How to predict protein structure with AlphaFold2 using ColabFold locally or on Google Colab

Question

I have 50 novel protein sequences from a genome annotation and want to predict their 3D structures. I don't have access to expensive cloud GPU instances. What is the most practical way to run AlphaFold2 predictions — is ColabFold better than the full AlphaFold2 pipeline for most researchers?

Admin · Accepted Answer

**ColabFold is the recommended approach for most researchers.** It uses the same AlphaFold2 model but with faster MSA generation via MMseqs2 instead of JackHMMER/HHblits — typically 10× faster per prediction.

**Option 1: ColabFold on Google Colab (free GPU)**
Open: `https://colab.research.google.com/github/sokrypton/ColabFold/blob/main/AlphaFold2.ipynb`
- Paste sequence, click Run All
- Free T4 GPU (~15 min per protein)
- Limited to ~1500 aa and ~5 free GPU hours/day

**Option 2: ColabFold locally (if you have a GPU)**
```bash
# Install
pip install 'colabfold[standard]'

# Run batch prediction
colabfold_batch sequences.fasta output_dir/ 
  --num-recycle 3 
  --num-models 5 
  --model-type alphafold2_multimer_v3  # for complexes
```

**Option 3: Full AlphaFold2 (for maximum accuracy, requires ~2 TB database)**
```bash
# After installing AlphaFold2 and downloading databases
python run_alphafold.py 
  --fasta_paths=sequence.fasta 
  --max_template_date=2024-01-01 
  --model_preset=monomer 
  --output_dir=./output 
  --data_dir=/databases/alphafold
```

**Interpreting output:**
- `ranked_0.pdb`: top-ranked model (usually best)
- `pLDDT score`: per-residue confidence (>90 = very high, 70–90 = high,