AlphaFold-NMR Protocol (under construction)

Scripts and data corresponding to Huang, Ramelot, Spaman, Kobayashi, Montelione (2025) "Hidden Structural States of Proteins Revealed by Conformer Selection with AlphaFold-NMR"

No required non-standard hardware

Software Requirements

R Packages:

bio3d: http://thegrantlab.org/bio3d/
cluster 

Python Dependences

numpy
pandas

RCI webserver for RCI and SCC calculation

https://www.randomcoilindex.ca/cgi-bin/rci_cgi_current.py

ASDP/RPF for Recall calculation in batch mode

https://github.rpi.edu/RPIBioinformatics/ASDP_public

RPF webserver for DoubleRecall analysis

https://montelionelab.chem.rpi.edu/rpf/

AFsample for enhanced sampling

https://github.com/bjornwallner/alphafoldv2.2.0

AlphaFold-NMR R and Python Scripts with Demo

Typical install times are several minutes.

*** all paths in the scripts need to be changed to your local computer environment

1. AI Enhanced sampling using AFsample

• run_afsample6000.sh
  • need to modify the path to fit your local computer system
  • calculate 6000 models and also relax all 6000 models.

Runtime: it can take day(s) to calculate 6000 models, depending on the size of the sequence and number of GPUs

• mergeChain.py and runMergedChain.py

  • merge two chains into one chain for PCA analysis

• FilterAF2.py: filters out bad models based on the AF log file. The python code is copied from here: https://github.rpi.edu/RPIBioinformatics/FilteringAF2_scripts

  Input sequence for AlphaFold: doc1_noN.fasta. 
  We excluded the long disordered tails and non-native tags from the input fasta sequence for AlphaFold modeling, to avoid potential influence on the pTM and <pLDDT> scores. 
     
  Commands: 
  
  #cd CDK2AP1-doc1 (set working dir: CDK2AP1-doc1)
  
  sbatch run_doc1_noN.sh (running with slrum) 
  This command calculates and relax all 6000 models using run_afsample6000.sh <br>
  The output models are here: AF_models_dropout/doc1_noN. pTM score is reported here: AF_models_dropout/scores.sc and log from AF: slurm-xxx.out 
  
  # python FilterAF2.py -log slurm-xxx.out -rel -inD AF_models_dropout/doc1_noN -outD filteredModels
  This command filters out bad models based on the AF log file (e.g. slurm-xxx.out). 

  Additional processing scripts: 
  
  Merge two chains into one chain for clustering analysis
  # python ../scripts/runMergedChain.py filteredModels mergedModels
  
  This command finds all pdb file in the fileredModels, merge two chains (using mergeChain.py) and save them in the mergedModels directory. 

Download pre-filtered AFsample models

https://zenodo.org/records/15015917 has 5984 models with one merged chain. Please unzip it and name it as CDK2AP1-doc1/ESmodels/ for the following analysis:

2. Clustering

• dmPCAClustering.R --> output: pc_dm_pdbs.RData, cluster_pc_dm.csv (in Rstudio, set the working dir to CDK2AP1-doc1 before running the R script)

We use "ward methods". To identify number of clusters --> by inspection of "Dendrogram" and pc plots.

Runtime: it can take hours (s) to clustering 6000 models, depending on the size of the sequence and number of GPUs

3. Scoring

Input files:

• ESmodels: all pdb files, support filenames with "relaxed*.pdb" or "relaxed*_new.pdb"
• NMRdata: input files to run RPF
• RCI1.csv: from RCI webserver, the sequence was edited to match with merged sequence from ESmodels

Scripts:

• runSCC.py: calulate SCC scores for all models, and write to file scc.sc

   # python ../scripts/runSCC.py RCI1.csv ESmodels > scc.sc

• runRPF.py and getRPF.py: calculate RPF scores for all models, and write to file rpf.sc. slow step - performance can be improved by only output recall, precision, f-measure and dp scores and skips others.

   working directory: NMRdata
   # python ../../scripts/runRPF.py control_RPF ../ESmodels rpfESmodels 
   need to set the RPFcommand in the runRPF.py script 
   # python ../../scripts/getRPF.py rpfESmodels > rpf.sc  

Runtime: mintues to hours for 6000 models, depends on the size of the protein sequence
output: NMRdata/rpfESmodels and NMRdata/rpf.sc

• getpLDDT.py: calculate scores for all models, and write to file pLDDT.sc

   # python ../scripts/getpLDDT.py ESmodels > pLDDT.sc 

• getScores.py: combine all scores.
  This script only works for models with the name "relaxed****.pdb" from AFsample. If your model name is different, you will need to change the script.

   # python ../scripts/getScores.py scores.sc pLDDT.sc scc.sc rpf.sc cluster_pc_dm.csv > scores.all  
   # python ../scripts/getScores.py > scores.all (as default, this command will also readin the above files)

4. State combination

• selectModles.py: select models based on p(model|NMR) scores, create a diretory with top5 models from each cluster
  

 # sh ../scripts/selectModels.py scores.all ESmodels/ selectedModels/ > selectedModels.txt 
 scores.all: output from step 3 - Scoring
   
 plotScores.R: R plots of these scores used for top 5 model selection from each cluster

output: selectedModels.txt and selectedModels/

• State combination is based on the SEM plot of RMSFvsRCI and CCC scores.

example R code: CDK2AP1-doc1/RMSF_RCIplot.R 

5. Validation by DoubleRecall analysis

For selected states, get RPF.zip file by runing RPF webserver (https://www.randomcoilindex.ca/cgi-bin/rci_cgi_current.py)

• DoubleRecall - ensemble A vs ensemble B: https://colab.research.google.com/drive/1IPXZVDdxG1xpn-6tgIErWJOSLa9Rx87A

• DoubleRecall - ensemble A vs ensembles B1+B2: https://colab.research.google.com/drive/1YShoUAQWdKu0EccMJgkLeeiojUWKT0SJ

Other codes that may be useful:

• RCItools -- tools we developed to generate SHIFTY input file to run RCI webserver
  • RPFtable2SHIFTY.py: convert the chemical shift file used by RPF (bmrbtable file) to SHIFTY format
  • nmrstar3toSHIFTY-fromBMRB.py: give the bmrb ID number, download chemical shift assignments from the BMRB database and convert to SHIFTY format
  • nmrstar3toSHIFTY.py: convert the local bmrb file in nmrstart 3.0 format to SHIFTY format
• group all pdb files in one directory into one pdb file R plots
• pLDDT_RCIplot.R