/protein-vae

Variational autoencoder for protein sequences - add metal binding sites and generate sequences for novel topologies

Primary LanguagePythonMIT LicenseMIT

Variational autoencoder for protein sequences

This repository provides code to accompany the paper:

Greener JG, Moffat L and Jones DT, Design of metalloproteins and novel protein folds using variational autoencoders, Scientific Reports 8:16189, 2018 - link

The work describes a variational autoencoder that can add metal binding sites to protein sequences, or generate protein sequences for a given protein topology.

Getting Started

These instructions will get a copy of this repo on your system and running so you can produce sequences with our models or expand on our work. These files have been cleaned and minimized so they can be run using python with as few dependencies as possible.

Dependencies

The follow are the packages needed to run our software. Our software uses python and leverages several python packages:

Follow the links to install the correct version of all of the packages. Note that installing pytorch can be a bit tricky but all of these packages can be easily install with Anaconda. Using the conda install command from the Anaconda package is highly recommended. The lark-parser package is only needed if you wish to generate protein sequences using the model that takes in a structure grammar to produce sequences (described more below). The neural networks are built using pytorch and some utility functions from sklearn are used for metrics, making proper installation critical.

If you wish to extend our work note that we have configured the general use scripts to run solely on CPU as they do not require the computational power of a GPU. However, the example training scripts provided can be run on GPU and contain switches and/or command line arguments to enable this.

Installation

Navigate to the directory you wish to store the software in and simply run the following from the command line.

user@computer:~$ git clone git@github.com:psipred/protein-vae.git
user@computer:~$ cd protein-vae/produce_sequences/

You are now in the directory to run the scripts to produce sequences. Producing sequences is described below.

Running the software

Before using our software please read the paper as linked to at the top of these instructions. There are three different methods of producing sequences that this software provides:

  1. Seq-to-Seq: Providing an initial sequence to return a similar sequence with some variation.
  2. Seq-to-Metal-Seq: Providing an initial sequence that does not bind to a metal and returning a the same sequence with variation that is more likely to bind a specified metal.
  3. Grammar-to-Seq: Providing a grammar string (see the paper) and producing a sequence that is likely to fold to the topology described in the grammar string.

Each one of these methods is run with its own python script. Before going any further make sure you are in the produce_sequences directory. Below we have provided examples for running each one of the scripts and what the input and output should look like. We have also provided example sequences (as found in the examples directory).

If you run one of the scripts without providing a sequence it will default use one of the example sequences. Below are examples of how to use each one of the scripts with the default examples.

1. Seq-to-Seq Example

1. Input File Preparation

Make sure you have a text file or fasta file with a single sequence in it. For example, in the seq2seq_example.txt you'll find the following:

AEVPSGEQLFNSNCSACHIGGNNVIISHKTLRKEALEKYAMNSLEAIRYQVVNGKNAMPAFGGRLNEEEIDAIATYVLGQAELD--------------------------------------------------------

Only provide one sequence in a given file (you can still output multiple based on the one). If you wish you can pad the sequence up to 140 characters long with a '-' symbol as above however this is not necessary (see the Seq-to-Metal-Seq example). Sequences longer than 140 will be chopped. If you do not provide a sequence the default example will be used.

2. Run the Script

Run the following command to produce sequences from the input file (the default has been used):

user@computer:~$ python seq_to_seq.py -infile examples/seq2seq_example.txt -numout 10

The -numout argument is an integer input for the number of sequences you want to produce. Here we are producing 10 examples. If you do not provide values for the two command line arguments the script defaults to producing 10 sequences from the default example.

