| Property | Value |
|---|---|
| Author | jglaser |
| Research Paper | bioRxiv preprint |
| Primary Task | Sentence Similarity / Binding Affinity Prediction |
| Framework | Sentence-Transformers |
What is protein-ligand-mlp-1?
protein-ligand-mlp-1 is an advanced machine learning model designed to predict binding affinities (pIC50 values) between proteins and chemical compounds. It utilizes a sophisticated sentence-transformer architecture to process both protein sequences and chemical SMILES notation, enabling accurate prediction of molecular interactions.
Implementation Details
The model implements a complex neural architecture combining multiple transformer layers and dense networks. It processes protein sequences with a 2048-length capable BERT model and ligand SMILES with a 512-length transformer, followed by multiple dense layers for prediction refinement.
- Protein encoding through a 1024-dimensional BERT transformer with custom pooling
- Ligand processing via 768-dimensional transformer with specialized tokenization
- Multiple GELU-activated dense layers for feature processing
- Ensemble capability for uncertainty estimation
Core Capabilities
- Protein sequence processing up to 2048 tokens
- SMILES notation handling up to 512 tokens
- Binding affinity prediction in pIC50 units
- Uncertainty quantification through ensemble predictions
- Feature extraction for both protein and ligand inputs
Frequently Asked Questions
Q: What makes this model unique?
The model uniquely combines protein and ligand processing in a single architecture, using specialized transformers for each input type and enabling end-to-end binding affinity prediction with uncertainty estimation.
Q: What are the recommended use cases?
This model is ideal for drug discovery applications, protein-ligand interaction studies, and computational chemistry workflows where accurate binding affinity predictions are crucial.