We evaluate existing AlphaFold-based, including AlphaLink2, Boltz2, and GRASP for their applicability towards integrative modeling.
- Kartik Majila, Shruthi Viswanath. Evaluation of methods for AlphaFold-based integrative modeling. (2026) bioRxiv.
- Data is deposited in Zenodo
- See
environment.ymlfor the dependencies. - Ensure Mamba is installed on the system.
mamba env create -n af_im -f environment.yml
We used the pl_upgrades branch in openfold git repo (for Cuda12). As in openfold directory.
Download the databases for AlphaFold/OpenFold. See instructions for downloading the databases on the OpenFold repository.
Add the path to the af_im repository to bash_profile and run,
source ~/.bash_profile
Update the following paths in the config.py file: db_dir, tool_base, openfold_params_dir
Clone the git repo from here.
We converted the Python2 implementation to Python3 using python2to3.com server.
For installation, run
python setup.py install
Install USalign as instructed here.
Post installation, add the executable to ./af_im/.
Download and install Phenix from here.
Clone and install AlphaLink2 as described here.
Clone and install Boltz2 as described here.
Clone and install GRASP as described here.
Post installation, activate the respective models environment and install the following packages:
pip install pandas, typing-extensions, ml_collections
mkdir benchmark
cd ./data/
Dataset creation occurs in multiple stages:
- Metadata collection
python prep_data2.py
This downloads the required metadata for all complexes (.pdb and .cif), followed by parsing the CIF file to obtain the sequence and residue numbers (seq_id).
Further it runs JWalk to obtain crosslinks for all complexes.
It creates the following 2 directories: {benchmark name}_benchmark/ and {benchmark name}_metadata/.
- Creating input files for modeling
python create_benchmark2.py
For all complexes selected in step 1, it creates a directory within {benchmark name}_benchmark/ containing the structure file (.cif or .pdb), data file (.csv file for crosslinks), and a JSON dict containing configs for modeling.
- Creating MSAs
python create_msas.py -b BENCHMARK_NAME -p
BENCHMRK_NAME could be either of crosslink or multistate. Use -p for multimers.
This script runs the OpenFold MSA creation pipeline for obtaining the MSA required for structure prediction.
For obtaining the input files for the multi-state proteins run the following script:
cd ../
python multi_state2.py
This script will create the input files for the multi-state proteins in the same format as for the crosslink benchmark. The directory structure is the same as above.
Activate the respective models environment and then run the following command for obtaining predictions,
python get_competing_method_preds.py -m MODEL -c CONFIG_NAME -b BENCHMARK_NAME -d DEVICE
| Flags | Description |
|---|---|
| -m | Model to use for prediction: alphalink2/boltz2/grasp |
| -c | Configs to use for obtaining predictions from a model. See model_configs.py for the available configs |
| -b | Name of the benchmark to obtain predictions for: crosslink/multistate |
| -d | Device to run predictions on: cpu/cuda:0/cuda:1 |
For the results shown in the paper we used the following configs: alpha, beta1, beta3, kappa1, kappa2, kappa3.
For the crosslink benchmark use the following command to run the analysis,
python analysis.py -m MODEL -b crosslink
For multistate benchmark run the following command,
python analysis.py -m MODEL -b multistate -ms
MODEL is the same as defined above.
This script runs the analysis for all specified configs for a given model in the model_configs.py.
To obtain the plots shown in the paper, run the following command,
python paper_figures.py
Author(s): Kartik Majila, Shruthi Viswanath
Date: MM DD, 2026
License: GPL v3
This work is licensed under the terms of the GNU General Public License,
Version 3, as published by the Free Software Foundation on 29 June 2007.
Testable: Yes
Parallelizeable: Yes
Publications: Majila K., Viswanath S. Evaluation of methods for AlphaFold-based integrative modeling. bioRxiv (2026), DOI.