We show how to run AF2Complex to predict the structures of three multimeric protein assembly targets from CASP14 as examples. The examples cover both the monomeric and multimeric deep learning models released by DeepMind, different MSA pairing modes, and the interface score evaluation.
It is recommended to use a computer with a modern GPU (e.g., Nvidia RTX, V100, A100) to test these examples because of fast result turnaround within minutes, except for Example 4 (which takes about 35 minutes on an A100 machine). But with patience you should be able to complete these examples on a modern CPU computer as well.
Before running these examples, please ensure that correct environment for AF2Complex has been properly set up. To do so, use a text editor to open an example shell script, and modify it accordingly. As instructed in the script, you must provide the correct conda environment of AlphaFold, if not already loaded; and provide the correct path to the weight parameter files of AlphaFold 2 neural network models. In AF's releases, these model parameters are named as params_model_x_ptm.npz (AF version 2.0.x), params_model_x_multimer.npz (AF version 2.1.x), and params_model_x_multimer_v2.npz (AF version 2.2.x). In the examples below, we employ two sets of parameters released in AF v2.0 and v2.2.
Predicting the structure of a heterodimer (H1065) using two AF-Multimer v2 DL models. In this example, the target is composed of two monomers (T1065s1 and T1065s2). AF2Complex first retrieves the input features of each monomer under the subdirectory af2c_fea, then pairs their MSAs using the MSA pairing option all. The example has two files,
example1.shThe shell script to run AF2Complex model inference.targets/example1.lstThe list file that defines the stoichiometry of the target.
Run the script with the command,
./example1.shThis run uses two multimer_v2 models to make complex structure prediction. The output structural models (in PDB format) are under the subdirectory af2c_mod/H1065. For the purpose of comparison, the experimental structures of these two examples are provided under subdirectory ground_truth.
You may also view the experimental structure of this target directly in the
Protein Data Bank.
To view your predicted structures, you may take advantage of the online molecular viewer Mol* in a simple or more advanced version by uploading your predicted structure coordinate files (.pdb). The other two types of output files are pickle files (.pkl), which contain detailed results of the prediction; and json files (.json), which contain statistics of the prediction. Example output is under af2c_mod_examples/H1065. Note the pickle files are not included in the example output because they are too large for github sharing.
Predicting the structure of a tetramer target (H1072) using two AF2 monomeric DL models in the unpaired MSA mode. The target has two monomers (T1072s1 and T1072s2), each contribute two copies to make the tetramer. The stoichiometry is A2:B2. The example has two files,
example2.shThe shell script to run AF2Complex model inference.targets/example2.lstThe list file that defines the target, such as stoichiometry and copy numbers in this case.
Run the script with the command,
./example2.shThis run uses two monomer_ptm models to make complex structure prediction. The output structural models (in PDB format) are under the subdirectory af2c_mod/H1072. Example output is under af2c_mod_examples/H1072. For the purpose of comparison, the experimental structures of these two examples are provided under subdirectory ground_truth. The experimental coordinates of this target are available in
the PDB.
Predicting the structure of the above two targets (H1065 and H1072) simultaneously using one AF-Multimer v2 DL model in unpaired MSA mode. The example has two files,
example3.shThe shell script to run AF2Complex model inference.targets/example3.lstThe target list file. Note that+instead of/is used to separate two complexes. The+specify the interface the iScore will evaluate.
Run the script with the command,
./example3.shIn this example, the interface score (iScore) evaluation is conducted between these two complexes, i.e., to assess whether there is any interaction between these two complexes. The resulting structure model comprises two protein complex clusters, one the dimer, and the other the tetramer.
