Learning the language of protein-protein interactions

🌿 Overview of MINT

MINT (Multimeric INteraction Transformer) is a Protein Language Model (PLM) designed for contextual and scalable modeling of interacting protein sequences. Trained on a large, curated set of 96 million protein-protein interactions (PPIs) from the STRING database, MINT outperforms existing PLMs across diverse tasks and protein types, including:

• Binding affinity prediction
• Mutational effect estimation
• Complex protein assembly modeling
• Antibody-antigen interaction modeling
• T cell receptor–epitope binding prediction

🔬 Why MINT?

✅ First PLM to be trained on large-scale PPI data

✅ State-of-the-art performance across multiple PPI tasks

✅ Scalable and adaptable for diverse protein interactions

🖥️ Installation

1. Create a new conda environment from the provided enviroment.yml file.

conda env create --name mint --file=environment.yml

2. Activate the enviroment and install the package from source.

conda activate mint
pip install -e .

3. Check if you are able to import the package.

python -c "import mint; print('Success')" 

4. Download the model checkpoint and note the file path where it is stored.

wget https://huggingface.co/varunullanat2012/mint/resolve/main/mint.ckpt

🚀 How to use

Generating embeddings

We suggest generating embeddings from a CSV file containing the interacting sequences like this one here. Next, simply execute the following code to get average embeddings over all input sequences.

import torch
from mint.helpers.extract import load_config, CSVDataset, CollateFn, MINTWrapper

cfg = load_config("data/esm2_t33_650M_UR50D.json") # model config
device = 'cuda:0' # GPU device
checkpoint_path = '' # Where you stored the model checkpoint

dataset = CSVDataset('data/protein_sequences.csv', 'Protein_Sequence_1', 'Protein_Sequence_2')
loader = torch.utils.data.DataLoader(dataset, batch_size=2, collate_fn=CollateFn(512), shuffle=False) 

wrapper = MINTWrapper(cfg, checkpoint_path, device=device)

chains, chain_ids = next(iter(loader)) # Get the first batch
chains = chains.to(device)
chain_ids = chain_ids.to(device)
embeddings = wrapper(chains, chain_ids)  # Generate embeddings
print(embeddings.shape) # Should be of shape (2, 1280)

However, we recommend using the sep_chains=True argument in the wrapper class for maximum performance on downstream tasks. This gets the sequence-level embedding for all sequences, and returns it concatenated in the same order as the input.

wrapper = MINTWrapper(cfg, checkpoint_path, sep_chains=True, device=device)

chains, chain_ids = next(iter(loader)) # Get the first batch
chains = chains.to(device)
chain_ids = chain_ids.to(device)
embeddings = wrapper(chains, chain_ids)  # Generate embeddings
print(embeddings.shape) # Should be of shape (2, 2560)

Binary PPI classification

We provide code and a model checkpoint to predict whether two input sequences interact or not. The downstream model, which is an MLP, is trained using the gold-standard data from Bernett et al..

import torch
from mint.helpers.extract import load_config, CSVDataset, CollateFn, MINTWrapper
from mint.helpers.predict import SimpleMLP

cfg = load_config("data/esm2_t33_650M_UR50D.json") # model config
device = 'cuda:0' # GPU device
checkpoint_path = 'mint.ckpt' # Where you stored the model checkpoint
mlp_checkpoint_path = 'bernett_mlp.pth' # Where you stored the Bernett MLP checkpoint

dataset = CSVDataset('data/protein_sequences.csv', 'Protein_Sequence_1', 'Protein_Sequence_2')
loader = torch.utils.data.DataLoader(dataset, batch_size=2, collate_fn=CollateFn(512), shuffle=False) 

wrapper = MINTWrapper(cfg, checkpoint_path, sep_chains=True, device=device)

# Generate embeddings 
chains, chain_ids = next(iter(loader)) 
chains = chains.to(device)
chain_ids = chain_ids.to(device)
embeddings = wrapper(chains, chain_ids) # Should be of shape (2, 2560)

# Predict using trained MLP
model = SimpleMLP() 
mlp_checkpoint = torch.load(mlp_checkpoint_path)
model.load_state_dict(mlp_checkpoint)
model.eval()
model.to(device)

predictions = torch.sigmoid(model(embeddings)) # Should be of shape (2, 1)
print(predictions) # Probability of interaction (0 is no, 1 is yes)

Finetuning

To finetune our model on a new supervised dataset, simply set the freeze_percent parameter to anything other than 1. Setting it to 0.5 means the last 50% of the model layers can be trained. For example,

import torch
from mint.helpers.extract import MINTWrapper

cfg = load_config("data/esm2_t33_650M_UR50D.json") # model config
device = 'cuda:0' # GPU device
checkpoint_path = '' # path where you stored the model checkpoint

wrapper = MINTWrapper(cfg, checkpoint_path, freeze_percent=0.5, device=device)
for name, param in wrapper.model.named_parameters():
    print(f"Parameter: {name}, Trainable: {param.requires_grad}")

Pre-training on STRING-DB

This section outlines the steps required to pretrain MINT on PPIs from STRING-DB. First, to create the train-validation splits we used, first download protein.physical.links.v12.0.txt.gz and protein.sequences.v12.0.fa.gz from STRING-DB.

Then, run the following commands to cluster the sequences using a 50% sequence similarity threshold using mmseqs.

mmseqs createdb protein.sequences.v12.0.fa DB100
mmseqs cluster DB100 clu50 /tmp/mmseqs --min-seq-id 0.50 --remove-tmp-files
mmseqs createtsv DB100 DB100 clu50 clu50.tsv

Then, run stringdb.py, ensuring that the filepaths in that script match the paths where you stored the protein.sequences.v12.0.fa, clu50.tsv (output of the previous step), and protein.physical.links.full.v12.0.txt.gz files.

Finally, run the training like this:

python train.py --batch_size 2 --crop_len 512 --model 650M --val_check_interval 320000 --accumulate_grad 32 --run_name 650M_nofreeze_filtered --copy_weights --wandb --dataset_split filtered

Examples

We provide several examples highlighting the use cases of MINT on various supervised tasks and different protein types in the downstream folder.

1. Predict whether two proteins interact or not
2. Predict the binding affinity of protein complexes
3. Predict whether two proteins interact or not after mutation
4. Predict the difference in binding affinity in protein complexes upon mutation

📝 Citing

@article{ullanat2026learning,
  title={Learning the language of protein-protein interactions},
  author={Ullanat, Varun and Jing, Bowen and Sledzieski, Samuel and Berger, Bonnie},
  journal={Nature Communications},
  year={2026},
  publisher={Nature Publishing Group UK London}
}