File size: 5,243 Bytes
9a21501
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
---
license: mit
---
---
language: en
tags:
- protein-function-prediction
- bioinformatics
- gene-ontology
- multi-label-classification
- esm-2
- CAFA-6
license: mit
datasets:
- CAFA-6
metrics:
- f1
- precision
- recall
---

# 🧬 CAFA 6 Protein Function Prediction

> *"Once I was zero epochs old, my model said to me... Go make yourself some predictions, don't wait for labeled data."*

**BioBERT, I'm coming for you!** πŸ”₯

## Model Description

State-of-the-art multi-label protein function prediction using ESM-2 embeddings. Predicts Gene Ontology (GO) terms across three ontologies from protein amino acid sequences.

### What This Model Does

Given a protein sequence like:
```
MKTAYIAKQRQISFVKSHFSRQLEERLGLIEVQAPILSRVGDGTQDNLSGAEKAVQVKVKALPDAQFEVVH...
```

It predicts:
- **Molecular Function (MFO)**: What the protein DOES (e.g., "protein binding", "kinase activity")
- **Biological Process (BPO)**: What pathways it's involved in (e.g., "signal transduction")
- **Cellular Component (CCO)**: WHERE it's located (e.g., "nucleus", "membrane")

## Files in This Repository

- `train_esm2_embeddings.pkl` (427 MB) - Pre-computed ESM-2 embeddings for 82,404 training proteins
- `test_esm2_embeddings.pkl` (1.16 GB) - Pre-computed ESM-2 embeddings for test proteins  
- `go_parser.pkl` (25.7 MB) - Gene Ontology hierarchy parser with 40,122 GO terms
- `.gitattributes` - Git LFS configuration for large files

## Dataset Statistics

### Training Data
- **Total proteins**: 82,404
- **Total annotations**: 537,027
- **Unique GO terms**: 26,125

### Selected Terms for Prediction
- **MFO**: 500 most frequent terms
- **BPO**: 800 most frequent terms  
- **CCO**: 400 most frequent terms

### Label Distribution
| Ontology | Proteins with Labels | Avg Labels/Protein | Sparsity |
|----------|---------------------|-------------------|----------|
| MFO      | 49,751 (60.4%)      | 54.2              | 89.2%    |
| BPO      | 44,382 (53.9%)      | 6.6               | 99.2%    |
| CCO      | 58,505 (71.0%)      | 36.5              | 90.9%    |

## Usage

### Requirements

```bash
pip install torch biopython transformers huggingface_hub numpy
```

### Quick Start - Load Embeddings

```python
from huggingface_hub import hf_hub_download
import pickle

# Download embeddings
embeddings_path = hf_hub_download(
    repo_id="nl45/Protein1",
    filename="train_esm2_embeddings.pkl"
)

# Load embeddings
with open(embeddings_path, 'rb') as f:
    embeddings = pickle.load(f)

# embeddings is a dict: {protein_id: embedding_vector}
print(f"Loaded embeddings for {len(embeddings)} proteins")
print(f"Embedding dimension: {list(embeddings.values())[0].shape}")
```

### Generate New Embeddings for Your Protein

```python
from transformers import AutoTokenizer, EsmModel
import torch

# Load ESM-2 model
tokenizer = AutoTokenizer.from_pretrained("facebook/esm2_t33_650M_UR50D")
model = EsmModel.from_pretrained("facebook/esm2_t33_650M_UR50D")

# Your protein sequence
sequence = "MKTAYIAKQRQISFVKSHFSRQLE..."

# Generate embedding
inputs = tokenizer(sequence, return_tensors="pt", padding=True)
with torch.no_grad():
    outputs = model(**inputs)
    embedding = outputs.last_hidden_state.mean(dim=1)  # Shape: [1, 1280]

print(f"Generated embedding shape: {embedding.shape}")
```

### Load GO Parser

```python
# Download GO parser
parser_path = hf_hub_download(
    repo_id="nl45/Protein1",
    filename="go_parser.pkl"
)

# Load parser
with open(parser_path, 'rb') as f:
    go_parser = pickle.load(f)

# Example: Get GO term information
term_info = go_parser.get_term_info("GO:0003674")
print(f"Term: {term_info['name']}")
print(f"Namespace: {term_info['namespace']}")
```

## Model Architecture

The prediction model uses a Multi-Layer Perceptron (MLP):

```
Input: ESM-2 Embeddings (1280-dim)
    ↓
[Dense 2048] β†’ BatchNorm β†’ ReLU β†’ Dropout(0.3)
    ↓
[Dense 1024] β†’ BatchNorm β†’ ReLU β†’ Dropout(0.3)
    ↓
[Dense 512] β†’ BatchNorm β†’ ReLU β†’ Dropout(0.3)
    ↓
[Dense Output] β†’ Sigmoid
    ↓
Multi-label Predictions
```

**Training Details:**
- Loss: Binary Cross-Entropy with Logits
- Optimizer: Adam
- Learning Rate: 0.001 with ReduceLROnPlateau
- Early Stopping: Patience of 10 epochs

## Data Processing Pipeline

1. **Raw Sequences** (FASTA format) β†’ Parse protein IDs and sequences
2. **ESM-2 Encoding** β†’ Generate 1280-dim embeddings using `facebook/esm2_t33_650M_UR50D`
3. **GO Annotations** β†’ Load and normalize GO terms
4. **Label Preparation** β†’ Create multi-label binary matrices with term propagation
5. **Model Training** β†’ Train separate models for MFO, BPO, CCO

## Citation

```bibtex
@misc{nl45_cafa6_2026,
  title={CAFA 6 Protein Function Prediction with ESM-2 Embeddings},
  author={nl45},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/nl45/Protein1}}
}
```

## Acknowledgments

- **CAFA Challenge**: Critical Assessment of Functional Annotation
- **ESM-2**: Evolutionary Scale Modeling from Meta AI
- **Gene Ontology Consortium**: For GO term annotations

## License

MIT License

## Contact

For questions or collaboration: [Create an issue](https://huggingface.co/nl45/Protein1/discussions)

---

**"BioBERT, I'm coming for you!"** πŸ”₯🧬