TL;DR
This paper introduces ISSEC, a novel deep learning approach that uses object detection to accurately infer contacts between protein secondary structure elements (SSEs). By treating contact maps as images and identifying characteristic "rectangle regions," ISSEC overcomes limitations of previous methods and significantly improves the prediction of inter-SSE contacts, which in turn aids in predicting overall protein structure.
What Problem Does This Paper Address?
Understanding how proteins fold into their complex three-dimensional structures is fundamental to biology and medicine. Protein structures are stabilized by interactions between their constituent parts, including local structures like alpha-helices and beta-strands (collectively known as secondary structure elements or SSEs). Predicting the contacts between these SSEs (inter-SSE contacts) provides crucial coarse-grained information about the protein's overall topology, making it a vital step in predicting the full 3D structure.
Previous methods for predicting inter-SSE contacts faced challenges:
- Noise in inter-residue contacts: Directly using predicted contacts between individual amino acids (inter-residue contacts) is often too noisy.
- Dependence on accurate SSE prediction: Some methods first predict the SSEs themselves and then infer contacts. However, errors in predicting SSE boundaries can propagate and hinder contact prediction.
- Limited scope: Many methods are designed for only one type of SSE contact (e.g., helix-helix or strand-strand), failing to handle proteins with mixed SSE types.
This paper aims to address these limitations by developing a more robust and versatile method for inferring inter-SSE contacts.
Methods
ISSEC employs a deep object detection framework, inspired by techniques used for image recognition. The core idea is to treat the predicted inter-residue contact map of a protein as an image. The authors observed that contacting SSEs often form characteristic rectangular patterns within these contact maps: parallel beta-strands form diagonal lines, anti-parallel beta-strands form anti-diagonal lines, and helices form dashed lines.
ISSEC's methodology involves three main steps:
- Feature Extraction: It uses a Feature Pyramid Network (FPN) to extract high-level features from the predicted inter-residue contact map. This helps overcome the noise inherent in raw contact predictions.
- Candidate Region Proposal: A Region Proposal Network (RPN) generates multiple candidate rectangular regions that are likely to contain contacting SSEs. This avoids the need for pre-defined SSE boundaries.
- SSE Identification: Mask R-CNN is then used to analyze these candidate regions. It determines if a region exhibits the characteristic patterns of inter-SSE contacts, classifies the type of contact (α-α, β-β parallel, or β-β anti-parallel), and estimates a confidence score.
A key innovation is ISSEC's ability to handle multiple SSE contact types simultaneously. It also uses a greedy selection strategy to identify non-overlapping regions with high confidence scores, and a multi-task loss function that includes classification, localization, and mask loss for training.
Datasets
The study utilized several datasets for training and evaluation:
- Training Set: Constructed from PDB25, with proteins having over 25% sequence identity to any protein in the test sets excluded, resulting in 9241 proteins.
- Test Sets:
- TEST1000: A newly constructed test set of 1000 proteins randomly selected from PDB25.
- PSICOV118: Consists of 118 proteins from the PSICOV dataset, excluding those without inter-SSE contacts.
- Mem30: 30 transmembrane proteins from MemConP.
- Mem11: 11 transmembrane proteins from a previous study.
- BetaSheet186: 186 proteins known to contain β-β contacts.
Key Insights and Ideas
- Contact Maps as Images: The novel framing of protein contact maps as images for object detection is a significant conceptual leap. This allows leveraging powerful deep learning architectures designed for visual recognition.
- Characteristic Patterns: The identification of specific visual patterns (diagonal lines, dashed lines) within contact maps corresponding to different SSE contact types is a key insight that ISSEC exploits.
- Robustness to Noise: By using deep convolutional features, ISSEC is inherently more robust to the noise present in predicted inter-residue contacts compared to methods that use raw probabilities.
- Eliminating SSE Boundary Dependency: ISSEC's ability to infer contacts without requiring accurate pre-defined SSE boundaries is a major advantage, as SSE prediction itself can be error-prone.
- Unified Prediction: The success in predicting multiple types of inter-SSE contacts simultaneously suggests a more generalizable approach to protein structure analysis.
Key Results
ISSEC demonstrated strong performance across various datasets:
- General Inter-SSE Prediction: ISSEC achieved competitive precision, recall, and F-measure scores on datasets with mixed SSE types (TEST1000 and PSICOV118), outperforming previous approaches in some cases.
- α-α Contact Prediction: On membrane protein datasets (Mem30, Mem11), ISSEC showed significant improvements, particularly in recall and F-measure, outperforming established methods like TMhhep, MemConP, and HHConPred. For instance, on Mem11, ISSEC exceeded HHConPred by 13% in precision, 10% in recall, and 3% in F-measure.
- β-β Contact Prediction: ISSEC performed well on the BetaSheet186 dataset, achieving a higher F-measure (62.1%) than bbcontacts (60.5%), despite bbcontacts having higher precision.
- Improved Inter-Residue Contact Prediction: Integrating ISSEC's predicted inter-SSE contacts into a deep learning model (DeepRN) led to significant improvements in predicting long-range inter-residue contacts, a notoriously challenging task. For example, it improved top L contacts by 4.2% on PSICOV118.
- Improved 3D Structure Prediction: When used to guide protein structure building within the CONFOLD framework (creating CONFOLD+ISSEC), ISSEC significantly improved the quality of predicted 3D structures, as measured by TM-score, for a majority of proteins tested. For protein 101zA, the TM-score improved from 0.44 to 0.55.
Limitations and Future Work
The authors acknowledge that ISSEC currently predicts three common types of inter-SSE contacts: α-α, β-β parallel, and β-β anti-parallel.
Future work directions include:
- Extending ISSEC to predict other types of contacts, such as α-β and β-turn contacts.
- Further refining the method to handle the "dashed line" patterns of α-α contacts more effectively, as the current mask mechanism is less suitable for them.
Summary updated