Explainable deep graph learning accurately modeling the peptide secondary structure prediction

Yi Jiang, Ruheng Wang, Jiuxin Feng, Junru Jin, Sirui Liang, Zhongshen Li, Yingying Yu, Anjun Ma, Ran Su, Quan Zou, Qin Ma, Leyi WeibioRxiv 2022 · DOI

TL;DR

This paper introduces PHAT, a novel deep graph learning framework for predicting peptide secondary structures, achieving state-of-the-art accuracy, especially for short peptides. PHAT stands out by incorporating explainability through interpretable hypergraph attention mechanisms and demonstrating its utility in downstream tasks like predicting peptide toxicity and aiding 3D structure reconstruction.

What Problem Does This Paper Address?

Accurately predicting the secondary structure of peptides is crucial for understanding their function and designing new therapeutic molecules. However, this task is challenging, particularly for short peptides, due to a lack of discriminative information. Existing methods often struggle with short sequences, rely heavily on manual feature engineering, and operate as "black boxes" lacking interpretability. Furthermore, protein-specific methods are not directly applicable to peptides due to inherent structural differences. This paper aims to bridge this gap by developing a peptide-specific, accurate, and interpretable deep learning model.

Methods

PHAT is a deep graph learning framework built upon three main modules:

  • Knowledge Transfer Module: Leverages a pre-trained protein language model (ProtT5) to transfer semantic knowledge from large protein datasets to peptides, capturing high-latent and long-term features.
  • Hypergraph Embedding Module (HyperGMA): This novel component uses a hypergraph structure to represent peptide sequences. Peptide fragments are treated as hyperedges, and residue groups as hypernodes. A multi-head attention mechanism within HyperGMA captures multi-scale relationships, integrating local and global information and preventing over-smoothing common in standard graph networks.
  • Fusion and Classification Module: Combines features from the previous two modules using element-wise multiplication. A Bidirectional Long Short-Term Memory (Bi-LSTM) network further integrates these features, followed by a Conditional Random Field (CRF) layer to capture correlations between secondary structures. Finally, a Viterbi algorithm is used for prediction.

A key innovation is the interpretability aspect, achieved by visualizing transition and emission matrices from the CRF layer, offering insights into the model's reasoning.

Datasets

The study utilized benchmark datasets commonly used in the field:

  • Three-state secondary structure dataset: Consists of 5,772 protein and peptide sequences derived from X-ray crystallography with a resolution of at least 2.5 Å, no chain breaks, and less than five unknown amino acids. Sequence similarity was reduced to 25%. After filtering, a dataset of 1,285 peptide sequences (length 30-99) was used for three-state prediction.
  • Eight-state secondary structure dataset: A dataset of 1,060 peptide sequences derived from the DSSP database was used for eight-state prediction.

The datasets were divided into training, validation, and testing sets, with specific considerations for sequences shorter than or equal to 50 residues.

Key Insights and Ideas

  • Hypergraph for Peptide Structure: The use of a hypergraph structure, where fragments are hyperedges and residue groups are hypernodes, effectively captures multi-scale relationships in peptide sequences.
  • Preventing Over-smoothing: The hypergraph construction and attention mechanism in HyperGMA are designed to mitigate the over-smoothing issue often encountered in traditional graph neural networks.
  • Transfer Learning from Proteins: Leveraging pre-trained protein language models (ProtT5) is highly effective for peptide analysis, transferring valuable semantic knowledge despite structural differences.
  • Interpretability through Visualization: Visualizing the transition and emission matrices of the CRF layer provides a window into the model's decision-making process, addressing the "black box" problem of deep learning.
  • Complementarity of Features: Fusion strategies, particularly element-wise multiplication, effectively combine information from different feature extractors (ProtT5 and HyperGMA), highlighting the complementary nature of sequential and structural information.

Key Results

  • State-of-the-Art Performance: PHAT achieved superior performance across various metrics (Acc, SOV, F1-score) compared to existing methods, including protein-specific predictors and peptide-specific predictors like PSSP-MVIRT, PROTEUS2, and RaptorX.
  • Excellent for Short Peptides: PHAT demonstrated particularly strong performance on shorter peptide sequences, outperforming other methods in various length intervals.
  • Downstream Task Improvement: Incorporating PHAT's predicted secondary structures significantly improved performance in downstream tasks such as peptide toxicity prediction, T-cell receptor interactions with MHC-peptide complexes, and protein-peptide binding site prediction.
  • 3D Structure Reconstruction Aid: PHAT's predicted contact and distance maps showed high accuracy, aiding in more realistic reconstruction of peptide 3D structures.
  • Online Server: An accessible online server for PHAT was developed to facilitate its use by researchers.

Limitations and Future Work

  • Performance on Longer Peptides: While PHAT excels with shorter peptides (<50 residues), its performance on sequences longer than 50 residues is not guaranteed to be as strong.
  • Interaction Context: The current model predicts secondary structures based solely on sequence patterns and does not account for how these structures might change due to interactions with other molecular targets (e.g., proteins, DNA, RNA).

Future work will focus on:

  • Extending PHAT's performance to longer peptide sequences.
  • Incorporating interaction information with other targets to better predict peptide secondary structures in diverse molecular contexts.

Summary updated

Abstract

Abstract Accurately predicting peptide secondary structures remains a challenging task due to the lack of discriminative information in short peptides. In this study, we propose PHAT, a deep graph learning framework for the prediction of peptide secondary structures. The framework includes a novel interpretable deep hypergraph multi-head attention network that uses residue-based reasoning for structure prediction. Our algorithm can incorporate sequential semantic information from large-scale biological corpus and structural semantic information from multi-scale structural segmentation, leading to better accuracy and interpretability even with extremely short peptides. Our interpretable models are able to highlight the reasoning of structural feature representations and the classification of secondary substructures. We further demonstrate the importance of secondary structures in peptide tertiary structure reconstruction and downstream functional analysis, highlighting the versatility of our models. To facilitate the use of our model, we establish an online server which is accessible via http://inner.wei-group.net/PHAT/ . We expect our work to assist in the design of functional peptides and contribute to the advancement of structural biology research.
BibTeX
@misc{jiang2022explainable,
  author = {Yi Jiang and Ruheng Wang and Jiuxin Feng and Junru Jin and Sirui Liang and Zhongshen Li and Yingying Yu and Anjun Ma and Ran Su and Quan Zou and Qin Ma and Leyi Wei},
  title = {{Explainable deep graph learning accurately modeling the peptide secondary structure prediction}},
  year = {2022},
  doi = {10.1101/2022.06.09.495580},
  howpublished = {bioRxiv},
}
Figure