AlphaFold2 / AlphaFold-Multimer Validation

Plain-language role: Use AlphaFold when you want a reference-grade structure prediction check for a designed sequence or complex.

Prerequisites

Requirement	Minimum	Recommended
Python	3.8+	3.10
CUDA	11.0+	12.0+
GPU VRAM	32GB	40GB (A100)
RAM	32GB	64GB
Disk	100GB	500GB (for databases)

How to run

First time? See Installation Guide to set up Modal and biomodals.

Option 1: ColabFold (recommended for multimer)

cd biomodals
modal run modal_colabfold.py \
  --input-faa sequences.fasta \
  --out-dir output/

GPU: A100 (40GB) | Timeout: 3600s default

Option 2: Local installation

git clone https://github.com/deepmind/alphafold2-multimer.git
cd alphafold2-multimer

python run_alphafold.py \
  --fasta_paths=query.fasta \
  --output_dir=output/ \
  --model_preset=monomer \
  --max_template_date=2026-01-01

Option 3: ESMFold (fast single-chain)

modal run modal_esmfold.py \
  --sequence "MKTAYIAKQRQISFVK..."

Key parameters

Parameter	Default	Options	Description
--model_preset	monomer	monomer/multimer	Model type
--num_recycle	3	1-20	Recycling iterations
--max_template_date	-	YYYY-MM-DD	Template cutoff
--use_templates	True	True/False	Use template search

Output format

output/
├── ranked_0.pdb           # Best model
├── ranked_1.pdb           # Second best
├── ranking_debug.json     # Confidence scores
├── result_model_1.pkl     # Full results
├── msas/                  # MSA files
└── features.pkl           # Input features

Extracting metrics

import pickle

with open('result_model_1.pkl', 'rb') as f:
    result = pickle.load(f)

plddt = result['plddt']
ptm = result['ptm']
iptm = result.get('iptm', None)  # Multimer only
pae = result['predicted_aligned_error']

Sample output

Successful run

$ python run_alphafold.py --fasta_paths complex.fasta --model_preset multimer
[INFO] Running MSA search...
[INFO] Running model 1/5...
[INFO] Running model 5/5...
[INFO] Relaxing structures...

Results:
  ranked_0.pdb:
    pLDDT: 87.3 (mean)
    pTM: 0.78
    ipTM: 0.62
    PAE (interface): 8.5

Saved to output/

What good output looks like:

• pLDDT: > 85 (mean, on 0-100 scale) or > 0.85 (normalized)
• pTM: > 0.70
• ipTM: > 0.50 for complexes
• PAE_interface: < 10

Decision tree

Should I use AlphaFold?
│
├─ What are you predicting?
│  ├─ Single protein → ESMFold (faster)
│  ├─ Protein-protein complex → AlphaFold/ColabFold ✓
│  ├─ Protein + ligand → Chai or Boltz
│  └─ Batch of sequences → ColabFold ✓
│
├─ What do you need?
│  ├─ Highest accuracy → AlphaFold/ColabFold ✓
│  ├─ Fast screening → ESMFold
│  └─ MSA-free prediction → Chai or ESMFold
│
└─ Which AF2 option?
   ├─ Local installation → Full control, slow setup
   ├─ ColabFold → Easier, MSA server
   └─ Modal → Recommended for batch

Typical performance

Campaign Size	Time (A100)	Cost (Modal)	Notes
100 complexes	1-2h	~$8	With MSA server
500 complexes	5-10h	~$40	Standard campaign
1000 complexes	10-20h	~$80	Large campaign

Per-complex: ~30-60s with MSA server.

---

Verify

find output -name "ranked_0.pdb" | wc -l  # Should match input count

---

Troubleshooting

Low pLDDT regions: May indicate disorder or poor design Low ipTM: Interface not confident, check hotspots High PAE off-diagonal: Chains may not interact OOM errors: Use ColabFold with MSA server instead

Error interpretation

Error	Cause	Fix
RuntimeError: CUDA out of memory	Sequence too long	Use A100 or split prediction
KeyError: 'iptm'	Running monomer on complex	Use multimer preset
FileNotFoundError: database	Missing MSA databases	Use ColabFold MSA server
TimeoutError	MSA search slow	Reduce num_recycles

---

Next: protein-design-qc for filtering and ranking.

Inputs

• One or more protein sequences in FASTA format, optionally grouped as a complex.
• Optional template structures, MSA settings, and recycle count overrides.
• A prediction workspace with enough disk for intermediate features and outputs.

Outputs

• Predicted structure files such as PDB/mmCIF plus per-model confidence JSON or PKL files.
• Model-level confidence metrics including pLDDT, pTM, ipTM, and PAE matrices.
• A ranked prediction set ready for protein-design-qc filtering or ipsae ranking.

Next Step

Run protein-design-qc to filter low-confidence models, then use ipsae when ranking binders for experiments.