3. Output Example

Running the script in step 2 outputs 10 sequences and provides the average sequence identity between the sequences produced and the original input sequence. This is outputed to stdout. Here is an example:

Average Sequence Identity to Input: 60.4%
ADLEAGEQIFSANCAACHGGGNNIIMPEKTLKKDALEENGMKSVEAITYQVTNGKNAMPAFGGRLSDEDIEDVANYVLSQAEKGW
ADLEHGAQIFSANCAACHAGGNNVIMPDKTLKKDALEKNGMNSIEAITYQVTNGKNAMPAFGGRLSDEDIEDVANYVLSQAEKGW
ADLENGGKVFSGACAACHIGGENIVRPEKTLKKDALEEGGMDSIEAITAQVTNGKNAAPAFGERLVDEDIEDVAEYVL
ADLAAGEQIFSANCAACHAGGNNVVMPDKTLKKDALEKYGMNSIEAITTQVTNGKNAMPAFGGRLEAEDIEDVAAYVLSQAEG
ADLEHGEQIFSANCAACHAGGNNVIMPEKTLKKDALEKYGMNSVEAITTQVTNGKNAMPAFGGRLEDEQIEDVANYVLSQSEW
ADIEHGEKIFSANCAACHAGGNNAIMRNKTLKKEALEPNGMNSIEAITYQVTNGKNAMPAFGGRLSDEDIEDVANYVLKQAEKGW
ADLAAGEQIFSANCAACHAGGNNIIMPEKTLKKEALEKYSMNSIEAITTQVTNGKNAAPAFGGRLSDEDIEDVANYVLSQAEKGW
ADIITGEQIFSANCAACHIGGNNAIRPEKTLKKPALETNGMNSVDAITTQVVNPKNAMPAFGGRLEDEDIEDVANYVLSQAEK
GDLEKGKGIFKFNCVACHSNGKNVIIIEKTLKKDALKANGMFSIDAITSQIANGKNAMPAFAGRLKDDLIELVAYYVLEKAEQW
ADLANGAKIFSANCAACHAGGGNAIMPTKTLKKNALEKNGMNSIEAITYQVTNGKNAMPAFKGRLSEEDIEDVAAYVLEQSEKGW

2. Seq-to-Metal-Seq

1. Input File Preparation

Note this is very similar to Seq-to-Seq example. Make sure you have a text file or fasta file with a single sequence in it. For example, in the seq2metalseq_example.txt you'll find the following:

DTDSEEEIKEAFKVFDKDGNGYISAAELRHVMTNLGEKLSDNEVDEMIREADVDGDGQINYEEFVKMMLSK

Only provide one sequence in a given file (you can still output multiple based on the one). If you wish you can pad the sequence up to 140 characters long with a '-' symbol however this is not necessary. Sequences longer than 140 will be chopped. If you do not provide a sequence the default example will be used.

You also need to decide which metal you want the model to try and insert a binding site for. For example, if you choose Iron it will produce sequences more likely to bind the metal specified. The 8 metals you can choose from are:

  • Fe
  • Zn
  • Ca
  • Na
  • Cu
  • Mg
  • Cd
  • Ni

2. Run the Script

Run the following command to produce sequences from the input file (the default has been used)

user@computer:~$ python seq_to_metalseq.py -infile examples/seq2metalseq_example.txt -numout 10 -metal Fe

The -numout argument is an integer input for the number of sequences you want to produce. Here we are producing 10 examples. The -metal argument is the two letter atomic code (one of the above 8) of the one metal you wish to use. If you do not provide values for the three command line arguments the script defaults to producing 10 sequences from the default example inserting Fe binding.

3. Output Example

Running the script in step 2 outputs 10 sequences and provides the average sequence identity between the sequences produced and the original input sequence. This is outputed to stdout. Here is an example:

