A variational expectation-maximization framework for balanced multi-scale learning of protein and drug interactions

Jiahua Rao, Jiancong Xie, Qianmu Yuan, Deqin Liu, Zhen Wang, Yutong Lu, Shuangjia Zheng, Yuedong Yang

TL;DR

This paper introduces MUSE, a novel framework that effectively integrates atomic-level structural data with broader molecular network information for predicting protein and drug interactions. By employing a variational Expectation-Maximization approach, MUSE overcomes the common challenge of imbalanced learning across different data scales, leading to significant improvements in prediction accuracy for both molecular interactions and atomic-level protein interfaces.

What Problem Does This Paper Address?

Understanding how proteins, drugs, and other biomolecules interact is fundamental to deciphering biological processes and developing new therapies. While computational methods exist, they often rely on either molecular network topology or detailed atomic structures, but rarely integrate both effectively. Existing multi-view learning approaches struggle because they tend to overemphasize one data scale while neglecting others, a problem attributed to the inherent "greediness" of learning algorithms and the imbalanced nature of the data. This leads to suboptimal performance and poor generalization. The authors aim to bridge this gap by developing a method that can harmoniously leverage information from different scales.

Methods

The core of MUSE is a variational Expectation-Maximization (EM) framework. This approach iteratively optimizes two key components:

  • Expectation Step (E-step): This step focuses on the atomic structure scale. It uses the detailed 3D structure of proteins and drugs to learn structural representations. Crucially, it also incorporates "pseudo-interactions" predicted by the other scale (M-step) to augment its learning.
  • Maximization Step (M-step): This step operates at the molecular network scale. It takes the structural embeddings from the E-step and the predicted interactions to learn and refine the overall molecular network.

This alternating procedure, with mutual supervision between the two scales, allows MUSE to balance their learning contributions. The framework uses graph neural networks (GNNs) for both scales: one for atomic structure representation and another for molecular network learning. The iterative optimization helps mitigate the imbalanced learning characteristics often seen in multi-scale approaches.

Datasets

MUSE was evaluated on several established benchmarks:

  • Protein-Protein Interactions (PPI): SHS27K dataset.
  • Drug-Protein Interactions (DPI): BioSNAP dataset.
  • Drug-Drug Interactions (DDI): DeepDDI dataset.
  • Protein Interface Contact Prediction: DIPS-Plus dataset.
  • Protein Binding Sites Prediction: ScanNets and Masif-site benchmarks.

The paper does not introduce new datasets but leverages existing ones to demonstrate MUSE's effectiveness across various interaction prediction tasks.

Key Insights and Ideas

  • The "Greediness" Problem: Multi-scale learning often suffers because one scale (usually the one with richer or more readily available data) learns much faster, dominating the learning process and hindering the effective utilization of information from other scales. MUSE directly tackles this.
  • EM for Balance: The EM framework provides a principled way to iteratively refine predictions from one scale using the other, creating a synergistic loop that balances their contributions.
  • Mutual Supervision is Key: The E-step uses predictions from the M-step (pseudo-interactions), and the M-step uses structural embeddings from the E-step. This mutual reliance ensures that both scales are actively contributing and learning from each other.
  • Beyond Interaction Prediction: MUSE's success in predicting protein interfaces at the atomic level demonstrates its versatility and the power of its multi-scale approach for tasks beyond just predicting whether things interact, but also how they interact at a fine-grained level.

Key Results

MUSE demonstrated state-of-the-art performance across multiple benchmarks:

  • Molecular Interactions: MUSE outperformed existing models on PPI, DPI, and DDI prediction tasks. For example, it achieved a 13.81% increase in micro-F1 score on the PPI dataset compared to the strong baseline HIGH-PPI. On the DDI dataset, it showed a 5.05% improvement over CGIB.
  • Protein Interface Prediction: MUSE also excelled in predicting protein interface contacts, achieving an AUROC of 0.92 and the highest precision on the DIPS-Plus benchmark.
  • Atomic Structure Scale: For protein binding site prediction, MUSE achieved a median AUPRC of 0.811 and a median AUROC of 0.938 on the ScanNets benchmark, outperforming other methods.
  • Utilization Rate Analysis: Experiments showed that MUSE significantly improved the "utilization rate" of both atomic structure and molecular network information compared to previous methods, indicating a more balanced and effective integration.

Limitations and Future Work

  • Handling Noisy Data: The authors acknowledge that MUSE's ability to handle noisy and incomplete multi-scale data could be further enhanced.
  • Extension to Other Scales: They suggest that the MUSE framework has potential for extension to other scales within computational drug discovery, such as atom and amino acid scales for protein representation learning.
  • Incorporating Prior Knowledge: Future work could explore combining MUSE with knowledge graphs and explainable AI techniques to further improve its robustness and interpretability.

Summary updated

Abstract

Protein functions are characterized by interactions with proteins, drugs, and other biomolecules. Understanding these interactions is essential for deciphering the molecular mechanisms underlying biological processes and developing new therapeutic strategies. Current computational methods mostly predict interactions based on either molecular network or structural information, without integrating them within a unified multi-scale framework. While a few multi-view learning methods are devoted to fusing the multi-scale information, these methods tend to rely intensively on a single scale and under-fitting the others, likely attributed to the imbalanced nature and inherent greediness of multi-scale learning. To alleviate the optimization imbalance, we present MUSE, a multi-scale representation learning framework based on a variant expectation maximization to optimize different scales in an alternating procedure over multiple iterations. This strategy efficiently fuses multi-scale information between atomic structure and molecular network scale through mutual supervision and iterative optimization. MUSE outperforms the current state-of-the-art models not only in molecular interaction (protein-protein, drug-protein, and drug-drug) tasks but also in protein interface prediction at the atomic structure scale. More importantly, the multi-scale learning framework shows potential for extension to other scales of computational drug discovery. Multi-scale learning still struggles with imbalanced information and greedy characteristics. Here the authors present MUSE, an Expectation-Maximization-based multi-scale framework, improving predictions across molecular interactions and atomic interfaces.
BibTeX
@inproceedings{rao2024variational,
  author = {Jiahua Rao and Jiancong Xie and Qianmu Yuan and Deqin Liu and Zhen Wang and Yutong Lu and Shuangjia Zheng and Yuedong Yang},
  title = {{A variational expectation-maximization framework for balanced multi-scale learning of protein and drug interactions}},
  booktitle = {Nature Communications},
  year = {2024},
  url = {https://openreview.net/forum?id=79RHXXgAnU},
}
Figure