Protein Representation Learning Achieves Enhanced Accuracy with Adaptive Multi-Perspective Graph Fusion

Protein representation learning benefits greatly from graph neural networks, which naturally model residue interactions, but current methods often rely on limited perspectives of these interactions, leading to incomplete protein models. Yusong Wang of the Guangdong Institute of Intelligence Science and Technology, alongside Jialun Shen, Zhihao Wu, and Shiyin Tan of the Institute of Science Tokyo, et al., introduce a novel framework called MMPG to address this challenge. MMPG constructs protein graphs from physical, chemical, and geometric viewpoints, then adaptively fuses these perspectives using a Mixture of Experts approach. This innovative use of MoE allows the model to specialise in capturing varying levels of interaction, from individual residue features to complex cross-perspective synergies, ultimately generating more robust and accurate protein representations. Demonstrating superior performance across four downstream protein tasks, MMPG represents a significant advancement in the field, promising improvements in protein function prediction and structural analysis.

The research addresses limitations inherent in current Graph Neural Network (GNN)-based methods, which typically rely on single-perspective graph construction and consequently capture incomplete information about residue interactions within proteins. MMPG overcomes this by building protein graphs representing physical, chemical, and geometric properties, offering a more holistic view of residue relationships. This innovative approach allows for a more comprehensive understanding of protein structure and function, ultimately leading to improved performance in downstream applications.

The team achieved this breakthrough by developing a Mixture of Experts (MoE) module, designed to adaptively fuse these multi-perspective graphs. This module dynamically routes different perspectives to specialized experts, enabling them to learn both perspective-specific features and the crucial synergies between them. Quantitative analysis confirms that the MoE module automatically specializes, effectively modelling interaction levels ranging from individual residue representations to complex, inter-perspective relationships and ultimately achieving a global consensus across all perspectives. By integrating this multi-level information, MMPG generates superior protein representations capable of discerning subtle but critical details.

This study unveils a new paradigm for protein representation learning, moving beyond the constraints of single-perspective approaches. Experiments show that MMPG effectively captures the complex interplay of physical, chemical, and geometric factors governing residue interactions, something previous methods often overlooked. The researchers quantitatively verified that the MoE module specializes in modelling distinct levels of interaction, from individual representations to pairwise synergies and a global consensus. This capability allows MMPG to produce more expressive protein representations, significantly enhancing performance on four distinct downstream protein tasks.

The research establishes a clear pathway towards more accurate and robust protein representation learning. By constructing graphs from physical, chemical, and geometric perspectives, and then intelligently fusing them with the MoE module, MMPG delivers a comprehensive understanding of residue interactions. This work opens new possibilities for applications in crucial areas such as drug discovery, functional annotation, and protein design, offering a powerful tool for advancing our understanding of biological systems and accelerating innovation in related fields. The superior performance across multiple protein tasks demonstrates the effectiveness and versatility of this innovative framework.

Multi-Perspective Graphs with Mixture of Experts

The research team pioneered a novel framework, MMPG, to address limitations in current protein representation learning (PRL) techniques. Existing graph neural network (GNN)-based methods typically rely on constructing protein interaction graphs from a single perspective, potentially overlooking crucial residue interactions and resulting in incomplete protein representations. MMPG overcomes this by constructing protein graphs from three distinct perspectives, physical, chemical, and geometric, each designed to characterise different properties of residue interactions and capture a more holistic view of protein structure and function. These perspectives are not treated equally, but rather integrated through an innovative Mixture of Experts (MoE) module.

Scientists engineered the MoE module to dynamically route each perspective to specialized experts, enabling the learning of both perspective-specific features and their complex synergies. This approach allows experts to model interactions at multiple levels, ranging from individual representations to pairwise inter-perspective relationships and ultimately, a global consensus across all perspectives. Quantitative verification confirmed that the MoE module automatically specialises these experts in modelling distinct levels of interaction, demonstrating its ability to effectively integrate multi-level information. The physical perspective utilizes knowledge-based potentials to capture interaction stability, while the chemical perspective encodes residue similarities based on biochemical properties and the geometric perspective models local spatial relationships.

Experiments employed these three graph constructions, physical-energetic, chemical-functional, and geometric-structural, to create a comprehensive representation of residue interactions. The system delivers a nuanced understanding of protein properties by recognising that these perspectives are interdependent; for example, a geometric graph identifies approximate neighbours, validated by the stability assessment of a physical-energetic graph. This multi-perspective integration significantly enhances the expressive power of learned protein representations, achieving advanced performance across four different downstream protein tasks and demonstrating a substantial improvement over single-perspective approaches. The methodology enables a more accurate and complete understanding of protein structure and function, paving the way for advancements in areas such as drug discovery and protein design.

Multimodal Protein Representation Learning Excels at Prediction

Scientists have developed a new framework, MMPG, for protein representation learning that achieves advanced performance across multiple protein-related tasks. The research demonstrates that constructing protein models from multiple perspectives, physical, chemical, and geometric, and adaptively fusing them using a Mixture of Experts (MoE) approach yields superior results. Experiments revealed that MMPG accurately models enzymatic function and cellular roles by integrating these diverse information sources. The team measured performance on four distinct protein tasks: Protein Fold Classification, Enzyme Reaction Classification, Gene Ontology Term Prediction, and Enzyme Commission number prediction.

Notably, MMPG achieved a peak score of 0.893 on the Enzyme Commission task and 0.489 on the Gene Ontology, Cellular Component prediction, indicating a high degree of accuracy in modelling enzymatic function and cellular roles. Data shows that MMPG outperforms single-perspective graph construction strategies, proving that combining physical, chemical, and geometric information creates more expressive protein representations. Further analysis confirmed the importance of each perspective used in the MMPG framework. Removing any single perspective, physical, chemical, or geometric, resulted in a significant performance decrease across all tasks.

The MoE module also proved crucial, outperforming alternative fusion strategies like edge stacking and simple concatenation. Tests prove that the MoE dynamically routes perspectives to specialized experts, learning intrinsic features and cross-perspective interactions at multiple levels, from individual representations to global consensus. Ablation studies validated each design component of MMPG, demonstrating the necessity of each perspective and the MoE fusion module. Parameter analysis showed that optimal graph construction hyperparameters align with the level of informational granularity required by each task, with energy thresholds ranging from -2.0 to -0.5 depending on the specific prediction goal. The work delivers a significant advancement in protein representation learning, offering potential for improved accuracy in a range of biological applications.