# GlycoBERT: Transformer-based Glycan Structure Prediction from Mass Spectrometry
## Model Description
GlycoBERT is a BERT-based transformer model designed for predicting glycan structures from tandem mass spectrometry (MS/MS) data. This model treats mass spectra as natural language sequences, called MS sentences, enabling accurate classification of MS/MS spectra into glycan structure classes.
## Key Features
- **State-of-the-art Performance**: Achieves 95.1% structural accuracy
- **Novel Tokenization**: Converts MS/MS spectra into "MS sentences" encoding experimental metadata, retention time, and m/z peaks
- **Contextual Analysis**: Uses transformer self-attention to capture complex relationships between spectral peaks
- **Robust Classification**: Consistent performance across diverse experimental conditions and glycan types
## Model Architecture
- **Base Model**: BERT with 12 transformer layers
- **Parameters**: 96 million parameters
- **Attention Heads**: 12 per layer
- **Hidd…
