Brewery: prediction of 1D protein structure annotations

The web server, train and test sets of Brewery are available at http://distilldeep.ucd.ie/brewery/.
The docker container is available at https://hub.docker.com/r/mircare/brewery (HOWTO).

The predictions of the UniProtKB entries for COVID-19 are available at http://distilldeep.ucd.ie/brewery/.
See https://github.com/mircare/Porter5 to predict protein secondary structure only.

Diagram of the pipeline we propose to gather and exploit deeper profiles.

Setup

$ git clone https://github.com/mircare/Brewery/ --depth 1 && rm -rf Brewery/.git

Requirements

1. Python3 (https://www.python.org/downloads/);
2. NumPy (https://www.scipy.org/scipylib/download.html);
3. HHblits (https://github.com/soedinglab/hh-suite/);
4. uniprot20 (http://wwwuser.gwdg.de/~compbiol/data/hhsuite/databases/hhsuite_dbs/old-releases/uniprot20_2016_02.tgz).

Optionally (for more accurate predictions):

1. PSI-BLAST (ftp://ftp.ncbi.nlm.nih.gov/blast/executables/blast+/LATEST/);
2. UniRef90 (ftp://ftp.uniprot.org/pub/databases/uniprot/uniref/uniref90/uniref90.fasta.gz).

How to run Brewery with/without PSI-BLAST

# To exploit HHblits only (for fast and accurate predictions)
$ python3 Brewery/Brewery.py -i Brewery/example/2FLGA.fasta --cpu 4 --fast

# To exploit both PSI-BLAST and HHblits (for very accurate predictions)
$ python3 Brewery/Brewery.py -i Brewery/example/2FLGA.fasta --cpu 4

How to run Brewery on multiple sequences

# To split a FASTA file with multiple sequences (Optionally)
$ python3 Brewery/split_fasta.py many_sequences.fasta

# To predict all the fasta files in a given directory (Fastas)
$ python3 Brewery/multiple_fasta.py -i Fastas/ --cpu 4 --fast

# To run multiple predictions in parallel (using a total of 8 cores)
$ python3 Brewery/multiple_fasta.py -i Fastas/ --cpu 4 --parallel 2 --fast

How to visualize the help of Brewery

$ python3 Brewery/Brewery.py --help
usage: Brewery.py [-h] [-input fasta_file] [--cpu CPU] [--fast] [--noSS]
                  [--noTA] [--noSA] [--noCD] [--distill] [--setup]

This is the standalone of Brewery5. Run it on a FASTA file to predict its
Secondary Structure in 3- and 8-classes (Porter5), Solvent Accessibility in 4
classes (PaleAle5), Torsional Angles in 14 classes (Porter+5) and Contact
Density in 4 classes (BrownAle).

optional arguments:
  -h, --help         show this help message and exit
  -input fasta_file  FASTA file containing the protein to predict
  --cpu CPU          How many cores to perform this prediction
  --fast             Use only HHblits (skip PSI-BLAST)
  --bfd              Harness also the BFD database (https://bfd.mmseqs.com/)
  --noSS             Skip Secondary Structure prediction with Porter5
  --noTA             Skip Torsional Angles prediction with Porter+5
  --noSA             Skip Solvent Accessibility prediction with PaleAle5
  --noCD             Skip Contact Density prediction with BrownAle5
  --distill          Generate useful outputs for 3D protein structure prediction
  --setup            Initialize Brewery5 from scratch (it is recommended when
                     there has been any change involving PSI-BLAST, HHblits,
                     Brewery itself, etc).

E.g., run Brewery on 4 cores: python3 Brewery5.py -i example/2FLGA --cpu 4

Use the docker image

# Set the absolute PATHs for databases and query sequences (stored locally)
$ docker run --name brewery -v /**PATH_to_uniprot20_2016_02**:/uniprot20 \
-v /**PATH_to_UniRef90_optional**:/uniref90 -v /**PATH_to_fasta_to_predict**:/Brewery/query \
--cap-add IPC_LOCK mircare/brewery sleep infinity &

# Run a prediction using 5 cores and HHblits only
$ docker exec brewery python3 Brewery.py -i query/2FLGA.fasta --cpu 5 --fast

