PPI-OMEGA

Methods

🟢 Overview

The PPI-OMEGA framework is designed to enhance protein-protein interaction (PPI) predictions by integrating multi-omics data into a graph-based learning model. We leverage a Variational Graph Autoencoder (VGAE) to capture complex dependencies in PPI networks, using RNA expression and immunohistochemistry (IHC) protein expression as node attributes.

🟢 Dataset & Preprocessing

We constructed our PPI dataset using the STRING database, which contains interaction confidence scores for protein pairs. The dataset was filtered to retain the top 5% high-confidence interactions and mapped to human gene identifiers.

Additional multi-omics features were integrated from The Human Protein Atlas:

• RNA Expression Data: Processed from 35 tissue types, normalized using TPM values.
• IHC Protein Expression Data: Derived from 45 tissues, discretized into 4 bins of different expression levels.

Principal component analysis (PCA) was conducted separately for normalized RNA expression and protein expression features to reduce dimensionality.

🟢 Graph Construction

Each protein was represented as a node, and edges were formed based on STRING database interactions. Node attributes included:

• Structural Features: Encoded from known PPI networks.
• Node Feature Data: Preprocessed RNA and/or IHC protein values as needed.

🟢 Variational Graph Autoencoder (VGAE) Model

Our model follows a standard VGAE framework:

• Encoder: Two-layer Graph Convolutional Network (GCN) extracts node embeddings.
• Latent Space: A probabilistic layer learns a Gaussian distribution for protein representations.
• Decoder: Dot product of embeddings reconstructs the adjacency matrix.

🟢 Hyperparameter Tuning

To optimize model performance, we performed a grid search over key hyperparameters, including the dropout rate, learning rate, and weight decay. Each configuration was evaluated on the validation set, selecting the combination that yielded the highest AUROC and AP scores.

• Dropout Rate (p): {0.3, 0.4, 0.5}
• Learning Rate (α): {0.001, 0.005, 0.01}
• Weight Decay (λ): {5e-4, 1e-3, 5e-3}

The best-performing configuration was found to be:

• Dropout Rate: 0.3
• Learning Rate: 0.01
• Weight Decay: 5e-4

This configuration resulted in the highest AUROC (0.9235) and AP (0.9318), significantly improving predictive accuracy compared to suboptimal hyperparameter choices.

To ensure robustness, we implemented an early stopping strategy based on validation AUROC. If no improvement of at least 0.01 was observed for 60 consecutive epochs, training was halted to prevent overfitting. Additionally, dropout regularization was applied to mitigate overfitting while ensuring stable convergence.

🟢 Training & Evaluation

The model is trained using a binary cross-entropy loss with a Kullback-Leibler (KL) divergence regularization term: \(\mathcal{L} = \mathcal{L}_{recon} + \beta \mathcal{L}_{KL}\). We used the Adam optimizer with a learning rate of 0.001. Training was stopped when validation loss did not improve by a minimum threshold within a set number of epochs.

🟢 Ablation Study

To assess the impact of multi-omics data integration, we conducted an ablation study during which we trained and evaluated the model under four different feature input conditions:

Table 1: Description of Different Feature Sets used in Ablation Study

Feature Set	Description
No Features	Only the network structure (adjacency matrix) was used as input.
RNA Only	RNA expression profiles were included as node attributes.
Protein Only	IHC protein expression levels were included as node attributes.
Combined (RNA + Protein)	Both RNA expression and IHC protein expression features were integrated.

The performance of each model variant was evaluated using AUROC and AP (Average Precision) scores.

Read more about our method architecture and training details on our report.

Results

To evaluate the impact of different input feature combinations on model performance, we conducted an ablation study. Table 2 presents AUROC and AP scores for four configurations: using no features, using only RNA expression, using only protein expression, and combining both features.

Table 2: Comparison of AUROC and AP from Ablation Study with Different Feature Sets

RNA Exp.	IHC Protein Exp.	AUROC	AP
❌	❌	0.8194	0.8280
✅	❌	0.8888	0.9026
❌	✅	0.9215	0.9297
✅	✅	0.9235	0.9318

🟢 Model Performance Analysis

Our results demonstrate that integrating both RNA and protein expression features led to the highest performance, with an AUROC of 0.9235 and an AP of 0.9318. This confirms that incorporating multi-omics data enhances PPI prediction compared to using RNA or protein expression alone.

Notably, the model trained with protein expression (IHC data) achieved an AUROC of 0.9215 and an AP of 0.9297, surpassing the RNA-only model (AUROC = 0.8888, AP = 0.9026). This suggests that protein expression features provide a stronger signal for identifying functional interactions. Meanwhile, the addition of RNA expression demonstrates diminishing return, which highlights needs to look for potential more relevant data.

