HYALINE predicts whether a G protein-coupled receptor (GPCR) structure is in an active or inactive conformational state using E(n)-equivariant graph neural networks. GPCRs are the largest family of drug targets, with ~35% of FDA-approved drugs acting on these receptors.

The model achieves 0.995 AuROC on cross-validation and 0.819 AuROC on a temporal held-out test set (structures released 2023–2024), outperforming sequence-only baselines by 6–12%. Notably, HYALINE achieves 87.2% accuracy on Class C GPCRs where sequence-based methods fail (39.4%).

For technical details, see the paper.

Installation

pip install git+https://github.com/Varosync/Hyaline.git

Or from source:

git clone https://github.com/Varosync/Hyaline.git
cd Hyaline; pip install -e .

For ESM3 embeddings (required for meaningful predictions):

pip install 'hyaline[esm]'
# or: pip install 'esm>=3.0.0'

Requirements: Python 3.10+, PyTorch 2.0+, PyTorch Geometric

Inference

hyaline predict structure.pdb --device cpu

The model checkpoint will be automatically downloaded on first use and cached at ~/.hyaline/checkpoints/hyaline.pt. You can also provide a checkpoint manually:

# Option 1: Pass directly
hyaline predict structure.pdb --checkpoint /path/to/hyaline.pt

# Option 2: Environment variable
export HYALINE_CHECKPOINT=/path/to/hyaline.pt
hyaline predict structure.pdb

# Option 3: Place in ~/.hyaline/checkpoints/hyaline.pt

Output:

HYALINE PREDICTION
  Score:       0.9521
  Prediction:  Active
  Confidence:  High

Scores >0.5 indicate active state, <0.5 indicate inactive state. Higher absolute values indicate greater confidence.

Batch Prediction

To score an entire directory of PDB files at once:

hyaline predict /path/to/pdb_directory/

This will process every .pdb file in the directory, print a ranked summary table sorted by activation score, and save a CSV file to the input directory:

RESULTS  (ranked by activation score)
======================================================================
  Rank  File                                      Score  State     Confidence
  1     7D7M_clean.pdb                           0.9712  Active    High
  2     6OS0_receptor.pdb                         0.9534  Active    High
  ...
  Active:    187 /  300
  Inactive:  113 /  300
======================================================================
  Results saved to: /path/to/pdb_directory/hyaline_results.csv

Use --output to specify a custom CSV path:

hyaline predict my_structures/ --output results.csv --device cpu

The CSV contains columns: rank, file, score, prediction, confidence — ready for downstream analysis in Python, R, or Excel.

ESM3 Model

HYALINE was trained using ESM3 small open v1 (esm3_sm_open_v1) embeddings (1536-dim). The same model is used at inference time. If ESM3 is not installed, you will see a clear error with installation instructions.

Architecture

HYALINE uses an enhanced E(n)-equivariant graph neural network with:

• ESM3 embeddings (1536-dim) for sequence features
• RBF distance encoding (96-dim) for spatial relationships
• Learned motif attention that autonomously prioritizes conserved activation switches (DRY, NPxxY, CWxP)

Data

The curated dataset contains 1,596 GPCR structures from the Protein Data Bank with activation state annotations from GPCRdb. The complete list of PDB IDs with train/test splits is provided in Supplementary_Data_1.csv.

Citation

If you use HYALINE in your research, please cite:

@article{hyaline2026,
  title   = {HYALINE: Geometric Deep Learning for Accurate Prediction of 
             G Protein-Coupled Receptor Activation States from Structure},
  author  = {Varosync},
  journal = {bioRxiv},
  year    = {2026},
  doi     = {10.64898/2026.01.05.697778}
}

License

MIT License