---
license: mit
language:
- en
library_name: peft
tags:
- ESM-2
- Proteins
- Binding Sites
- QLoRA
- biology
---

# ESM-2 QLoRA for Predicting Binding Sites

## QLoRA Info

```
trainable params: 208322 || all params: 17382365 || trainable%: 1.198467527289871
```

## Testing for Overfitting

```python
Train metrics:
{'eval_loss': 0.09572703391313553,
'eval_accuracy': 0.9670769479865963,
'eval_precision': 0.3970221190232079,
'eval_recall': 0.9411011487595375,
'eval_f1': 0.5584507515735834,
'eval_auc': 0.9543828770020467,
'eval_mcc': 0.5996252550053665}

Test metrics:
{'eval_loss': 0.1680256575345993,
'eval_accuracy': 0.943313091525589,
'eval_precision': 0.2342637814982173,
'eval_recall': 0.7618306193745306,
'eval_f1': 0.35833816875074714,
'eval_auc': 0.8544971814140561,
'eval_mcc': 0.40290081143832884}
```

The metrics on the PDB datasets from [this paper](https://github.com/hamzagamouh/pt-lm-gnn) can be 
[found here](https://huggingface.co/AmelieSchreiber/esmt12_35M_qlora_binding_sites_v0/blob/main/pdb_structure_metrics.txt).