DPLM: Dynamics-aware Protein Language Model

DPLM is a dynamics-aware protein language model that enriches protein sequence representations with information learned from molecular dynamics (MD) trajectories. DPLM is pretrained using unsupervised contrastive learning, aligning sequence embeddings with embeddings of real MD trajectories.

The resulting representations capture protein flexibility and dynamics while requiring only sequence input at inference time.

Features

This repository provides code for:

• Extracting DPLM protein representations
• Running intrinsic disorder region (IDR) prediction
• Running protein stability change (ΔΔG) prediction

🔗 Resources

• Code repository: https://github.com/yuexujiang/DPLM_release
• Pretrained checkpoints (Hugging Face): https://huggingface.co/Yuexuhug/DPLM/tree/main

Installation

We recommend using a conda environment.

conda create -n dplm python=3.9 -y
conda activate dplm
pip install -r requirements.txt

Pretrained Checkpoints

Download the required checkpoints from Hugging Face and place them in the checkpoint/ directory.

Task	Checkpoint file
Representation extraction	checkpoint_best_val_rmsf_cor.pth
IDR prediction	idr.pth
ΔΔG prediction	ddt.pth

Usage

1️⃣ Extracting DPLM Protein Representations

DPLM produces per-sequence embeddings that encode dynamic information.

Example

from utils.utils import *

model_location = './checkpoint/checkpoint_best_val_rmsf_cor.pth'
model_config = './config/config_vivit3.yaml'

input_data = [
    "MKTVRQERLKSIVRILERSKEPVSGAQLAEELSVSRQVIVQDIAYLRSLGYNIVATPRGYVLAGG",
    "KALTARQQEVFDLIRDHISQTGMPPTRAEIAQRLGFRSPNAAEEHLKALARKGVIEIVSGASRGIRLLQEE"
]

model, alphabet = load_model(model_config, model_location)

embeddings = []
for seq in input_data:
    embeddings.append(extract_emb_perseq(model, alphabet, seq))

Each entry in embeddings is a dynamics-aware protein representation.

2️⃣ Intrinsic Disorder Region (IDR) Prediction

DPLM can be adapted to residue-level IDR prediction using a lightweight prediction head.

Example

from model_idr import load_model_idr

config_path = './config/idr_config_30CAID2_trainfix_adp16_adp4.yaml'
model_location = './checkpoint/idr.pth'

model = load_model_idr(config_path, model_location)

sequence = "MKTVRQERLKSIVRILERSKEPVSGAQLAEELSVSRQVIVQDIAYLRSLGYNIVATPRGYVLAGG"
result = model([sequence])

Output: a per-residue disorder probability.

3️⃣ Protein Stability Change (ΔΔG) Prediction

DPLM can also be adapted for mutation-induced stability change prediction.

Example

from model_ddt import load_model_ddt

config_path = './config/ddt_config_adapterH16_adapterH4.yaml'
model_location = './checkpoint/ddt.pth'

model = load_model_ddt(config_path, model_location)

wild_seq = [
    "MKTVRQERLKSIVRILERSKEPVSGAQLAEELSVSRQVIVQDIAYLRSLGYNIVATPRGYVLAGG"
]
mut_seq = [
    "MKTVRQERLKSIVRILERSKEPVSGAQLKEELSVSRQVIVQDIAYLRSLGYNIVATPRGYVLAGG"
]

result = model(wild_seq, mut_seq)

Output: predicted ΔΔG value for the mutation.

Citation

If you use DPLM in your research, please cite:

[Add citation information here]