protein-detective

Python package to detect proteins in EM density maps.

Have you ever had an coarse Electron Microscopy (EM) density map and was unable to identify part of it? This package can help you to identify the protein in the density map by searching for protein structures in Uniprot, PDBe and AlphaFold DB and fitting them into the density map.

It uses

• protein-quest to search, retrieve and filter protein structures from Uniprot, PDBe and AlphaFold DB.
• powerfit to fit protein structure in a Electron Microscopy (EM) density map.
• cyclopts for command line interface
• molviewspec to visualize the fitted structures and density map in a web browser.
• dask to run powerfit in parallel on multiple CPU cores or GPUs.
• Research Object Crate and rocrate-action-recorder to keep track of commands and their input/output files/directories.
• duckdb to query CSV files like powerfit/*/*/solutions.out files.

Diagram how protein-detective calls protein-quest and powerfit:

flowchart TB
    subgraph search [protein-detective search]
        direction TB
        S2[protein-quest search uniprot]
        S3[protein-quest search alphafold]
        S4[protein-quest search pdbe]
        S5[protein-quest search pdbe-quality]
        S6[protein-quest search complexes]
        S2 --> S3 & S4
        S4 --> S5
        S2 -.-> S6
        S6 -.-> S4
    end
    search -- "UniProt accessions & PDB ids" --> retrieve

    subgraph retrieve [protein-detective retrieve]
        direction TB
        R2[protein-quest retrieve pdbe]
        R3[protein-quest retrieve alphafold]
    end
    retrieve -- "mmcif_files" --> filter

    subgraph filter [protein-detective filter]
        direction TB
        F2[protein-quest convert structures --uniprots]
        F3[protein-quest filter chain]
        F4[protein-quest filter combined]
        F5[protein-quest filter secondary-structure]
        F2 --> F3 --> F4 -.-> F5
    end
    filter -- "fewer mmcif_files" --> P1[protein-detective powerfit run]

    I1[protein-detective import-structures]
    I1 -. "mmcif files" .-> P1

    E1[protein-detective powerfit report]
    M1[protein-detective powerfit fit-models]
    P1 -- "**/solutions.out" --> E1 & M1

    classDef dashedBorder stroke-dasharray: 5 5;
    S6:::dashedBorder
    F5:::dashedBorder
    I1:::dashedBorder

(Dashed nodes are optional) (The Mermaid figure might not be rendered, see https://www.bonvinlab.org/protein-detective/ for rendered version)

Install

pip install protein-detective

Or to use the latest development version:

pip install git+https://github.com/haddocking/protein-detective.git

By default OpenCL support is included, but if you want to use CUDA, you can install with:

# For CUDA version 13
pip install "protein-detective[cuda13]"
# or for CUDA version 12
pip install "protein-detective[cuda12]"

Usage

The main entry point is the protein-detective command line tool which has multiple subcommands to perform actions.

Search Uniprot for structures

protein-detective search \
    --taxon-id 9606 \
    --reviewed \
    --subcellular-location-uniprot nucleus \
    --subcellular-location-go GO:0005634 \
    --molecular-function-go GO:0003677 \
    --limit-uniprot 100 \
    --pdbe.limit 100 \
    ./mysession

(GO:0005634 is "Nucleus" and GO:0003677 is "DNA binding")

In ./mysession directory, you will find the search results.

protein-detective search --verbose \
    --taxon-id 9606 \
    --reviewed \
    --subcellular-location-uniprot nucleus \
    --subcellular-location-go GO:0005634 \
    --molecular-function-go GO:0003677 \
    --interaction.seed A8MT69 \
    --interaction.exclude Q9BTP7 \
    --limit-uniprot 100 \
    --pdbe.limit 100 \
    ./mysession2

To retrieve a bunch of structures

protein-detective retrieve ./mysession

In ./mysession directory, you will find mmCIF files from PDBe and PDB files and AlphaFold DB.

