---
library_name: pytorch
license: apache-2.0
tags:
- protein-structure-prediction
- bioinformatics
- pytorch-lightning
- deep-learning
datasets:
- protein-secondary-structure
metrics:
- f1
---

# Protein Secondary Structure Prediction Models

BiRNN and BiLSTM models for predicting Q8 (8-state) and Q3 (3-state) protein secondary structures.

## Model Performance

| Model | Val F1 Q8 | Val F1 Q3 | Harmonic F1 |
|-------|-----------|-----------|-------------|
| BiRNN | 0.547 | 0.700 | **0.6222** |
| BiLSTM | 0.570 | 0.718 | **0.6352** |

## Model Architecture

### BiRNN
- Embedding: 128-dim
- Bidirectional RNN: 2 layers, hidden_dim=256
- Dropout: 0.3
- Parameters: 601K

### BiLSTM  
- Embedding: 128-dim
- Bidirectional LSTM: 2 layers, hidden_dim=256
- Dropout: 0.3
- Parameters: 2.4M

## Files
- `birnn_best.ckpt`: BiRNN checkpoint
- `bilstm_best.ckpt`: BiLSTM checkpoint

## Usage
import torch
import pytorch_lightning as pl
from huggingface_hub import hf_hub_download

Download checkpoint
checkpoint_path = hf_hub_download(
repo_id="yogesh-2003/protein-structure-nppe2",
filename="bilstm_best.ckpt"
)

Load model (requires ProteinStructurePredictor and BiLSTM classes)
model = ProteinStructurePredictor.load_from_checkpoint(
checkpoint_path,
model=BiLSTM()
)