Average Sequence Identity to Input: 80.1%
DTDREEEIREAFRVFDKDGNGFISAAELRHVMTNLGEKLTDEEVDEMIREADIDGDGQVNYEEFVKMMEAK
MTDTEEEIDEAFRVFDKDGNGYDSAAELRHVMTNLGEKLTDEEVDEMIREADIAGDGQVNYEEFVTMMTAK
DTSSEEEIDEAFRVFDKDGNGFISAAELRHVMTNLGEKLTDEEVDEMIREADNAGDGQDNYEEFVTMMTVK
DTDEEEKIREAFRDFDKSDNEFDSAAELRHVMTAGGEKLTDEEVDEMIDGADMDDDGQDFDEEFDGMMTAK
DSDTEEEIKEAFRVFDKDGNGYISAAELRHVMTNVGEKLTDEEVDEMIREADIDGDGQVNYEEFVVMMTAK
DTDSEEEIREAFRVFDKDGNGFISAAELRHVMTNLGEKLTDEEVDEMIREADIAGDGQVNYEEFVKMMTAK
DTDSEEEIREAFRVFDKSDNGFISAAELRHLMTNLGEKLTDEEVDEMIREADIDGDGQINYEEFVKMMLAK
DTDREEEIREAFRVFKKSGNELISAAELRHVMTPLGEKLTDEEVDEMIREAIIDGDGQVNYEEFVGMMKDK
DTDSEEEIREAFRVFDKDGNGFISAAELRHVMTNLGEKLTDEEVDEMIREADIAGDGQVNYEEFVGMMTAK
DTDSESELKEAFRVADKDRNGPDSACKLRHVMLNGIEKLTDKEVDEMIREADIAEDGQVNYEEFVMT

3. Grammar to Seq

1. Input File Preparation

Make sure you have a text file that contains a single grammar string in it as defined in the linked paper. For example, in the gram2seq_example.txt you'll find the following:

+B+0-C+0+B+2-B+1

Only provide one grammar string in a given file (you can still output multiple sequences based on the one grammar string). If you do not provide a grammar string the default example will be used.

2. Run the Script

Run the following command to produce sequences from the input file (the default has been used)

user@computer:~$ python gram_to_seq.py -infile examples/gram2seq_example.txt -numout 10

The -numout argument is an integer input for the number of sequences you want to produce. Here we are producing 10 examples. If you do not provide values for the two command line arguments the script defaults to producing 10 sequences from the default example.

3. Output Example

Running the script in step 2 outputs 10 sequences and provides these outputed to stdout. Here is an example:

ESGYAVVCDTTCSYDGECNNECTCCCLKVKQKGNDGGYCWLWECGCLCLGAPVLVPEDTKCK
KKGCLVSRGTGCGSGCSNNNCAKGLKISNGAKGKEGHRGYKCGCGCFCWPDR
CDGYLVESKTGCGFCGLNNSCCNLCCNKNGAKAGYCACGYKCKCECLPLPLPN
RDGYPVHDKGCKISCFGNNYCWKECKKKGKSKGYCYCWWLACWCYGLPDPEKVWDYA
KKGYPVVSDDCCKYCCLNNKYCNYCCNKCGAKSGYCAWCCKSGCACWCLDLPK
ERDGYIADPTNCGYTCANNSCCNGLCTKNGAKAGYCAWIGPYGKACWCIPLPDKVP
KDYYPKDDKTCCSCCFNNNYCNKECKKEGKASGYCYGWCPACWCWCLPDDE
KKGKYINDGTNCKYTCANNAKNNCCDKKCGAKGGYGHWGYPFGKACWCFPLPE
KRGYLVVKNTNCKYSCFNLGYCNYCCTKCGAKSGYCSWGYCYGNACWCKPLPDKVPIRPPGKC
DRGYLVVSDTGCKYVCYNNSYNKYCDRKCKNKAEYYGFGWLFGYGCWCLPLPEPVWIKIVDC

The grammar is specified using the Lark package and can be found here.

Editing the Software

If you'd like to edit the software we have provided training scripts that are self contained, along with trained models that can be loaded in. The directories metal_gen and fold_gen contain the scripts for experiments involving generating metal binding proteins and experiments producing sequences for a given fold, respectively. These scripts also contain the ability to perform inference however this must be specified by altering the script. Instructions are contained with the script.

Datasets

The datasets used for training are available as numpy binaries to be read in with numpy.load:

  • assembled_data_mb.npy is the metal binding training data, size (147842, 3088). Each row is an example consisting of 22x140=3080 values for the sequence and 8 values for the metal binding flags (order as above).
  • assembled_data_fold.npy is the topology training data, size (104845, 4353). Each row is an example consisting of 22x140=3080 values for the sequence, 23x55=1265 values for the topology encoding and 8 values for the metal binding flags (order as above).

The assignments of Taylor topology strings to SCOP folds we used can be found in topology_data.