This repository contains the code to produce the results in the paper "Uncovering distinct protein conformations using coevolutionary information and AlphaFold" by Yongkai Chen, Samuel W.K. Wong, and S. C. Kou.

• cuda version >=12
• cudnn version >=9
• GCC >= 12
• python-3.10.12

Via mamba (recommended): mamba env create -f SMICE.yml

or via conda: conda env create -f SMICE.yml

• This takes 10-15 mins to build

• HHsuite: refer to https://github.com/soedinglab/hh-suite for installation

• colabFold: refer to https://github.com/sokrypton/ColabFold for installing localColabFold

• Foldseek: refer to https://github.com/steineggerlab/foldseek for installation

• AFcluster and Random Sampling (optional, only to reproduce results from these methods): refer to https://github.com/HWaymentSteele/AF_Cluster for installation and implementation

• Obtain and unzip the MSA files MSA_cov75_all.zip to the base directory

• Unzip the pdb files of the true conformations pdbs_92.zip

• Unzip PDB_annotations.txt.zip

The instructions in this section assume access to a Slurm-based high-performance computing cluster. For instructions on how to run an example on a local workstation or laptop, please skip to the "Run Locally" section at the end of this README.

cd bash/benchmark_exp

chmod +x *.sh

See bash/benchmark_exp/README.md for instructions on updating all *.slurm files

See config/README.md for details of setting up the configuration file

Set the paths in config/config_SMICE_benchmark.json accordingly

./run_SMICE_exmp.sh

./run_SMICE_all.sh

cd experiment/validation

conda activate SMICE

python all_calculate_TMscores.py

We use the TMscore results of AF-Cluster and Random Sampling as saved in AFclust_random_res.zip.

unzip AFclust_random_res.zip

Compare the TMscores of the top predictions:

python compare_TopPred.py

Compare the overall prediction accuracy of the prediction sets:

python compare_OverallPred.py

Analyze relationships between confidence metrics and TMscores:

cd experiment/analysis

python confidence_metric_analysis.py

All generated files are organized as follows:

• Job Outputs are saved in a subfolder (named by the PDB ID) within the directory specified by base_output_dir in config/config_SMICE_benchmark.json. This includes:

  • Sampled MSA subsets and predicted PDB files for different steps of SMICE
  • RepStructure.zip (contains the extracted representative structures)
  • Clustering_Res (contains the detailed clustering results)
  • outputs_SMICE.json.zip (the file paths of the predicted PDB files, the corresponding MSA file paths, and the corresponding confidence scores)

• Results & Analysis are saved in the directory specified by base_result_dir in config/config_SMICE_benchmark.json. This includes summary figures for:

  • TM-score scatter plots
  • Comparison results
  • Confidence metric analysis

The SMICE code for running locally is provided, which can be downloaded and unzipped on a local workstation or laptop.

• Follow the same "Installation" and "Dataset" instructions as above.

• Set the paths in config/config_SMICE_benchmark.json to match your local folder setup.

• Activate the SMICE environment, and set your CONDA_PREFIX accordingly in line 33 of run_SMICE_exmp_local.sh in the bash/benchmark_exp directory.

• Colabfold will use the GPU if available. Otherwise, CPU-based colabFold will run ~10 times slower on a system without an Nvidia GPU/CUDA driver, so we reduce the MSA sampling size of SMICE for demonstration purposes. Note that this could take over 20 hours on a desktop computer, but will run significantly faster if a recent GPU is available. This demonstration version for an example protein can be executed by running the following script from the bash/benchmark_exp directory:

  ./run_SMICE_exmp_local_toy.sh

• To instead run the full version of SMICE for this example, run the following script from the bash/benchmark_exp directory:

  ./run_SMICE_exmp_local.sh