To filter structure

Filter structures based on

• For PDBe structures the chain of Uniprot protein is written as chain A.
• For AlphaFold structures filter by confidence (pLDDT) threshold
• Number of residues in chain A
  • For AlphaFold structures writes new files with low confidence residues (below threshold) removed
• Number of residues in secondary structure (helices and sheets)

Also uncompresses *.cif.gz files to *.cif files for compatibility with powerfit.

protein-detective filter \
    --min-confidence 50 \
    --min-residues 100 \
    --max-residues 1000 \
    ./mysession

# or to filter on secondary structure having some helices
protein-detective filter mysession --secondary.abs-min-helix-residues 40

Import filtered structures

If you have a directory of structures ((optionally gzipped) PDB/mmCIF files), each with a single chain called A and a single UniProt accession. You can import them into a new protein detective session with:

protein-detective import-structures ./mysession/filtered ./mysession3

Imported structures can be used to run powerfit.

Powerfit

Rotate and translate the prepared structures to fit and score them into the EM density map using powerfit.

protein-detective powerfit run ../powerfit-tutorial/ribosome-KsgA.map 13 ./mysession
# or for with some flags
protein-detective powerfit run --workers-per-gpu 2 --angle 40 --powerfit-run-id myrun1 ../powerfit-tutorial/ribosome-KsgA.map 13 ./mysession4

This will use dask-distributed to run powerfit for each structure in parallel on multiple CPU cores or GPUs.

pip install dask-jobqueue
python3

from dask_jobqueue import SLURMCluster

cluster = SLURMCluster(cores=8, processes=4, memory="16GB", queue="normal")
print(cluster.scheduler_address)
# Prints something like: 'tcp://192.168.1.1:34059'
# Keep this Python process running until powerfits are done

protein-detective powerfit run ../powerfit-tutorial/ribosome-KsgA.map 13 docs/session1 --scheduler-address tcp://192.168.1.1:34059

protein-detective powerfit run --batch-size 50 --gpu-backend cuda ../powerfit-tutorial/ribosome-KsgA.map 13 ./mysession

protein-detective powerfit commands ../powerfit-tutorial/ribosome-KsgA.map 13 docs/session1 > commands.txt
parallel --jobs 4 < commands.txt

To list the completed powerfit runs, you can use:

protein-detective powerfit list-runs mysession

Which will output something like:

powerfit_run_id,density_map,run_dir,options
myrun1,mysession/powerfit/myrun1/ribosome-KsgA.map,mysession/powerfit/myrun1,--workers-per-gpu 2 --angle 40 --powerfit-run-id myrun1 ../powerfit-tutorial/ribosome-KsgA.map 13 ./mysession

To print top 1 solution per template structure to the terminal, you can use:

protein-detective powerfit report mysession

Outputs something like:

powerfit_run_id,structure,rank,cc,fishz,relz,translation,rotation,template_file,uniprot_accessions,structure_id,is_alphafold
myrun1,3i8z_updated_A2A.cif.gz,1,0.598,0.69,12.959,239.46:187.27:211.83,-0.238:0.322:0.916:0.916:-0.238:0.322:0.322:0.916:-0.238,mysession/combined_output/3i8z_updated_A2A.cif.gz,O00257,3I8Z,False
myrun1,6mzc_updated_E2A.cif.gz,1,0.547,0.614,14.671,199.55:214.9:165.78,1.0:0.0:0.0:0.0:-1.0:0.0:0.0:0.0:-1.0,mysession/combined_output/6mzc_updated_E2A.cif.gz,O00268,6MZC,False
...

To generate model PDB files rotated/translated to top 1 solution per template structure, you can use:

protein-detective powerfit fit-models mysession

Outputs something like:

powerfit_run_id,structure,rank,fitted_model_file,unfitted_model_file
myrun1,3i8z_updated_A2A.cif.gz,1,mysession/powerfit/myrun1/3i8z_updated_A2A.cif.gz/fit_1.pdb,mysession/combined_output/3i8z_updated_A2A.cif.gz
myrun1,6mzc_updated_E2A.cif.gz,1,mysession/powerfit/myrun1/6mzc_updated_E2A.cif.gz/fit_1.pdb,mysession/combined_output/6mzc_updated_E2A.cif.gz
...

Where the fitted_model_file column is the structure file of the fitted model and unfitted_model_file is the original structure file.

The results can also be visualized see visualization.ipynb for an example.

Contributing

For development information and contribution guidelines, please see CONTRIBUTING.md.