CF-random

CF-random predicts alternative protein conformations and fold-switching behavior using AlphaFold2 with varied MSA depths.

To run CF-random in a Colab notebook:

CF-random depends on ColabFold (structure prediction) and Foldseek (structure search). A convenience script sets up the conda environment and required packages:

chmod +x install.sh
./install.sh

CF-random supports three modes:

• FS — fold-switching prediction with two reference PDBs
• AC — alternative conformation prediction with two reference PDBs
• blind — alternative conformation or fold-switching without reference PDBs

• FS and AC modes require two reference PDB files (fold1 and fold2).
• Each target needs its own MSA directory; do not reuse output folders across runs.
• All required PDB files, MSA files, and Python scripts must be in the same directory.
• PDB files should contain a single chain. Multi-chain PDBs may be converted automatically in some workflows, but providing single-chain PDBs avoids issues.
• FS mode requires a range_fs_pairs_all.txt file describing the fold-switching region (see Fold-switching mode below for format details). ColabFold uses 1-based residue indexing; ensure ranges match your PDB/sequence.
• --num_msa and --num_ens apply to all modes except blind, which does not support --num_ens.
• Activate the conda environment before running:

conda activate cf-random

Predicts RfaH using two reference structures and an MSA.

cf-random --fname 2oug_C-search/ --pdb1 2oug_C.pdb --pdb2 6c6s_D.pdb --option FS

Input files:

• 2oug_C.pdb — dominant reference
• 6c6s_D.pdb — alternative reference
• 2oug_C-search/0.a3m — MSA
• range_fs_pairs_all.txt — fold-switching region definition

range_fs_pairs_all.txt format:

Each line defines the fold-switching region for a pair of reference structures: pdb1, pdb2, XXX-XXX, XXX-XXX, XXX-XXX, XXX-XXX

Fields are: PDB1 name, PDB2 name, residue range of reference 1, residue range of reference 2, residue range of prediction 1, residue range of prediction 2. ColabFold generates residue indices starting from 1; ensure your ranges match the residue numbering in your PDB files.

Generates 200 structures; takes under 30 minutes on an A100 GPU.

Output files (written to predictions_all/):

MSA depth encoding: 0=1:2, 1=2:4, 2=4:8, 3=8:16, 4=16:32, 5=32:64, 6=64:128

Predicts Lactococcal OppA using two reference structures and an MSA.

cf-random --fname P71447-search/ --pdb1 2wfa_A.pdb --pdb2 2wf5_A.pdb --option AC --num_msa 5

Input files:

• 2wfa_A.pdb — dominant reference
• 2wf5_A.pdb — alternative reference
• P71447-search/P71447_converted.a3m — MSA

Generates 200 structures; takes under 70 minutes on an A100 GPU (~250 residue protein).

Output files (written to predictions_all/):

MSA depth encoding: 0=1:2, 1=2:4, 2=4:8, 3=8:16, 4=16:32, 5=32:64, 6=64:128

Predicts alternative conformations or fold-switching without reference PDBs. Uses Foldseek to cluster predicted structures and identify structures of interest.

cf-random --pname Mad2_test --fname 2vfx_L-search/ --option blind

Note: Before running blind mode, make a symbolic link to the Foldseek PDB libraries in the directory where you run the command.

Input files:

• 2vfx_L-search/0.a3m — MSA

Generates 200 structures + 200 Foldseek result files; takes under 70 minutes on an A100 GPU.

Output files (written to blind_prediction/Mad2_test/):

Lee, M., Schafer, J.W., Prabakaran, J. et al. Large-scale predictions of alternative protein conformations by AlphaFold2-based sequence association. Nat Commun 16, 5622 (2025). https://doi.org/10.1038/s41467-025-60759-5

See LICENSE.md.