Abstract
Post-translational modifications (PTMs) either enhance a protein’s activity in various sub-cellular processes, or degrade their activity which leads towards failure of intracellular processes. Tyrosine nitration (NT) modification degrades protein’s activity that initiate and propagate various diseases including Neurodegenerative, Cardiovascular, Autoimmune diseases, and Carcinogenesis. Identification of NT modification support development of novel therapies and drug discoveries for associated diseases. Identification of NT modification in biochemical labs is expensive, time consuming, and error-prone. To supplement this process, several computational approaches have been proposed. However these approaches remain fail to precisely identify NT modification, due to the extraction of irrelevant, redundant and less discriminative features from protein sequences. The paper in hand presents NTpred framework competent in extracting comprehensive features from raw protein sequences using four different sequence encoders. To reap the benefits of different encoders, it generates four additional feature spaces by fusing different combinations of individual encodings. Furthermore, it eradicates irrelevant and redundant features from eight different feature spaces through a Recursive Feature Elimination process. Selected features of four individual encodings and four feature fusion vectors are used to train eight different Gradient Boosted Tree classifiers. The probability scores from the trained classifiers are utilized to generate a new probabilistic feature space, that is utilized to train a Logistic Regression classifier. On BD1 benchmark dataset, the proposed framework outperform existing best performing predictor in 5-fold cross validation and independent test evaluation with combined improvement of 13.7% in MCC and 20.1% in AUC. Similarly, on BD2 benchmark dataset, the proposed framework outperform existing best performing predictor with combined improvement of 5.3% in MCC and 1.0% in AUC.
1. Introduction
Proteins are essential cellular components whose functions are often regulated by Post-Translational Modifications (PTMs). While many PTMs are beneficial, Tyrosine nitration (NT) is a deleterious modification characterized by the attachment of a nitro group ($NO_2$) to the Tyrosine residue. This process, often triggered by oxidative stress or inflammation, acts as a form of biological "rust," leading to protein degradation and structural damage. NT is a significant biomarker for several severe conditions, including Alzheimer’s disease, Parkinson’s disease, and various cancers.
Traditional laboratory identification of NT sites is resource-intensive and prone to errors. Consequently, computational "in-silico" methods have become vital. However, existing predictors often suffer from high false-positive rates due to their inability to capture the complex sequence patterns surrounding nitration sites.
2. Methodology: The NTpred Framework
NTpred addresses these challenges through a multi-perspective feature engineering and ensemble learning approach. The framework operates in four distinct stages:
2.1 Multi-Perspective Encoding
To capture diverse biological signals, NTpred utilizes four sequence encoders:
- Kgap & CKSAAP: These capture the frequency of amino acid pairs separated by varying gaps.
- ASDC: This method normalizes pair frequencies to account for variations in protein length.
- DistancePair: This encoder analyzes both individual amino acids and their spatial pairings.
2.2 Feature Fusion and Selection
To maximize the information extracted, the framework generates four additional "fused" feature spaces by combining different encoder outputs. To ensure the model remains efficient and avoids overfitting, Recursive Feature Elimination (RFE) is applied to prune redundant or noisy data, retaining only the most discriminative features.
2.3 Stacked Ensemble Architecture
The core of NTpred is a two-stage classification pipeline. In the first stage, eight Gradient Boosted Tree (GBT) classifiers are trained independently on each feature space. In the second stage, the probability scores from these models are fed into a Multivariate Logistic Regression meta-learner. This stacking strategy allows the model to learn the optimal weighting for each encoder's predictions.
3. Experimental Setup
The framework was validated using two benchmark datasets:
- BD1: A large-scale dataset derived from UniProt, filtered to minimize sequence similarity and prevent memorization.
- BD2: A smaller, more heterogeneous dataset compiled from specialized databases like dbPTM and GPS-YNO2. Both datasets included independent test sets with significant class imbalance, mirroring the real-world scarcity of nitration sites relative to non-modified residues.
4. Results and Discussion
NTpred demonstrated substantial performance gains over existing state-of-the-art (SOTA) predictors. On the BD1 dataset, the framework achieved a 13.7% improvement in the Matthews Correlation Coefficient (MCC) and a 20.1% increase in the Area Under the Curve (AUC). On the more challenging BD2 dataset, it showed improvements of 5.3% in MCC and 1.0% in AUC.
The success of NTpred highlights the importance of combining absolute and relative sequence representations. By leveraging a meta-learner instead of simple averaging, the system effectively mitigates sequence bias and handles the inherent complexity of protein sequence data.
5. Conclusion
NTpred represents a significant advancement in the computational identification of Tyrosine nitration sites. While highly effective, future iterations could benefit from integrating deep-learning-based embeddings from protein language models. Such advancements would allow the framework to capture even more nuanced "grammatical" patterns within amino acid sequences, further refining the precision of PTM site prediction.
Summary updated