DynamicMPNN

DynamicMPNN is a deep learning inverse folding model that generates protein sequences compatible with multiple conformational states. Built on GVP-GNN architecture, it pools structural information across conformations during encoding and uses autoregressive decoding to predict amino acid sequences. We introduce a multi-state self-consistency metric using template-based AlphaFold3 structure prediction with decoy normalization to evaluate design success.

Installation

conda env create -f environment.yml
conda activate dynamicmpnn
pip install -e .

Configuration

Create a .env file from the template and update paths for your machine:

cp .env.example .env

Required paths:

• PROJECT_PATH - Path to this repository
• PUBLIC_DB - Directory containing processed .pt files (e.g., data/)
• RUNS_PATH - Where training runs/checkpoints are saved

Optional (for AF3 evaluation):

• AF3_EXECUTABLE, AF3_SCRIPT, AF3_MODEL_DIR, AF3_DB_DIR

Data

Download processed .pt files from Zenodo and extract to data/:

# Multi-chain
data/train_pt_multi_chain/
data/val_pt_multi_chain/
data/test_pt_multi_chain/

# Single-chain
data/train_pt_single_chain/
data/val_pt_single_chain/
data/test_pt_single_chain/

Update PUBLIC_DB in .env to point to data/.

Checkpoints

Pre-trained model checkpoints are available in the checkpoints/ directory. Each checkpoint contains model weights, hyperparameters, and optimizer state.

Checkpoint	Description
multi_chain_reload.ckpt	Multi-chain mode, 2-state with full PDB encoding, per-epoch cluster resampling
multi_chain_no_reload.ckpt	Multi-chain mode, 2-state with full PDB encoding, static cluster sampling
single_chain_k*.ckpt	Single-chain mode, supports up to 5 conformational states (k=2,3,5 available)

Training

Setup

1. Download processed .pt files from Zenodo to data/
2. Configure paths in .env

Experiments

Experiment	Description
seq30_reload	Multi-chain, per-epoch cluster resampling
seq30_no_reload	Multi-chain, static cluster sampling
single_chain_k_conf	Single-chain, k conformations pooling

Commands

cd src/dynamicmpnn

# Full training
python train.py experiment=seq30_reload

# Quick test run
python train.py experiment=seq30_reload trainer.max_epochs=1 trainer.limit_train_batches=10

# Resume from checkpoint
python train.py experiment=seq30_reload ckpt_path=path/to/checkpoint.ckpt

Evaluation

Sequence Sampling

Generate sequences for a multi-state target:

dynamicmpnn-evaluate \
  eval=1bdt_1qtg \
  eval.model_ref=checkpoints/multi_chain_reload.ckpt \
  output_dir=runs/1bdt_1qtg

Key parameters:

• eval.num_samples: Number of sequences to generate (default: 25)
• eval.temperature: Sampling temperature

Output: samples/samples.csv and samples/samples.fasta

See example configs in src/dynamicmpnn/configs/eval/ for custom targets.

AF3 Self-Consistency Metrics

Enable AlphaFold3 structure prediction to compute self-consistency metrics:

dynamicmpnn-evaluate \
  eval=1bdt_1qtg \
  eval.model_ref=checkpoints/multi_chain_reload.ckpt \
  eval.af3_evaluate=true \
  output_dir=runs/1bdt_1qtg

Requires AF3 paths configured in .env.

Metrics computed:

• TM-score: Template modeling score (Kabsch-aligned)
• LDDT: Local distance difference test
• RMSD: Root mean square deviation
• pLDDT: Predicted local distance difference test

Decoy normalization is used for fair comparison. Output: summary_af3_results.csv

Data Preprocessing (optional)

To regenerate .pt files from scratch, start from the AlphaFold mmCIF database. See src/dynamicmpnn/scripts/README.md for the full pipeline.

Citation

@inproceedings{abrudan_dynamicmpnn,
  title={Multi-state Protein Sequence Design with DynamicMPNN},
  author={Abrudan, Alex and Ojeda, Sebastian Pujalte and Joshi, Chaitanya K and Greenig, Matthew and Engelberger, Felipe and Khmelinskaia, Alena and Meiler, Jens and Vendruscolo, Michele and Knowles, Tuomas},
  booktitle={International Conference on Learning Representations (ICLR)},
  year={2026}
}