MULTICOM_qa

Estimate (predict) the quality of protein multimer structure models. It uses the pairwise similarity between structural models and a deep learning-based interface contact probability score to predict the quality of the models. It was ranked first in predicting the global accuracy of strucural models of protein multimers/assemblies/complexes in the 15th Critical Assessment of Techniques for Protein Structure Prediction (CASP15) in 2022 (see the SCORE ranking at https://predictioncenter.org/casp15/zscores_EMA.cgi).

Requirements:

1. CDPred
2. MMalign
3. Biopython
4. Futures
5. Numpy

Installation

1. Download Multicom_qa

   git clone https://github.com/BioinfoMachineLearning/MULTICOM_qa.git

2. Download and setup the CDPred.

3. Activate the CDPred enviroment.

4. Pip install -r requirements.txt.

5. Copy the file "run_CDFold.sh" inside the CDPred directory

6. Download the MMalign

7. Open the "config" file and edit the path of the following variables with the appropiate path:

   1. MM_ALIGN_PATH = "path of the MMalign tools"

      e.g. MM_ALIGN_PATH = "/home/multicom4s_tool/MMalign"

   2. CDPred = "path of the installed CDPred directory"

      e.g. CDPRED_PATH = "/home/multicom4s_tool/CDpred_mm_eva/CDPred/run_CDFold.sh

Usage

python multi_eva.py <fasta_file in casp format> <dir of the structure for evaluation > <stoichiometry> <dir of the tertiary structures (Alphafold 2 recommended)> <CPU_count > <out put directory>

e.g python multi_eva.py ./examples/H1143/H1143.fasta ./examples/H1143/decoys/ A1B1 ./examples/H1143/H1143_alphafold_monomer/ 4 ./examples/H1143/output/

Final output would be a csv file with the final score of all the structures in the output directory, e.g "H1143.csv". The csv will contain "Final Rank" which is the weighted average of the Interchain Contact Probability Score (ICPS) rank and Pairwise Similarity Score (PSS) RANK. And will also contain "Final_Score" which is the weighted average of the ICPS and PSS. The expected final output is provided in the expected_output directory.

Predicted score vs True score

The predicted result coincides with true one as displayed below.

Model	Final Rank	Final Score	True TM-Score
H1143TS035_5	1.4	0.4712	0.9528
H1143TS035_3	1.6	0.4676	0.8534
H1143TS035_4	3	0.4227	0.6771
H1143TS399_4	4	0.41186	0.6509
H1143TS037_3	5.4	0.36549	0.5926
H1143TS493_2	5.6	0.36231	0.5919

Reference

Roy, R. S., Liu, J., Giri, N., Guo, Z., & Cheng, J. (2023). Combining pairwise structural similarity and deep learning interface contact prediction to estimate protein complex model accuracy in CASP15. bioRxiv.