MD-Transformer: Multimodal Integration of ProtBERT Embeddings and Physicochemical Descriptors for Protein-Protein...

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Machine Learning

Background:

• Accurate prediction of protein-protein interaction (PPI) interface residues is crucial for understanding molecular mechanisms and guiding drug design.
• Existing methods often rely solely on sequence or structural information, potentially missing complementary signals.

Purpose of the Study:

• To develop a multimodal framework, MD-Transformer, that integrates contextual sequence representations with structure-related physicochemical information for improved PPI interface residue prediction.
• To evaluate the contribution of physicochemical descriptors to prediction accuracy and identify their role in reducing false positives.

Main Methods:

• MD-Transformer combines residue-level ProtBERT embeddings with physicochemical descriptors (B-factor, solvent-accessible surface area (SASA), hydrophobicity).
• A hybrid fusion module aligns features, followed by Transformer encoding and cross-modal attention for multimodal integration.
• The model was evaluated on the DB5.5 benchmark using complex-level and homology-aware split protocols.

Main Results:

• MD-Transformer achieved a high AUPRC of 0.564 on the complex-level split, outperforming an ablation model lacking physicochemical descriptors by 0.159.
• The model maintained strong performance (AUPRC=0.480, MCC=0.242) under the homology-aware split, demonstrating robustness to reduced sequence similarity.
• SASA was identified as a key descriptor for reducing false-positive predictions, particularly for exposed residues, though a precision-recall trade-off was observed.

Conclusions:

• Integrating contextual sequence representations with residue-level physicochemical descriptors provides complementary signals for accurate PPI interface prediction.
• MD-Transformer offers a powerful approach for enhancing the understanding of molecular recognition and supporting structure-guided design.
• Physicochemical properties, especially SASA, play a significant role in refining predictions and reducing errors in diverse residue exposure environments.