glycotrans 0.1.0

A python package for glycan structure prediction from mass spectrometry data

glycoTrans

A repository for transformer-based models for glycan structure prediction from mass spectrometry data. Presently, the repository includes two models: GlycoBERT and GlycoBART.

(Ejas Althaf Abtheen, Arun Singh, Shyam Sriram, Changyou Chen, Sriram Neelamegham, Rudiyanto Gunawan bioRxiv 2025.07.02.662857; doi: https://doi.org/10.1101/2025.07.02.662857)

Overview

glycoTrans introduces state-of-the-art transformer architectures to glycomics, enhancing how we predict glycan structures from tandem mass spectrometry (MS/MS) data. By treating mass spectra and glycan structures as sequences, called MS and Glycan sentence, respectively, the models are trained to capture complex contextual relationships in spectral data.

Key Models

• GlycoBERT: A BERT-based sequence classifier for high-accuracy glycan structure prediction
• GlycoBART: A BART-based generative model capable of de novo glycan structure inference

Installation

From PyPI

pip install glycotrans

From GitHub

pip install git+https://github.com/CABSEL/glycotrans.git

Features

Superior Performance

• 95.1% accuracy on test datasets, outperforming state-of-the-art CNN-based methods like CandyCrunch
• Robust performance across diverse MS analysis parameters and glycan types

De Novo Discovery

• GlycoBART's generative capability enables prediction of novel glycan structures not present in training data
• Overcome database-dependent limitations

Transformer Architecture

• Self-attention mechanisms capture long-range dependencies in spectral data
• Bidirectional processing using BERT and BART for comprehensive context understanding
• Custom tokenization for MS spectra and glycan structures

Technical Innovation

MS Sentence Representation

Our novel approach converts MS/MS spectra into "MS sentences" containing:

• Experimental metadata (LC type, ion mode, fragmentation method, etc.)
• Normalized retention time
• Precursor m/z and fragment ion information
• Peak intensity encoding through positional embeddings

Glycan Sentence Format

Glycan structures are represented as sequences of constituent antennae:

• Terminal-to-core monosaccharide ordering
• Linkage information preservation

Model Architecture

GlycoBERT

• Base: BERT encoder with 12 transformer layers
• Parameters: 96 million trainable parameters
• Task: Multi-class classification (3,590 glycan classes)
• Attention: 12 attention heads per layer
• Embedding: 768-dimensional representations

GlycoBART

• Base: BART encoder-decoder architecture
• Parameters: 207 million trainable parameters
• Task: Conditional sequence generation
• Architecture: 12-layer encoder + 12-layer decoder
• Attention: 16 attention heads per layer
• Generation: Beam search with 32 beams

Performance Metrics

Accuracy Levels

1. Mass Accuracy: Monoisotopic mass matching
2. Composition Accuracy: Monosaccharide composition matching
3. Topological Accuracy: Branching pattern recognition
4. Structural Accuracy: Complete linkage-specific identification

Benchmark Results

Model	Mass	Composition	Topology	Structure
GlycoBERT	98.8%	98.8%	96.7%	95.1%
GlycoBART (top-1)	93.2%	93.1%	90.4%	89.1%
GlycoBART (top-5)	95.5%	95.5%	93.3%	93.1%
CandyCrunch	94.1%	94.0%	91.2%	90.3%

Example

Example_Inference.ipynb

The Google colab notebook shows an example inference that can be modified for specific use cases.
The example file used in the notebook can be found at

Data and Models

Training data are available on Zenodo under .

The full version of GlycoBERT and GlycoBART are available on HuggingFace .

Funding

This work was supported by funding from NHLBI (HL103411)