Predicting the structure of a homo-dodecamer using one AF2 monomeric DL model
in the cyclic MSA paring mode. The target H1060v4 is an homo-oligomer, and the input monomer features are under af2c_fea/H1060s3. This example takes about half an hour using an A100 GPU. Eight recycles are recommended for this example. If you can only run short recycles, checkpoint option is activated such that you may continue the modeling in multiple runs. The example has two files,
example4.shThe shell script to run AF2Complex model inference.targets/example4.lstThe target list file called by the shell script
Run the script with the command,
./example4.shAs of Aug 2022, the expreimental structure of this target have not been released. But it is known that the target structure is a circular ring. The cyclic MSA mode provided by AF2Complex gives you a better chance to obtain such a ring-like structure, see an example output under af2c_mod_examples/H1060v4. It may take several tries to get it, and you may also see an ellipse structure or multiple rings. You may also try multimer v2 models with or without the cyclic MSA mode, or multimer v1 models in the unpaired MSA mode. Note that the radius of these circular rings resulted from different models may not be the same in this challenging example.
The general script for AF2Complex model inference is provided in the script
run_af2comp.sh, which processes targets specified in test.lst. The target list files defined multiple targets, each with its own stoichiometry, domain ranges, copy numbers, etc.
Run the script with the command,
./run_af2comp.shYou can calculate the interface score (iScore) using a result pickle file generated by AF2Complex model inference. In this example, we re-calculate the interface scores using the pickle files generated in running Example 2.
calc_interface_score.shThe shell script to calculate the interface score.targets/test_iscore.lstThe target list file called by the shell script.
It conducts two iScore evaluations as specified in targets/test_iscore.lst on each pickle file. The first iScore is to evaluate all intefaces of the full tetrameric complex. The second is to evaluate only the interactions between two homodimers. In the latter case, it uses a magic + to divide individual monomeric chains into two sets of superchains, each with multiple chains. When + is present in the stoichiometry of a target, iScore is calculated only on components separated by +, instead of all interfaces by default without using any +. This is useful for PPI screening purpose, e.g., screening a protein complex with multiple chains against arbitrary library entries, either monomers or known multimers. We consider predictions of medium, high, and very high confidence, progressively, at iScore cutoffs of 0.4, 0.5, 0.7.
If you apply this package to a new target. The first step is to generate input features. run_fea_gen.sh is an example shell script to derive input features for subsequent model inference. To run it, you need to first install third-party sequence tools and libraries as required by AlphaFold 2. And modify this script to reflect correct installation path for sequence alignment accordingly.
There are three feature_mode to choose from:
-
monomer+speciesModified monomer data pipeline to include additionally species information. We recommend that you use this mode, which enables a flexibility using either unpaired MSAs or paired MSAs in various paring modes available in AF2Complex v1.3.0. The monomer features generated with these two monomer modes can be used to predict arbitrary combinations of the monomers for modeling a multimeric target. -
monomerAF v2.0 data pipeline to generate input features for predicting monomer structures. The features generated by this option are used by AF2Complex v1.2.2 and below. -
multimerAF v2.1 data pipeline to generate input features for predicting multimer structures. The multimer features generated by this option can only be used in combination with the "multimer" model preset. It is provided such as you can run the official DeepMind's procedure. Certain AF2Complex features, such as domain cropping, are not supported in this model preset.
The input features used by examples above were all generated by using the option monomer+species.
An example shell script of model relaxation is provided in run_relaxation.sh.
To test it after you have successfully run the example above,
./run_relaxation.shThis script will launch the relaxation protocol on all un-relaxed structural models generated by running the examples above. Note that the relaxation is a molecular dynamics minimization procedure, which cannot eliminate severe clashes as observed in some models of large complexes generated by deep learning models.
examples.shRun through Examples 1 to 3targets/examples.lstInput targets and parameters called byexamples.shaf2c_feaDirectory contains pre-generated features of monomers and complex targets. For a monomer, the input features contains species information. See feature generation aboveaf2c_mod_examplesExamples of output modelsaf2c_modOutput directory created by running the example scriptsground_truthCoordinates (PDB format) of target H1065 and H1072 determined by experimentstargets/test_custom_msa.lstandtest.adjAn example of customized MSAs, an experimental feature.