yk0/tuna-r-tuna

TUnA Protein-Protein Interaction Predictor

Model Description

TUnA is a sequence-based protein–protein interaction (PPI) predictor that uses a Transformer backbone with a last-layer Gaussian process (LLGP). It produces a probability score for whether two proteins interact.

Intended Use

• Rapid inference for candidate PPI scoring.
• These weights can be loaded in from the TUnA-R repository on Github.

Training Data

Trained on the Bernett dataset (https://academic.oup.com/bib/article/25/2/bbae076/7621029) composed of Human PPI interactions.

Model Architecture

Backbone: tuna.models._transformer.Transformer

• protein_dim: 640
• hid_dim: 64
• ff_dim: 256
• n_layers: 1
• n_heads: 8
• dropout: 0.2
• llgp: True
• use_spectral_norm: True
• out_targets: 1
• gp_config:
  • rff_features: 4096
  • gp_cov_momentum: -1
  • gp_ridge_penalty: 1
  • likelihood: "binary_logistic"

Evaluation Results (Test Set)

• AUROC: 0.70
• MCC: 0.30
• Accuracy: 0.65
• AUPRC: 0.69
• Precision: 0.65

Citation

If you use this model, please cite:

• TUnA: https://academic.oup.com/bib/article/25/5/bbae359/7720609
• Bernett dataset: https://academic.oup.com/bib/article/25/2/bbae076/7621029

License

MIT