UdonPred: Untangling Protein Intrinsic Disorder Prediction

UdonPred is a set of models for predicting disorder based on different definitions of disorder, reaching state-of-the-art performance using simple, ProstT5-based embeddings. The ability of models trained on one definition of disorder to generalise to others was accessed in the corresponding publication.

The training data can be found on here.

For the old version submitted to CAID3 go to https://github.com/jschlensok/udonpred.

Installation

1. git clone https://github.com/DavidWagemann/UdonPred.git
2. cd UdonPred
3. uv sync

Usage

uv run predict.py {path to fasta} {path to weights}

You can use the following options:

• --target: chooses the model trained on the specified dataset (trizod, chezod, softdis, pdbflex, atlas, plddt, disprot). The default is trizod.
• --output: sets the output directory path. Each sequence will be saved as a .caid file. The output will be written to the terminal if this is not set.
• --batch-size: sets the total sequence length per batch. Try reducing this if you get an out of memory error.
• --device: sets the device used for inference (cpu or cuda). Uses cuda by default if available.
• --smooth: Applies gaussian smoothing with the give sigma to the results in order to remove prediction noise.

Training a Model

UdonPred can be retrained by placing the required data as jsonl files in a data/ subfolder and pointing to it in config/data.yaml. The training configuration and architecture can be changed in config/config.yaml and config/architecture.yaml respectively. To start the training process, run uv run run.py train.

After training is complete, a checkpoint can be exported for use with the prediction script using uv run export.py {path to checkpoint}. For export options, see uv run export.py --help.

How to Cite

@article {UdonPred,
	author = {Schlensok, Julius and Wagemann, David and Senoner, Tobias and Haak, Markus and Rost, Burkhard},
	title = {UdonPred: Untangling Protein Intrinsic Disorder Prediction},
	elocation-id = {2026.01.26.701679},
	year = {2026},
	doi = {10.64898/2026.01.26.701679},
	publisher = {Cold Spring Harbor Laboratory},
	URL = {https://www.biorxiv.org/content/early/2026/01/28/2026.01.26.701679},
	eprint = {https://www.biorxiv.org/content/early/2026/01/28/2026.01.26.701679.full.pdf},
	journal = {bioRxiv}
}