A comprehensive review and comparison of existing computational methods for protein function prediction

Baohui Lin, Xiaoling Luo, Yumeng Liu, Xiaopeng JinBriefings in Bioinformatics 2024 · DOI

TL;DR

This paper provides a comprehensive review and comparison of computational methods for predicting protein function, categorizing them into sequence-based, structure-based, PPI network-based, and hybrid approaches. It highlights the evolution of these methods, their strengths and weaknesses, and offers insights into the current state and future directions of this critical field in bioinformatics.

What Problem Does This Paper Address?

Understanding protein function is fundamental to deciphering biological processes, disease mechanisms, and developing new therapeutics. While experimental methods are accurate, they are time-consuming and costly. This has led to the development of numerous computational methods for predicting protein function. However, the rapid growth of these methods necessitates a systematic review and comparison to understand their relative merits, limitations, and to guide future research. This paper aims to bridge this gap by providing an updated, comprehensive overview and comparative analysis of existing computational approaches.

Methods

The paper categorizes computational methods into four main groups:

  • Sequence-based methods: These methods leverage the amino acid sequence of a protein to infer its function. Techniques include traditional sequence alignment (like BLAST) and more advanced deep learning approaches such as Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and Transformer-based models (e.g., protein language models). These methods often use embeddings like one-hot encoding, SPMap, or Word2Vec.
  • 3D structure-based methods: These approaches utilize the three-dimensional structure of a protein. Proteins are often represented as contact maps, and features are extracted from these structures. Graph Convolutional Networks (GCNs) and Graph Attention Networks (GATs) are commonly employed, often integrating sequence information as well. The accuracy of predicted structures (e.g., from AlphaFold) is crucial for these methods.
  • PPI network-based methods: These methods focus on protein-protein interaction (PPI) networks. Algorithms like GeneMANIA and Mashup integrate network information to predict function. Techniques include label propagation, network diffusion, and autoencoders for feature extraction.
  • Hybrid information-based methods: This category represents the current trend, integrating multiple data sources such as sequences, structures, PPI networks, and domain information (e.g., from InterPro). These methods aim to leverage the complementary nature of different data types to improve prediction accuracy. Deep learning architectures, including multi-layer perceptrons, CNNs, and GNNs, are frequently used for feature fusion and prediction.

Datasets

The paper discusses several key databases crucial for protein function prediction:

  • UniProt: A comprehensive protein sequence and functional information database.
  • Gene Ontology (GO): Provides a structured vocabulary for describing gene and protein functions, categorized into Biological Process (BP), Molecular Function (MF), and Cellular Component (CC).
  • GOA (Gene Ontology Annotation): Links GO terms to proteins.
  • STRING: A database of known and predicted protein-protein interactions.
  • InterPro: Integrates protein families, domains, and functional sites.
  • RCSB PDB: A repository for experimentally determined 3D protein structures.
  • AlphaFold Database: Provides high-accuracy predicted protein structures.

For performance evaluation, the authors used a benchmark dataset that strictly followed the standards of the CAFA challenge, utilizing manually curated and reviewed proteins from UniProtKB/Swiss-Prot with experimental evidence codes.

Key Insights and Ideas

  • Hybridization is Key: The most significant insight is that integrating multiple data sources (sequence, structure, PPI, etc.) consistently leads to improved protein function prediction accuracy, outperforming methods relying on a single data type.
  • Structure Matters: 3D structure information, especially when accurately predicted by tools like AlphaFold, significantly enhances prediction accuracy, particularly for Molecular Function and Cellular Component aspects.
  • PPI Networks Inform Biological Processes: PPI network information is particularly valuable for predicting Biological Processes, likely due to its role in understanding cellular signaling and regulation.
  • Evolution of Methods: There's a clear trend from simpler sequence-based methods to complex deep learning architectures that can integrate diverse and rich biological data.
  • Benchmark Challenges: The lack of standardized benchmark datasets and the reliance on different evaluation criteria across studies make direct comparison of methods challenging.

Key Results

The paper presents a comprehensive performance comparison of the four categories of methods.

  • Overall Performance: Hybrid information-based methods generally outperform the other three categories across various metrics (Fmax and AUPR) for MFO, BPO, and CCO.
  • Specific Method Performance:
    • PFresGO achieved the best performance for BPO (Fmax of 0.527).
  • HNetGO showed top performance for MFO (Fmax of 0.523) and CCO (Fmax of 0.618), and also led in AUPR for all three categories (MFO: 0.570, BPO: 0.512, CCO: 0.669).
  • Category Contributions: Sequence and structure information contribute to MFO and CCO prediction, while PPI information is more impactful for BPO.
  • Structure Prediction Accuracy: AlphaFold demonstrated superior contact accuracy compared to other methods like RaptorX, ProSPr, and trRosetta on the CASP13 test set.
  • Unannotated Proteins: Current best-performing computational methods show moderate accuracy (max ~37%) in predicting functions for unannotated proteins, indicating room for improvement.

Limitations and Future Work

  • Data Limitations: The paper highlights several limitations in current datasets:
  • A vast majority of proteins in UniProt are not yet manually reviewed.
  • Experimental structure data is limited.
  • PPI data can be noisy and contain false positives.
  • Lack of standardized benchmark datasets hinders direct comparison.
  • Feature Extraction: Developing new, more comprehensive, and accurate feature extraction methods is crucial.
  • Algorithms: The paper suggests that while deep learning has advanced the field, further research is needed to effectively utilize and combine these models. Transfer learning is proposed as a solution for predicting functions in species with sparse data.
  • Missing Data: Fusing information from different data sources can be challenging when data is missing for certain proteins.
  • Computational Resources: Integrating multiple data sources often requires more sophisticated algorithms and can lead to increased computational costs.

Future work should focus on creating better benchmark datasets, developing novel feature extraction techniques, exploring advanced deep learning architectures, and addressing the challenge of missing data in hybrid approaches.

Summary updated

BibTeX
@article{lin2024comprehensive,
  author = {Baohui Lin and Xiaoling Luo and Yumeng Liu and Xiaopeng Jin},
  title = {{A comprehensive review and comparison of existing computational methods for protein function prediction}},
  journal = {Briefings in Bioinformatics},
  year = {2024},
  doi = {10.1093/bib/bbae289},
}
Figure