# Run AlphaFold2 on Instance

## Prerequisites

* Instance(GPU recommended). Full AlphaFold databases require ~2+ TB disk; reduced DBs use much less.
* Conda installed.
* Read/write access to chosen data and output directories (for example /mnt/model/vdb).

## 1. System packages (Ubuntu)

Update and install basic build tools and utilities:

```bash
sudo apt update -y && sudo apt upgrade -y
sudo apt install -y build-essential cmake git hmmer kalign tzdata wget libstdc++6
```

## 2. Create and activate Conda environment

Create a Python 3.8 environment and activate it:

```bash
conda create -n alphafold python=3.8 -y
conda activate alphafold
```

## 3. Install Python dependencies

Install key packages; follow AlphaFold README for full dependency list:

```bash
pip install jaxlib==0.4.26
conda install -c conda-forge pdbfixer libstdcxx-ng -y
```

(Install any additional pip packages listed in the AlphaFold repo README.)

## 4. Helper file required by AlphaFold

AlphaFold expects a helper text file under alphafold/alphafold/common/:

```bash
mkdir -p alphafold/alphafold/common/
wget -q -P alphafold/alphafold/common/ https://git.scicore.unibas.ch/schwede/openstructure/-/raw/7102c63615b64735c4941278d92b554ec94415f8/modules/mol/alg/src/stereo_chemical_props.txt
```

## 5. Clone AlphaFold repo

```bash
git clone https://github.com/google-deepmind/alphafold.git
```

## 6. cd alphafold

Follow the repository README for any extra setup steps or environment pins.

## 7. Prepare databases

Place required databases on your instance and note the root data directory (example: /mnt/model/vdb/database). Essential subpaths include pdb70, uniref90, mgnify, bfd (or small_bfd for reduced), uniref30, pdb_mmcif. If disk or time is constrained, use the reduced DB preset or precomputed MSAs.

## 8. Run AlphaFold (full DBs)

Update paths (FASTA, output, data) and run the full DB preset as a single-line command:

```bash
python3 run_alphafold.py --use_gpu_relax=True \
  --fasta_paths=/mnt/model/vdb/input/T1049.fasta \
  --output_dir=/mnt/model/vdb/output \
  --data_dir=/mnt/model/vdb/database \
  --pdb70_database_path=/mnt/model/vdb/database/pdb70/pdb70 \
  --uniref90_database_path=/mnt/model/vdb/database/uniref90/uniref90.fasta \
  --mgnify_database_path=/mnt/model/vdb/database/mgnify/mgy_clusters_2022_05.fa \
  --bfd_database_path=/mnt/model/vdb/database/bfd/bfd_metaclust_clu_complete_id30_c90_final_seq.sorted_opt \
  --uniref30_database_path=/mnt/model/vdb/database/uniref30/UniRef30_2021_03 \
  --template_mmcif_dir=/mnt/model/vdb/database/pdb_mmcif/mmcif_files \
  --obsolete_pdbs_path=/mnt/model/vdb/alphafold/pdbs/pdb.json \
  --max_template_date=2025-03-04 \
  --db_preset=full_dbs \
  --model_preset=monomer \
  --use_precomputed_msas=False
```

## 9. Run AlphaFold (reduced DB option)

If disk/runtime is constrained, use reduced DBs and the small BFD; single-line command:

```bash
python3 run_alphafold.py --use_gpu_relax=True \
  --fasta_paths=/mnt/model/vdb/input/T1049.fasta \
  --output_dir=/mnt/model/vdb/output \
  --data_dir=/mnt/model/vdb/database \
  --template_mmcif_dir=/mnt/model/vdb/database/pdb_mmcif/mmcif_files \
  --obsolete_pdbs_path=/mnt/model/vdb/alphafold/pdbs/pdb.json \
  --max_template_date=2025-03-04 \
  --db_preset=reduced_dbs \
  --model_preset=monomer \
  --use_precomputed_msas=False \
  --small_bfd_database_path=/mnt/model/vdb/database/small_bfd/bfd-first_non_consensus_sequences.fasta
```

## 10. Parameter highlights

* `--fasta_paths`: input FASTA (one or more).
* `--data_dir`: root directory containing databases.
* `--output_dir`: where predictions and metrics are written.
* `--max_template_date`: restrict templates by date.
* `--db_preset`: full_dbs or reduced_dbs.
* `--model_preset`: monomer (or multimer if configured and set up).
* `--use_precomputed_msas`: set True if you have precomputed MSAs.

## 11. Viewing predictions

After completion, PDB files and metrics are in the output directory. Visualize PDBs with NCBI iCn3D:

Open [https://www.ncbi.nlm.nih.gov/Structure/icn3d/full.html](https://www.ncbi.nlm.nih.gov/Structure/icn3d/full.html) → File → Open Files → select PDB(s).

## 12. Common issues & troubleshooting

* **CUDA OOM**: switch to monomer, reduce DB size, use a instance with more GPU memory, or enable swap if appropriate for your environment.
* **Missing files / path errors**: verify FASTA and all DB paths are correct and accessible.
* **Permission errors**: ensure the runtime user can read/write the data and output directories (use chown/chmod as needed).
* **Long runtime / disk constraints**: use reduced_dbs or precomputed MSAs to speed runs.

## 13. Example outputs

AlphaFold writes predictions and auxiliary outputs (ranked models, metrics) into the output directory.

Example header text:

```plaintext
prediction_1 | score=... | model_preset=monomer
```

## 14. Resources and links

* AlphaFold official repo: [https://github.com/google-deepmind/alphafold](https://github.com/google-deepmind/alphafold)


---