PPLM is a protein–pair language model that learns directly from paired sequences through a novel attention architecture, explicitly capturing inter-protein context. Building on PPLM, we developed PPLM-PPI, PPLM-Affinity, and PPLM-Contact for predicting protein–protein interactions, estimating binding affinity, and identifying interface residue contacts, respectively.

Authors: Jun Liu, Hungyu Chen, and Yang Zhang

Contact: junl_sg@nus.edu.sg

License: PolyForm Noncommercial License

Citation:
Jun Liu, Hungyu Chen, Yang Zhang. A Paired Sequence Language Model for Protein-Protein Interaction Modeling. Nature Communications (2026). https://doi.org/10.1038/s41467-026-70457-5

System Requirements

x86_64 machine
Linux Kernel OS

Software & Dataset Requirements (for PPLM-Contact)

HH-suite3 for MSA Search: Install HH-suite3 and update the "hhsuite_dir" parameter in the "pplm_contact/config.py" file.
Uniclust Database: Download the Uniclust30 database, unzip it on your machine, and update the "UniRef_database" parameter in the "pplm_contact/config.py" file.
CCMpred for DCA: Install ccmpred, or use the pre-packaged version in the "pplm_contact/external_tools" directory. Set the "ccmpred" parameter in the "pplm_contact/config.py" file. You may need to grant permission by running 'chmod +x pplm_contact/external_tools/ccmpred'.
LoadHHM for PSSM Calculation: Download LoadHHM.py and place the file in the "pplm_contact" directory of the PPLM package, or use the pre-packaged version within the "pplm_contact" directory.
ESM-MSA for Feature Generation: Install the ESM package, or use the pre-packaged version within "pplm_contact/external_tools" directory. Download the pre-trained ESM-MSA model and set the "esm_msa_model" parameter in the "pplm_contact/config.py" file.

Download parameters

You can download the pre-trained weights for PPLM and its downstream models from the links below and place them in the weights/ directory:

PPLM: https://zhanggroup.org/PPLM/bin/weights/pplm_t33_650M.pt or google_drive
PPLM-PPI: https://zhanggroup.org/PPLM/bin/weights/ppi_models.pkl or pplm-ppi_weights
PPLM-Affinity: https://zhanggroup.org/PPLM/bin/weights/affinity_models.pkl or pplm-affinity_weights
PLM-Contact: https://zhanggroup.org/PPLM/bin/weights/pplm_contact_models.pkl or pplm-contact_weights
PLM-Contact2: https://zhanggroup.org/PPLM/bin/weights/pplm_contact_models2.pkl or pplm-contact2_weights

Usage

1. Install environment

conda env create -f environment.yml

2. Activate environment

conda activate PPLM

3. Run PPLM-PPI

python run_pplm-ppi.py example/seq1.fasta example/seq2.fasta

You can also run PPLM-PPI for two individual sequences:

python pplm_ppi/predict.py example/seq1.fasta example/seq2.fasta

4. Run PPLM-Affinity

python run_pplm-affinity.py example/receptor.fasta example/ligand.fasta

5. Run PPLM-Contact

For homodimer

python run_pplm-contact.py example/protein.pdb example/protein.pdb example/homo_example

For heterodimer

python run_pplm-contact.py example/protein1.pdb example/protein2.pdb example/hetero_example

6. Run PPLM-Contact2

For homodimer

python run_pplm-contact2.py example/homodimer.afm.pdb example/homodimer.af3.pdb example/homodimer.dmf.pdb example/homo_example2

For heterodimer

python run_pplm-contact2.py example/heterodimer.afm.pdb example/heterodimer.af3.pdb example/heterodimer.dmf.pdb example/hetero_example2

7. Generate embeddings and attention matrices for other applications

python run_pplm.py example/seq1.fasta example/seq2.fasta example/seq1-seq2.pplm.pkl

Example Outputs

PPLM-PPI

Command:

python python run_pplm-ppi.py example/seq1.fasta example/seq2.fasta

Output: Predicted interaction probability printed to the command line：

Predicted interaction score: 0.9431089

PPLM-Affinity

Command:

python run_pplm-affinity.py example/receptor.fasta example/ligand.fasta

Output: Predicted binding affinity printed to the command line：

Predicted binding affinity: -7.6090136

PPLM-Contact

Command:

python run_pplm-contact.py example/protein.pdb example/protein.pdb example/homo_example

Output: The predicted contacts are saved in example/homo_example/homo_example.pred_contact.txt:

Format:
Rank      ResIdx1   ResType1  ResIdx2   ResType2  Contact_Probability
1         23:A      MET       26:B      CYS       0.976151
2         26:A      CYS       23:B      MET       0.974481
3         22:A      ILE       26:B      CYS       0.971633
4         23:A      MET       30:B      GLN       0.971191
5         30:A      GLN       22:B      ILE       0.970514
6         27:A      GLY       23:B      MET       0.970334
7         22:A      ILE       30:B      GLN       0.970124
8         30:A      GLN       23:B      MET       0.96919
9         23:A      MET       27:B      GLY       0.966725
10        23:A      MET       23:B      MET       0.966512
...

Troubleshooting (MKL and libperl.so)

On some systems, users may encounter MKL- or libperl.so-related errors due to local library and environment differences. We recommend installing PPLM in a fresh conda environment using the provided environment.yml:

conda create -n pplm python=3.10
conda activate pplm
conda env update -n pplm -f environment.yml

If MKL errors persist (e.g. import errors for numpy/pytorch), please try reinstalling MKL or recreating the environment:

conda install mkl

If you see an error about libperl.so not found when running PPLM-Contact or PPLM-Contact2, first ensure perl is installed and that the library path is visible:

conda install -c conda-forge perl
ls $CONDA_PREFIX/lib | grep libperl
cd $CONDA_PREFIX/lib
ln -s libperl.so.5.xx libperl.so   # replace with the actual version
export LD_LIBRARY_PATH="$CONDA_PREFIX/lib:$LD_LIBRARY_PATH"

README.md Unescape Escape

PPLM: A Paired Sequence Language Model for Protein-Protein Interaction Modeling