| --- |
| language: |
| - en |
|
|
| license: mit |
|
|
| library_name: pytorch |
|
|
| tags: |
| - bioinformatics |
| - protein |
| - protein-protein-interaction |
| - interface-prediction |
| - partner-aware |
| - sequence-based |
| - esm2 |
| - cross-attention |
|
|
| pipeline_tag: token-classification |
| --- |
| |
| # SPPIDER-seq |
|
|
| **SPPIDER-seq** is a deep learning framework for **partner-aware prediction of protein–protein interaction (PPI) sites directly from amino acid sequences**. Given a pair of interacting proteins, the models predict residue-level interface propensities while explicitly accounting for the identity of the interaction partner. |
|
|
| ## Models |
|
|
| This repository contains two complementary pretrained models: |
|
|
| - **Peptide-centric model** – predicts interface residues on the peptide/ligand protein using the receptor sequence as context. |
| - **Receptor-centric model** – predicts interface residues on the receptor protein using the peptide/ligand sequence as context. |
|
|
| ## Model Architecture |
|
|
| Both models share the same neural network architecture described in the SPPIDER-seq publication: |
|
|
| - ESM-2 (650M) protein language model embeddings |
| - Multi-head cross-attention network |
| - Multi-layer perceptron (MLP) prediction head |
|
|
| ## Software |
|
|
| SPPIDER-seq software, command-line interface, and notebooks are available at: |
|
|
| https://github.com/aporollo-lab/SPPIDER-seq |
|
|
| ## Citation |
|
|
| If you use these models in your research, please cite: |
|
|
| > Porollo A, Jadhav O, Alvarez A, Chen J. **SPPIDER-seq: sequence-based partner-aware predictor of protein–protein interaction sites.** *Bioinformatics*. 2026;42(7):btag388. https://doi.org/10.1093/bioinformatics/btag388 |
|
|
| ## Model Version |
|
|
| **Version:** v2.2.0 |
|
|
| ## Notes for Version 2.2.0 |
|
|
| The neural network architecture is unchanged from the published SPPIDER-seq model. |
|
|
| Compared with the models accompanying the original publication, the v2.2.0 models were retrained using the complete curated datasets described in the manuscript. Training employed a 90%/10% train/validation split to maximize the amount of data available for model fitting while retaining an independent validation set for model selection. |
|
|
| Although overall predictive performance is expected to remain comparable to that reported in the publication, the updated models generally produce higher interface propensity scores. In addition, the peptide-centric model exhibits improved sensitivity toward interaction sites located within intrinsically disordered regions, consistent with observations made during internal benchmarking. |
|
|
