xiRT 2.0

Retention-time prediction for linear and crosslinked peptides using PyTorch. The default model learns modification-aware peptide embeddings, shares a recurrent encoder between crosslink branches, and predicts continuous RTs or fraction labels. It runs on CPU or CUDA.

Install and run

Requires Python 3.12–3.14; the default development interpreter is 3.13. The PyPI distribution is xirt-torch; the Python import and CLI remain xirt.

uv sync --locked --extra cpu
uv run --no-sync xirt -p xirt_params.yaml -s learning_params.yaml
# Edit the generated configurations for your input columns and targets.
uv run --no-sync xirt -i psms.csv -o results -x xirt_params.yaml -l learning_params.yaml

Replace --extra cpu with --extra cuda for the locked CUDA backend. Optional extras provide local MLflow tracking (tracking), TensorBoard (tensorboard), Lion (experiments), and local ESM-C embeddings (esm, Python 3.12 only). ESM model access and licensing are separate.

Fine-tuning

Native recurrent models support full fine-tuning, frozen-encoder head training, and GRU-weight LoKr or LoRA. Checkpoints retain architecture and adapter settings. See fine-tuning for the API, YAML and measured comparison.

TensorFlow is no longer an application dependency or supported backend. Historical TF and PyTorch predictions, checkpoint provenance, tables and plots are preserved under benchmark-tf-torch/<cohort>/results/. PyTorch had lower mean held-out RMSE on two of five evaluated cohorts, not all five. See benchmark results. Keras .h5 models cannot be loaded by xiRT 2.0.

Examples

CPU notebooks cover training, checkpoint-based prediction and all four fine-tuning strategies with 2,048 observed peptide pairs. Training uses up to 150 epochs with validation-based early stopping:

uv run --no-sync marimo edit examples/train_predict.py --host 127.0.0.1
uv run --no-sync marimo edit examples/fine_tune.py --host 127.0.0.1
# Run both without a notebook server:
make examples

Development

make sync
make check
make build

pixi run sync bootstraps the same locked uv environment. make help lists documentation, notebook, test and release commands. Benchmark code lives in xirt/data_pipelines/benchmarks/ and is excluded from the production wheel.

Documentation · Configuration · Migration · Development

Citation

Giese, S.H., Sinn, L.R., Wegner, F. et al. Retention time prediction using neural networks increases identifications in crosslinking mass spectrometry. Nature Communications 12, 3237 (2021). https://doi.org/10.1038/s41467-021-23441-0

Contributors: Sven Giese and Ludwig Sinn.