Performances of Secondary Structure Predictors in 3 classes

Method	Q3 per AA	SOV'99 per AA	Q3 per protein	SOV'99 per protein
Brewery	83.81%	80.41%	84.32%	81.05%
SPIDER 3	83.15%	79.43%	83.42%	79.79%
Brewery HHblits only	83.06%	79.49%	83.68%	80.26%
SSpro 5.1 with templates	82.58%	78.54%	83.94%	80.29%
PSIPRED 4.01	81.88%	77.36%	82.48%	78.22%
RaptorX-Property	81.86%	78.08%	82.57%	78.99%
Porter 4	81.66%	78.05%	82.29%	78.61%
SSpro 5.1 ab initio	81.17%	76.87%	81.10%	76.92%
DeepCNF	81.04%	76.74%	81.16%	76.99%

Reference: Table 1 in https://doi.org/10.1101/289033.

Performances of Secondary Structure Predictors in 8 classes

Method	Q8 per AA	SOV8_refine per AA	Q8 per protein	SOV8_refine per protein
Brewery	73.02%	72.09%	73.92%	72.64%
SSpro 5.1 with templates	71.91%	70.72%	74.46%	73.45%
Brewery HHblits only	71.8%	71.16%	72.83%	71.74%
RaptorX-Property	70.74%	69.65%	71.78%	70.03%
DeepCNF	69.76%	68.5%	70.14%	68.06%
SSpro 5.1 ab initio	68.85%	67.54%	69.27%	67.91%

Reference: Table 2 in https://doi.org/10.1101/289033.

Performances of Solvent Accessibility Predictors in up to 4 classes

Method	Q2 per AA	Q3 per AA	Q4 per AA
ACCpro 5.1 with templates	80.5%	N.A.	N.A.
Brewery	80.48%	66.41%	56.46%
PaleAle 4	78.21%	N.A.	52.53%
SPIDER 3	77.91%	61.19%	49.01%
ACCpro 5.1 ab initio	76.6%	N.A.	N.A.
RaptorX-Property	N.A.	63.25%	N.A.

Performances of Torsion Angles Predictors in 14 classes

Method	Q14 per AA	Q14 per protein
Brewery	69.93%	70.59%
SPIDER 3	66.58%	66.27%
Porter+	64.73%	66%

Performances of Contact Density Predictors in 4 classes

Method	Q4 per AA	Q4 per protein
Brewery	50.01%	48%
BrownAle	46.5%	N.A.

Citation

If you use Brewery, please cite our Bioinformatics paper:

@article{torrisi_brewery_2020,
	title = {Brewery: Deep Learning and deeper profiles for the prediction of 1D protein structure annotations},
	doi = {10.1093/bioinformatics/btaa204},
	journal = {Bioinformatics},
	author = {Torrisi, Mirko and Pollastri, Gianluca}
}

References

Brewery: Deep Learning and deeper profiles for the prediction of 1D protein structure annotations,
Bioinformatics, Oxford University Press; Mirko Torrisi and Gianluca Pollastri;
Guest link: https://academic.oup.com/bioinformatics/advance-article/doi/10.1093/bioinformatics/btaa204/5811232?guestAccessKey=9a73ae2a-2cb6-4fe1-b333-a4f3261f02cf.

Protein Structure Annotations; Essentials of Bioinformatics, Volume I. Springer Nature;
Mirko Torrisi and Gianluca Pollastri; Post-print: https://www.researchgate.net/publication/332048741_Protein_Structure_Annotations.

Deeper Profiles and Cascaded Recurrent and Convolutional Neural Networks for state-of-the-art Protein Secondary Structure Prediction, Scientific Reports, Nature Publishing Group; Mirko Torrisi, Manaz Kaleel and Gianluca Pollastri;
doi: https://doi.org/10.1038/s41598-019-48786-x.

PaleAle 5.0: prediction of protein relative solvent accessibility by deep learning, Amino Acids, Springer
Manaz Kaleel, Mirko Torrisi, Catherine Mooney and Gianluca Pollastri; Guest Link: https://rdcu.be/bNlXS.

License

This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.

Email us at gianluca[dot]pollastri[at]ucd[dot]ie if you wish to use it for purposes not permitted by the CC BY-NC-SA 4.0.