🟢 Training Dynamics

To evaluate model convergence and performance stability, we tracked training dynamics by monitoring AP (Average Precision) scores across epochs (Figure 3). The multi-omics model, which integrates both RNA and protein expression features, demonstrated faster convergence and higher final AP scores compared to single-feature models.

The results indicate that:

• The multi-omics model achieves the highest AP score and converges faster than RNA-only and protein-only models.
• The protein-only model performs almost as well as the multi-omics model, reinforcing the importance of protein expression features in PPI prediction.
• The RNA-only model lags in both convergence speed and final AP, suggesting that RNA expression alone provides limited predictive power.
• The model trained without node features struggles with stability and achieves the lowest AP score, confirming that incorporating biological features significantly improves PPI predictions.

🟢 Latent Representation Analysis

To assess the biological relevance of the learned protein embeddings, we examined how the latent representations evolved during training. The model successfully captures biologically meaningful relationships, distinguishing between housekeeping and context-dependent proteins and structuring embeddings based on network connectivity.

As shown in Figure 4, we visualized the UMAP projections of protein embeddings at different training epochs, classifying proteins as either housekeeping or context-dependent based on functional annotations. The results show that:

• The separation between housekeeping and context-dependent proteins increases over training, indicating that the model progressively refines biologically meaningful representations.
• Some overlap remains, likely reflecting functional interactions between proteins that span both categories.

To further investigate how network structure influences embedding formation, we analyzed the variance of the latent representation mean (μ) for highly connected versus sparsely connected proteins (Figure 5). The findings suggest that:

• Highly connected proteins converge faster, with their embeddings stabilizing early in training due to strong structural constraints.
• Sparsely connected proteins exhibit greater variance fluctuations before stabilizing, requiring more training epochs to reach a meaningful latent space representation.

Read more detailed results and interpretation on our report.

Discussion & Future Work

🟢 Enhancing PPI Prediction with Multi-Omics

Predicting protein-protein interactions (PPIs) is essential for understanding biological systems, disease mechanisms, and drug discovery. Traditional models primarily rely on sequence data, but our study demonstrates that incorporating RNA and protein expression significantly improves prediction accuracy. This multi-omics approach captures functional relationships beyond what sequence data alone can reveal. Future work will include integrating protein sequence features alongside multiomics data to build a more comprehensive and accurate PPI prediction framework.

🟢 Handling Dropped or Unused Data

Preprocessing decisions impact the quality of PPI predictions. While thresholding and PCA reduce noise, they may also remove informative low-confidence PPIs or rare expression patterns. To improve data retention and maintain biological signal integrity, future work will explore:

• Adaptive thresholding to refine data inclusion dynamically.
• Data imputation to recover missing but meaningful information.
• Feature selection to balance dimensionality reduction and biological relevance.

🟢 Improving Protein Expression Encoding

Our reliance on immunohistochemistry (IHC) data introduces variability in protein quantification. More precise techniques will enhance the robustness of our model, including:

• Using mass spectrometry-based proteomics for accurate protein quantification.
• Normalizing RNA-protein expression values across datasets.
• Incorporating time-series expression data to capture dynamic PPI interactions.
• Exploring single-cell transcriptomics to improve resolution in cellular studies.

🟢 Integrating Protein Sequence Features

While our study focused on multi-omics data, protein sequence information remains a fundamental component of PPI prediction. Many models leverage structural motifs and conserved domains for interaction analysis. Future work will integrate sequence-based features with multi-omics data to enhance accuracy and uncover novel PPIs relevant to disease research.

🟢 Advancing Model Architecture

Improving model interpretability and predictive power requires architectural refinements. Planned enhancements include:

• Implementing attention mechanisms to prioritize biologically significant interactions.
• Expanding multi-omics integration to include epigenetic and metabolic profiles.
• Addressing class imbalance to minimize false positives in real-world applications.
• Enhancing generalizability across tissues for precision medicine applications.
• Exploring graph-based deep learning to model complex biological networks.

🟢 Real-World Impact

More accurate PPI prediction models have broad applications in biomedical research. Our findings contribute to:

• Drug discovery – Identifying novel drug targets and reducing off-target effects.
• Disease modeling – Mapping interaction networks in genetic disorders and cancers.
• Precision medicine – Personalizing treatments based on molecular profiles.
• Systems biology – Providing deeper insights into cellular mechanisms.

Moving forward, we will expand our datasets, refine our methodologies, and collaborate with experimental biologists to validate findings. By bridging computational predictions with experimental validation, we aim to create a biologically informed AI-driven tool for PPI analysis, driving new discoveries in molecular biology and therapeutic research.