Source code for Seq2Symm: Rapid and accurate prediction of protein homo-oligomer symmetry, published here:

Meghana Kshirsagar, Artur Meller, Ian R. Humphreys, Samuel Sledzieski, Yixi Xu, Rahul Dodhia, Eric Horvitz, Bonnie Berger, Gregory R. Bowman, Juan Lavista Ferres, David Baker, Minkyung Baek, Nature Communications

https://www.nature.com/articles/s41467-025-57148-3

Seq2Symm takes a protein sequence as input and predicts the protein symmetry, such as 'C1', 'C2', 'C5', 'D2', 'D3', 'D5', 'H', 'O', 'T', 'I' etc. for 17 different symmetry types.

There are two options: do inference using Google Colab or a local installation where you can do both inference and/or training

This self-contained notebook will do inference on input protein sequences

https://colab.research.google.com/drive/1ptQTyC22ExxJ3BnSK6dnPaCeg7J8i3le#scrollTo=CJm12VgOiDfj

Dependencies are in the yaml file esm2_finetune.yaml

conda env create --name esm2 --file=esm2_finetune.yaml

0. Latest model trained in March 2025 (please look under 3. for the model used for all the results in the paper) https://drive.google.com/file/d/1Hk4idiCGNnn5B0KUTCsnMtv1vtvWb6lh/view?usp=sharing

1. the predictions from the model on various datasets, predictions on proteomes http://files.ipd.uw.edu/pub/seq2symm/predictions.zip

2. Seq2Symm model from the Nature Communications, 2025 paper http://files.ipd.uw.edu/pub/seq2symm/ESM2_model.ckpt

3. data download links are here https://github.com/microsoft/seq2symm/tree/main/datasets

   All code, datasets, model, predictions are also available on Zenodo: http://doi.org/10.5281/zenodo.14659968

python src/finetune.py --meta_data_file ../datasets/homomer_pdbids_hash_clusterid_labels_sampled.csv --data_dir ../datasets/ --model_dir models/ --output_model seq2symm --output_dir outputs/seq2symm --bs 16 --data_splits_file ../datasets/train_val_test_splits.pkl --limit 65536 --granularity 3 --n_classes 20 --n_epoch 100 --weighted_sampler 1

A jupyter notebook is available at src/load_chkpt_and_predict.ipynb that shows examples of how this is done for two different file formats

This script is available courtesy of Moritz Ertlet:

The src/predict_oligmerization.py script predicts the oligomerization states of protein sequences provided in a FASTA file, outputting the probabilities for each symmetry state with a probability >= 1%.

python predict_oligomerization.py -input_file <input_fasta> -chkpt_file <model_checkpoint> [-output_file <results.csv>] [-batch_size <n>]

• -input_file: Path to the input FASTA file containing the sequences for prediction (required).
• -chkpt_file: Path to the model checkpoint file (required).
• -output_file: Path to save the prediction results as a CSV file (default: ./results.csv).
• -batch_size: Batch size for processing sequences (default: 1).

Example output:

fasta_id, predicted_oligomerization_state
ID1, "{'C1': 0.9, 'C2': 0.1, 'Other': 0.001}"
ID2, "{'C4': 0.85, 'C6': 0.15, 'D3': 0.01}"

Which you can load with pandas like this:

import pandas as pd
import ast
df = pd.read_csv("results.csv")
df['predicted_oligomerization_state'] = df['predicted_oligomerization_state'].apply(ast.literal_eval)