Cavity analysis

Cluster an MD trajectory (or accept pre-computed representative structures) and run Fpocket cavity detection + filtering on each representative structure.

Description

The input can be either (1) a trajectory and topology or (2) a folder of representative structures from an external clustering. With a trajectory, the workflow clusters it to find representative structures. With structures, it uses them directly.

The workflow then runs Fpocket on each representative structure. Fpocket detects pockets as clusters of alpha spheres. It reports a score, a druggability score, and a volume for every pocket.

  • Step 1: Conversion of the trajectory from Amber-compatible formats to GROMACS xtc (automatic when the input trajectory is not GROMACS-compatible).

  • Step 2: Creation of the atom-selection index file used for clustering.

  • Step 3: Clustering of the trajectory (gmx_cluster).

  • Step 4: Extraction of the most populated centroids.

  • Step 5: Cavity analysis of each centroid with Fpocket (fpocket_run).

  • Step 6: Filtering of cavities by score, druggability score and volume (fpocket_filter).

  • Step 7: Filtering of cavities by distance of their center of mass to a residue selection (filter_residue_com).

Usage

conda activate biobb_vs
cavity_analysis --structures_path data/receptor/receptor.pdb --filtering_selection "resid 31 or resid 21"

The config.yml is auto-generated from the CLI arguments into --output. --restart resumes from the last completed step when re-run against the same output folder. Run cavity_analysis --help for the full option list.

Options

Inputs

Provide either a trajectory (--traj_path + --top_path) or a folder of representative structures (--structures_path).

Flag

Default

Description

--traj_path

None

Input trajectory (GROMACS or AMBER formats); use with --top_path.

--top_path

None

Input topology/structure (gro, pdb); use with --traj_path.

--structures_path

None

Folder of representative structures (PDB); skips clustering.

Parameters

Flag

Default

Description

--filtering_selection

None

Atom selection (MDAnalysis syntax) used to filter pockets by the distance of their center of mass.

--distance_threshold

None

Distance threshold (Å) for the center-of-mass filter.

--num_clusters

20

Most-populated clusters to extract and analyze; ignored when --structures_path is given.

--clustering_method

linkage

gmx cluster method (linkage, jarvis-patrick, monte-carlo, diagonalization, gromos).

--clustering_cutoff

0.1

Clustering cutoff; reduce to increase the number of clusters.

--gmx_bin

gmx

GROMACS binary (gmx single-node, gmx_mpi multi-node).

--restart

False

Restart from the last completed step.

--output

working_dir_path

Output directory.

Recommendations

  • Clean the input first. Remove ligands, ions, and waters from the trajectory. Clustering and Fpocket work on the protein. Extra molecules can cause errors or noisy pockets.

  • Clustering uses the whole protein. To cluster on a specific region instead, do the clustering yourself and pass the representative structures with --structures_path.

  • Number of clusters. Lower --clustering_cutoff to get more clusters, raise it to get fewer (default 0.1). If too many clusters are obtained the workflow will re-launch the clustering increasing the cut-off value 50% up to 10 times. --num_clusters sets how many of the most populated clusters are analyzed (default 20); it is ignored when you pass --structures_path.

  • Distance filter. --filtering_selection and --distance_threshold keep only pockets whose center of mass is close to the chosen residues. A threshold of 10 Å is reasonable when the residues are part of the pocket.

Output

The top-level output folder contains three YAML summaries of the results, one per ranking:

  • summary_by_drug_score.yml

  • summary_by_score.yml

  • summary_by_volume.yml

Each summary lists the models (clusters). For every model it gives its pockets with their score, druggability score, and volume. Models are ranked by their best pocket. When the workflow did the clustering, each model also shows its population (how many trajectory frames it represents).

The druggability score goes from 0 to 1. A higher value means the pocket is more likely to bind a drug-like molecule. Volume is in ų.

Each model also has its own subfolder with the structure (model.pdb) and the detected pockets.

Limitations

  • Clustering uses the whole protein. The fit and cluster selection is fixed to Protein. You cannot restrict it to a region from the command line.

  • Fpocket detection settings are fixed: alpha-sphere radius 3–6 Å, and at least 35 spheres per pocket.

  • Filter thresholds are fixed: score 0.4–1, druggability score 0.4–1, volume 200–5000 ų. Pockets outside these ranges are dropped. These cannot be changed from the command line.