DiffPPI: Conditional Denoising Diffusion Framework for Brain-Specific Protein-Protein Interaction Prediction

Pratibha Joshi

doi:10.1016/j.jocs.2026.103041

DiffPPI: Conditional Denoising Diffusion Framework for Brain-Specific Protein-Protein Interaction Prediction

2026-10-02

0

PRE

AI

DOI：10.1016/j.jocs.2026.103041

Original

Original request for help

Save

Abstract

En 中文

The protein-protein interactions form the molecular foundation of cellular organization and signaling, yet brain-specific interactomes remain incomplete due to experimental limitations and context-dependent variability. The existing computational approaches largely formulate PPI prediction as deterministic binary classification, which restricts their ability to model uncertainty in sparse and context-dependent biological data. To address this gap, we propose DiffPPI, a conditional diffusion-based framework for brain-specific PPI prediction that integrates graph-based protein embeddings with generative denoising. In this work, a curated brain PPI network comprising several thousand proteins and experimentally supported interactions is modeled as an undirected graph. The node-level representations are learned using a pre-trained GraphSAGE model, capturing both local topology and relational context. Pairwise features are encoded using concatenation, element-wise interaction, and absolute difference operators, and are used to condition a denoising diffusion model with 200 diffusion steps. The proposed DiffPPI learns to reconstruct clean interaction signals from noisy observations, enabling uncertainty-aware prediction. The framework is evaluated against classical machine-learning models and graph neural network baselines using accuracy, F1-score, and ROC-AUC. The proposed DiffPPI achieves superior performance, with accuracy 0.8397 ± 0.0019 and F1-score close to 0.8291 ± 0.0031 and a ROC-AUC of approximately 0.9023 ± 0.0014, outperforming all baselines trained on identical data splits. Additionally, explainability analysis using Integrated Gradients highlights the dominant contribution of embedding difference features.

Keywords:

Protein-Protein Interaction Prediction

Diffusion Models

Graph Neural Networks

Protein Embeddings

Integrated Gradients

Explainable Artificial Intelligence

DiffPPI: Conditional Denoising Diffusion Framework for Brain-Specific Protein-Protein Interaction Prediction

Abstract

Journal

Organization

Cited Papers