Deep learning model for protein multi-label subcellular localization and function prediction based on multi-task collaborative training

Peihao Bai, Guanghui Li, Jiawei Luo, Cheng Liang

TL;DR

This paper introduces DeepMTC, a novel deep learning model that simultaneously predicts protein subcellular localization and function. By employing a multi-task collaborative training strategy and integrating structural and sequence information, DeepMTC achieves state-of-the-art performance, particularly for predicting functions of newly discovered proteins without prior annotations.

What Problem Does This Paper Address?

Understanding protein function and location is fundamental to biology, impacting drug discovery and disease research. Traditional methods for determining these properties often rely on laborious wet-lab experiments or extensive databases like Gene Ontology (GO). Existing computational models face challenges: sequence-based methods often ignore crucial structural information, structure-based methods struggle with long-range dependencies, and database-dependent methods fail for novel proteins. Furthermore, effectively integrating diverse protein information (sequence, structure, function) remains a significant hurdle. This paper aims to bridge this gap by developing a unified model that leverages multi-task learning to predict both subcellular localization and function, while reducing reliance on external databases.

Methods

DeepMTC employs a multi-task collaborative training framework. It starts by using a pre-trained language model (like ESM-Fold) to obtain a protein's 3D structure and extract sequence features. These features are then processed by a Graph Transformer (GT) module, which is designed to capture long-range dependencies within protein structures more effectively than traditional graph neural networks. Following this, a Graph Autoencoder (Gae_block) is used to extract functional features from the protein's graph representation. The core innovation lies in the Functional Cross-Attention (Fun_attention) block, which adaptively combines these learned functional features (Biological Process, Cellular Component, Molecular Function) to predict multi-label subcellular localization. This approach bypasses the need for direct GO database lookups during prediction.

Datasets

The study utilized data from the SwissProt and TrEMBL databases within UniProt for multi-label subcellular localization. GO annotations were sourced from EMBL-EBI. The dataset was curated to include proteins from human and mouse, with sequences up to 1200 amino acids, and focused on GO annotations with a frequency of ≥20 and 10 subcellular localization labels. This resulted in a dataset of 6083 protein sequences. The data was split into training (80%), validation (10%), and testing (10%) sets.

Key Insights and Ideas

  • Multi-task learning as a synergy: The paper demonstrates that jointly training for protein function prediction and subcellular localization not only improves performance on both tasks but also allows the model to learn more robust and generalizable protein representations.
  • Graph Transformer for structural context: The use of a Graph Transformer is crucial for effectively capturing long-range dependencies in protein structures, overcoming a common limitation of traditional graph neural networks.
  • Database-free functional integration: DeepMTC's ability to leverage learned functional features for subcellular localization, rather than relying on direct GO database queries, makes it particularly powerful for predicting the properties of novel or unannotated proteins.
  • Interpretability through attention: The model's attention mechanisms allow for visualization of residue importance, correlating well with known functional domains, which enhances trust and understanding of its predictions.

Key Results

DeepMTC achieved state-of-the-art performance on both protein function prediction and multi-label subcellular localization tasks. On the independent test set for subcellular localization, DeepMTC achieved an Average Precision (AP) of 0.8426, AUROC of 0.8124, and AUPR of 0.8415, outperforming established methods like DeepLoc 2.0 and GPSFun. For protein function prediction, AUROC values exceeded 80% for all three sub-ontologies (BP, CC, MF), with BP and CC reaching over 90%. Ablation studies showed that removing the Graph Transformer module had the most significant negative impact on performance, highlighting its importance.

Limitations and Future Work

The authors acknowledge that DeepMTC's performance is dependent on the accuracy of the 3D structures generated by pre-trained models. To improve efficiency, they plan to explore using only 3D structures during training to speed up testing. Future work will also involve incorporating protein-protein interaction (PPI) data to further enhance the model's understanding of protein function and exploring predictions for less frequent GO annotations to improve generalizability. The ultimate goal is to develop a generalized model that integrates multiple aspects of protein analysis, including function, localization, ligand binding, solubility, and dynamics.

Summary updated

BibTeX
@article{bai2024deep,
  author = {Peihao Bai and Guanghui Li and Jiawei Luo and Cheng Liang},
  title = {{Deep learning model for protein multi-label subcellular localization and function prediction based on multi-task collaborative training}},
  journal = {Briefings in Bioinformatics},
  year = {2024},
  url = {https://openreview.net/forum?id=1ZcrUQtP74},
}
Figure