Skip to content

Latest commit

 

History

History
172 lines (146 loc) · 4.68 KB

prediction.md

File metadata and controls

172 lines (146 loc) · 4.68 KB
Raw

Prediction

Requirements:

  • DSSP docker image: 
    git clone https://github.com/cmbi/dssp ~/dssp
pushd ~/dssp
git checkout 697deab74011bfbd55891e9b8d5d47b8e4ef0e38
docker build -t dssp .
popd
  • Uniref50 database: 
    wget -q 'ftp://ftp.uniprot.org/pub/databases/uniprot/uniref/uniref50/uniref50.fasta.gz' 
gzip --decompress 'uniref50.fasta.gz'
rm 'uniref50.fasta.gz'
  • A GraphQA checkpoint, for example this checkpoint

Prediction

Input files:

  • Sequences in Fasta format in file called sequences.fasta: 
    >T1234
ACKIENI.......RFDLLTR

>T5678
NDVKKQQ.......GAPVPKQ
  • Decoys in PDB format in a decoys folder with this structure (note the .pdb extension): 
    decoys
├── T1234
│   ├── first_decoy.pdb
│   ├── another_decoy.pdb
│   └── third_decoy.pdb
└── T5678
    ├── first_decoy.pdb
    ├── another_decoy.pdb
    └── third_decoy.pdb

Preprocess the decoys:

UNIREF_50='path/to/uniref50.fasta'
python -m graphqa.data.preprocess . "${UNIREF_50}"

2020-11-19 11:05:30 | INFO | graphqa.data.dssp:run_dssp:239 - Running on 6 pdb files
2020-11-19 11:05:32 | INFO | graphqa.data.msa:run_msa:146 - Running on 2 sequence(s)
2020-11-19 11:18:12 | INFO | __main__:main:121 - Processing T1234
2020-11-19 11:18:12 | INFO | __main__:main:121 - Processing T5678

This will create the following files:

decoys
├── T1234
│   ├── first_decoy.pdb
│   ├── first_decoy.dssp
│   ├── another_decoy.pdb
│   ├── another_decoy.dssp
│   └── third_decoy.pdb
│   └── third_decoy.dssp
└── T5678
  ├── first_decoy.pdb
  ├── first_decoy.dssp
  ├── another_decoy.pdb
  ├── another_decoy.dssp
  └── third_decoy.pdb
  └── third_decoy.dssp

processed
├── T1234.pth
└── T5678.pth

Run GraphQA:

STAGE=1
CHECKPOINT='path/to/graphqa.ckpt'
python -m graphqa.eval "${CHECKPOINT}" . --stage "${STAGE}"

2020-11-19 11:21:59 | DEBUG | graphqa.dataset:__init__:25 - Starting to load graphs from 2 pth files
2020-11-19 11:21:59 | DEBUG | graphqa.dataset:__init__:69 - Done loading 6 graphs, skipped 0

This will create the following files:

predictions
├── global.csv
├── local.csv
├── T1234.stage1.qa
└── T5678.stage1.qa

The file global.csv contains all decoy-level predictions in this format:

target_id decoy_id tm gdtts gdtha lddt cad
T1234 first_decoy 0.53 0.45 0.29 0.46 0.31
T1234 another_decoy 0.29 0.46 0.31 0.53 0.45
... ... ... ... ... ... ...

The file local.csv contains all residue-level predictions in this format:

target_id decoy_id residue_idx lddt cad
T1234 first_decoy 0 0.47 0.26
T1234 first_decoy 1 0.54 0.39
T1234 first_decoy 2 0.50 0.29
... ... ... ... ...
T5678 third_decoy 268 0.47 0.26
T5678 third_decoy 269 0.54 0.39
T5678 third_decoy 270 0.50 0.29

Note that residue_idx is 0-indexed.

CASP emails

CASP emails come in a standard format that can be easily processed.

We assume the file emails.txt contains some emails in this format:

TARGET=T1031
SEQUENCE=ACKIE...LLTR
[email protected]
STOICHIOMETRY=A1
TARBALL_LOCATION=http://predictioncenter.org/.../T1031.stage1.3D.srv.tar.gz

TARGET=T1032
SEQUENCE=NDVKKQ...LRGMVFGAPVPKQ
[email protected]
STOICHIOMETRY=A2
TARBALL_LOCATION=http://predictioncenter.org/.../T1032.stage1.3D.srv.tar.gz

Download all decoys and extract the tar files:

grep 'TARBALL_LOCATION=' 'emails.txt' | while read -r line ; do
  curl -L "${line/TARBALL_LOCATION=/}" | tar xz --directory 'decoys'
done

Add .pdb to each decoy file:

for f in decoys/*/*; do
  mv "$f" "${f%.pdb}.pdb"
done

Extract sequences in fasta format:

cat 'emails.txt' |
  sed '/REPLY_EMAIL=/d' | 
  sed '/STOICHIOMETRY=/d' | 
  sed '/TARBALL_LOCATION=/d' | 
  sed 's/TARGET=/>/' | 
  sed 's/SEQUENCE=//' > sequences.fasta

Run GraphQA:

conda activate proteins

STAGE=1
CHECKPOINT='path/to/graphqa.ckpt'
UNIREF_50='path/to/uniref50.fasta'

python -m graphqa.data.preprocess . "${UNIREF_50}"
python -m graphqa.eval "${CHECKPOINT}" . --stage "${STAGE}"