AMS

AMS predicts 88 single-amino-acid post-translational modification (PTM) types in protein sequences using machine learning models trained on integrated experimental modification data.

Key Features:

• Extensive PTM Prediction: Predicts a diverse set of 88 different single-amino-acid PTM types in protein sequences.
• Data Sources and Training: Utilizes experimentally confirmed modification data from UniProt and PhosphoELM to train predictive models.
• Physico-Chemical Feature Representation: Represents each modified residue's sequence vicinity with physico-chemical features encoded as high-quality indices (HQI) derived from the AAindex database via automatic clustering.
• Machine Learning Models: Employs Multi-Layer Perceptron (MLP) pattern classifiers trained separately for each PTM type.
• Classifier Optimization and Metrics: Classifiers are optimized to improve recall, precision, and area under the ROC curve (AUC).
• Brainstorming Consensus Meta-Learning: Integrates multiple machine learning instances and data representations using a "brainstorming" consensus approach to enhance prediction accuracy.
• Performance Enhancement vs Other Methods: Achieves an average AUC improvement of approximately 7% over single-method approaches such as artificial neural networks (ANNs) or support vector machines (SVMs) and up to ~32% improvement on challenging sequence motifs.
• Consensus Meta-Learning Performance: The brainstorming consensus meta-learning methodology raises the reported average AUC score to around 89%.

Scientific Applications:

• Proteomics Research: Enables identification of PTMs across proteomes to support studies of protein function and regulation.
• Disease Mechanism Studies: Supports investigation of PTM involvement in disease pathways to aid identification of potential biomarkers or therapeutic targets.
• Functional Annotation of Proteins: Assists annotation of proteins with functional information based on predicted post-translational modifications.

Methodology:

Protein sequences are dissected into overlapping short segments, each segment is encoded in a high-dimensional space using physico-chemical features, and models are trained and validated using leave-one-out cross-validation (LOOCV).