only. To explicitly enable VESPA add --vespa (which will run both models), and to disable VESPAl add --no-vespal. To generate predictions, execute the following code (in the VESPA folder):

• Both (if you computed the conservation prediction and the log-odds):

  vespa_run --vespa data/conspred_probs.h5 data/sequences.fasta --T5_input data/logodds.h5 --output predictions/
  

• Only VESPA (if you computed the conservation prediction and the log-odds):

  vespa_run --vespa --no-vespal data/conspred_probs.h5 data/sequences.fasta -T5_input data/logodds.h5 --output predictions/
  

• Only VESPAl (if you only computed the conservation prediction):

  vespa_run data/conspred_probs.h5 data/sequences.fasta --output predictions/
  

Optional: In case you want to output a single [.h5 file](#.h5 files) in addition to the default .csv files with predictions, add --h5_output followed by the output location. If you are only interested in an .h5 file, add --no_csv and drop --output predictions/. More details can be found under output.

Note: The format of the results file is described below at VESPA and VESPAl output.

Additional Information

Extracting raw reconstruction probabilities

You might be interested in extracting the raw reconstruction probabilities for each mutation position from ProtT5. These raw reconstruction probabilities are used to calculate the log-odds ratio and is explained in more detail in the corresponding publication. To do so, use:

vespa_logodds -r --reconstruction_output data/reconstruction_probas data/sequences.fasta -o data/logodds.h5 

Optional: Add a mutations file.

The generated datasets in the .h5 file contain probability vectors that determine the reconstruction probabilities for all amino acids sorted according to the MUTATION_ORDER in config.py. If a particular mutation was not computed (i.e., the position was not present in the mutation.txt) it contains -1.

File Specifications

This section describes relevant file formats for VESPA and VESPAl:

Mutations file

A simple text file with a protein ID and one mutation per-line (i.e., mutations separated by \n for newline). Every mutation should be specified by <PROTEIN_ID>_<SAV-String> separated by an underscore. The default <SAV-String> has the following format <fromAA><0-basedSequencePosition><toAA>.

The sequence ID needs to be equivalent to the one in the sequence FASTA file. The SAV string has the format: <Original Amino Acid><Position><Replacement Amino Acid>

Example:

ENSP00000355206_I0L
ENSP00000355206_I0V
ENSP00000355206_I0L
ENSP00000355206_I0K
ENSP00000355206_I0T

Optional: In case you have a file with 1-based mutations, add the flag --one_based_mutations.

.h5 files

Multiple scripts generate .h5 files. These files follow the hdf5-standard and can be processed in python using the library h5py. Generally, the files are segmented into datasets that can be accessed using the protein accession in the FASTA file. The datasets are explained in more detail in the respective output sections.

Conservation output

The conservation prediction will output an .h5 file with the predicted 9-state conservation probabilities. For each ID of the FASTA file, the respective dataset contains a matrix of size 9xL, with L being the length of the protein sequence, and 9 being the predicted conservation class (index 0 = very variable; index 8 = very conserved).

Additionally, you can extract a file with per-residue conservation class predictions for each sequence directly.

Log-odds ratio output

The default output is an .h5 file that contains one Lx20 matrix per sequence. -1 means no log-odds ratio was computed for the respective mutation.

When running --single_csv data/single_logodds.csv, the log-odds ratio output file contains all mutations (or the mutations determined by the mutations file) as <PROTEIN_ID>_<SAV-String>, followed by a ; and the log-odds ratio for a particular mutation per line.

Example:

B3VI55_LIPSTSTABLE_A438Q;0.053829677402973175
B3VI55_LIPSTSTABLE_A438N;0.061238136142492294
B3VI55_LIPSTSTABLE_A438Y;0.012603843584656715
B3VI55_LIPSTSTABLE_A438M;0.012212143279612064
B3VI55_LIPSTSTABLE_A438H;0.0241163931787014
B3VI55_LIPSTSTABLE_A438W;0.004520446062088013

In case the predictions are written into a directory, e.g. by specifying --csv_dir in Step 3, the script will create one file per sequence, named by sequence ID (note: the ID will be normalized, i.e, each special character will be replaced by _). The file will contain one <Mutation-String>;score per-line.

Example file called B3VI55_LIPSTSTABLE:

A438Q;0.053829677402973175
A438N;0.061238136142492294
A438Y;0.012603843584656715
A438M;0.012212143279612064
A438H;0.0241163931787014
A438W;0.004520446062088013

VESPA and VESPAl output

The default output of VESPA and/or VESPAl will generate one .csv (with semicolon separator) file per protein in the output folder of vespa_run_directory. If you specified an output directory with --output, you will find the .csv (with semicolon separator) files there. To circumvent naming issues due to long sequence ID's, the csv files will be numbered by sequence occurrence in the FASTA file. A lookup file map.json will be created in the output directory containing a dictionary mapping from number to sequence ID.

Example map.json:

{    "0": "B3VI55_LIPSTSTABLE",
    "1": "BF520_ENV",
    "2": "BG_STRSQ",
    "3": "BG505_ENV",
    "4": "HG_FLU",
    "5": "MTH3_HAEAESTABILIZED"}

The individual files will contain rows with the mutations (format: <fromAA><0-basedSequencePosition><toAA>) along with the respective predictions of VESPA and/or VESPAl.

Example 0.csv:

Mutant;VESPAl;VESPA
M0A;0.4457732174287125;0.3520255108578212
M0L;0.3191178420567241;0.2717188481387661
M0G;0.5355136080284415;0.4110670843315182
M0V;0.3594337197937546;0.2971971641898669
M0S;0.4457732174287125;0.35202555423053
M0R;0.4457732174287125;0.35202621644931126

If you chose to generate an .h5 output, the ID's will be the original FASTA ID's. The file contains one 20xL-shaped matrix per sequence-ID. L is the length of the respective protein length, and 20 is the possible number of amino acid variants (including self). The order of amino acids is determined by MUTANT_ORDER in predict/config.py (ALGVSREDTIPKFQNYMHWC). Empty fields that were not calculated/specified (f.e., wildtype substitutions) contain a -1.

Development Roadmap

• Write comprehensive tests
• Publish pypi package
• Install from github release
• Contributing guidelines

Installation from current Github Release

WARNING Experimental: To install the current release from github you can use:

python -m pip install https://github.com/Rostlab/VESPA/releases/download/v0.9.0-beta/vespa-0.9.0b0.tar.gz

Cite

If you want to credit us, feel free to cite

Marquet, C., Heinzinger, M., Olenyi, T. et al. Embeddings from protein language models predict conservation and variant effects. Hum Genet (2021). https://doi.org/10.1007/s00439-021-02411-y

@article{Marquet2021,
  doi = {10.1007/s00439-021-02411-y},
  url = {https://doi.org/10.1007/s00439-021-02411-y},
  year = {2021},
  month = dec,
  publisher = {Springer Science and Business Media {LLC}},
  author = {C{\'{e}}line Marquet and Michael Heinzinger and Tobias Olenyi and Christian Dallago and Kyra Erckert and Michael Bernhofer and Dmitrii Nechaev and Burkhard Rost},
  title = {Embeddings from protein language models predict conservation and variant effects},
  journal = {Human Genetics}
}