ResNetKhib: a novel cell type-specific tool for predicting lysine 2-hydroxyisobutylation sites via transfer learning

Xiaoti Jia, Pei Zhao, Fuyi Li, Zhaohui Qin, Haoran Ren, Junzhou Li, Chunbo Miao, Quanzhi Zhao, Tatsuya Akutsu, Gensheng Dou, Zhen Chen, Jiangning Song

TL;DR

This paper introduces ResNetKhib, a novel deep learning tool that accurately predicts lysine 2-hydroxyisobutylation (Khib) sites, specifically tailored to different cell types. By employing transfer learning and a ResNet-inspired architecture, ResNetKhib significantly outperforms existing methods, offering a valuable resource for understanding this crucial protein modification.

What Problem Does This Paper Address?

Lysine 2-hydroxyisobutylation (Khib) is a vital post-translational modification (PTM) involved in numerous biological processes. Identifying the specific lysine residues (Khib sites) that undergo this modification is crucial for deciphering its molecular mechanisms. While experimental methods like mass spectrometry can identify these sites, they are often time-consuming and expensive. Computational tools have been developed to predict Khib sites, but a significant gap exists: these tools generally treat all Khib sites equally, failing to account for the observation that Khib site characteristics can vary significantly between different cell types within the same species or across different species. This lack of cell type-specificity limits their predictive power and hinders a nuanced understanding of Khib's diverse roles. Furthermore, there was a specific need for a predictor for mouse Khib sites.

Methods

The researchers developed a deep learning model called ResNetKhib. Its core innovation lies in combining:

  • ResNet Architecture: Inspired by the residual connections in ResNet, which help train very deep neural networks by allowing gradients to bypass layers, this architecture facilitates learning complex patterns.
  • 1D Convolutional Neural Networks (CNNs): These are used to process the protein sequence data, effectively capturing local sequence motifs around potential Khib sites.
  • Transfer Learning: This technique is crucial for addressing data scarcity in specific cell types. A model pre-trained on a larger, more general dataset (e.g., a merged human dataset) is fine-tuned for a specific cell type. This allows the model to leverage learned features and achieve better performance even with limited cell-type-specific data.

ResNetKhib was implemented in two versions: one using word embedding for amino acid representation (ResNetKhibwe) and another using one-hot encoding (ResNetKhibone-hot). The word embedding version, which captures richer amino acid properties, ultimately demonstrated superior performance.

Datasets

The study curated 10 datasets from three species: Homo sapiens (human), Mus musculus (mouse), and Oryza sativa (rice). These datasets were categorized into:

  • Eight cell type-specific datasets: Including human lung (A549), uterus (HeLa), oral cavity (SCC), and kidney (HEK293) cells; mouse liver (C57) cells; and rice leaves, seeds, and flowers.
  • Two general datasets: Merged datasets for human and rice.

These datasets were split into training (for 10-fold cross-validation) and independent test sets. The data preprocessing involved sequence redundancy reduction using CD-HIT.

Key Insights and Ideas

  • Cell Type Specificity is Crucial: The analysis of motif conservation clearly demonstrated that the amino acid preferences surrounding Khib sites differ significantly across various cell types, even within the same species. This strongly supports the need for cell type-specific prediction tools.
  • Transfer Learning Boosts Performance: For cell types with limited experimental data, transfer learning proved highly effective, significantly improving prediction accuracy and accelerating model training. This is particularly important for rare or less-studied cell types.
  • Deep Learning Outperforms Traditional Methods: The ResNet-based models consistently outperformed traditional machine learning methods like Random Forest (RF), highlighting the power of deep learning for capturing complex sequence-based patterns in biological data.
  • ResNet Architecture is Well-Suited: The residual connections in the ResNet architecture appear to be particularly effective for learning the intricate features associated with Khib sites.

Key Results

  • Superior Predictive Performance: ResNetKhib achieved high Area Under the Receiver Operating Characteristic (AUROC) values, ranging from 0.807 to 0.901 across different cell types and species. This significantly outperformed existing methods, including the state-of-the-art DeepKhib predictor.
  • Benchmarking Success: On independent test datasets, ResNetKhib consistently achieved better performance metrics (Accuracy, Sensitivity, Specificity, MCC, F1-score) compared to RF-based models and DeepKhib. For instance, on the human_K dataset, ResNetKhib achieved an AUROC of 0.855 compared to DeepKhib's 0.827.
  • Web Server Availability: A user-friendly web server has been developed and made publicly accessible, allowing researchers to easily predict Khib sites for their protein sequences.

Limitations and Future Work

  • Data Availability: While the study curated extensive datasets, the performance of cell type-specific models is inherently dependent on the availability of high-quality experimental data for each cell type.
  • Generalizability: While the models show promise, further validation on a wider range of cell types and species would strengthen their generalizability.
  • Experimental Validation: The paper suggests that the predicted Khib sites can serve as candidates for hypothesis-driven experimental validation, implying that further experimental work is needed to confirm these predictions.

The authors envision ResNetKhib as a valuable tool for the research community, facilitating the identification of Khib sites and advancing the understanding of Khib's biological roles.

Summary updated

BibTeX
@article{jia2023resnetkhib,
  author = {Xiaoti Jia and Pei Zhao and Fuyi Li and Zhaohui Qin and Haoran Ren and Junzhou Li and Chunbo Miao and Quanzhi Zhao and Tatsuya Akutsu and Gensheng Dou and Zhen Chen and Jiangning Song},
  title = {{ResNetKhib: a novel cell type-specific tool for predicting lysine 2-hydroxyisobutylation sites via transfer learning}},
  journal = {Briefings in Bioinformatics},
  year = {2023},
  doi = {10.1093/bib/bbad063},
}
Figure