# AI use cases for Alphafold3 (google-deepmind/alphafold3)

Repo: [google-deepmind/alphafold3](https://github.com/google-deepmind/alphafold3) | Use cases: 3 | Page: https://dedupedai.com/use-cases/google-deepmind-alphafold3/

AlphaFold 3 is Google DeepMind's inference pipeline for predicting the structure of biomolecular complexes. You describe the complex in a JSON input file and run run_alphafold.py, usually inside the provided container, with separate flags for the CPU heavy data pipeline and the GPU inference step. Model parameters are not in the repo: you must request them from Google and follow their terms of use.

This file holds every use case of this repo. Each one starts with a "## " heading and ends with its URL.

- Predict a biomolecular complex from a JSON input: https://dedupedai.com/use-cases/google-deepmind-alphafold3/predict-a-complex-structure-with-alphafold-3/
- Run the CPU search and the GPU fold on different machines: https://dedupedai.com/use-cases/google-deepmind-alphafold3/split-the-alphafold-3-data-pipeline-from-inference/
- Pick the right AlphaFold 3 weights for your project: https://dedupedai.com/use-cases/google-deepmind-alphafold3/choose-an-alphafold-3-weights-variant/

## How to predict a biomolecular complex from a JSON input with Alphafold3

Repo: [google-deepmind/alphafold3](https://github.com/google-deepmind/alphafold3) | Difficulty: Advanced | Time: 4 hours | Basis: From the README

Write a fold_input.json for your sequences and run the container to get a predicted structure.

**Who it is for:** A structural biologist who needs a prediction for a specific complex.

**The problem:** Public servers limit ligands, modifications and throughput.

**The outcome:** A predicted structure written to your output directory.

### Steps

1. Follow the installation documentation in docs/installation.md and obtain the model parameters from Google
2. Write a fold_input.json with name, sequences, modelSeeds, dialect Alphafold3 and version 1, as in the README example
3. Run the docker command in the README with volumes for af_input, af_output, the model parameters and the public databases, calling `python run_alphafold.py` with --json_path, --model_dir and --output_dir
4. Read docs/output.md to interpret the files you get back

### Prompt to paste into your agent

```text
Using google-deepmind/alphafold3, write a fold_input.json for <my sequences>, run the documented docker command with my model and database directories, and explain the output files.
```

### Caveat

The repo contains the inference pipeline only; it does not include the model parameters.

Risk, credentials: Model parameters may only be used if received directly from Google and are subject to their terms of use.

Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/

URL: https://dedupedai.com/use-cases/google-deepmind-alphafold3/predict-a-complex-structure-with-alphafold-3/

## How to run the CPU search and the GPU fold on different machines with Alphafold3

Repo: [google-deepmind/alphafold3](https://github.com/google-deepmind/alphafold3) | Difficulty: Advanced | Time: 3 hours | Basis: From the README

Use the run flags to do genetic and template search on a CPU box and inference on a GPU box.

**Who it is for:** A lab sharing one GPU across many predictions.

**The problem:** The data pipeline is slow and does not need the GPU, so it wastes GPU hours.

**The outcome:** A split workflow where the expensive search runs where it is cheap.

### Steps

1. Install AlphaFold 3 as described in docs/installation.md
2. Run `python run_alphafold.py --help` to list all the available flags
3. Run the data pipeline alone with `--run_data_pipeline` set to true and `--run_inference` set to false on a machine without a GPU
4. Run inference alone on the GPU machine with `--run_data_pipeline` false and `--run_inference` true, reading docs/performance.md for tuning

### Prompt to paste into your agent

```text
With google-deepmind/alphafold3, show me how to run the data pipeline on my CPU server and inference on my GPU server using --run_data_pipeline and --run_inference, and what to copy between them.
```

### Caveat

The data pipeline needs the genetic and template databases, which are large.

Risk, credentials: Keep the model parameters and their terms of use in mind when copying files between machines.

Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/

URL: https://dedupedai.com/use-cases/google-deepmind-alphafold3/split-the-alphafold-3-data-pipeline-from-inference/

## How to pick the right AlphaFold 3 weights for your project with Alphafold3

Repo: [google-deepmind/alphafold3](https://github.com/google-deepmind/alphafold3) | Difficulty: Beginner | Time: 30 minutes | Basis: From the README

Compare the standard weights with the SynthID Bio-structure and Leaving Atom variants before you run.

**Who it is for:** A team deciding how to run AlphaFold 3 for commercial or watermarked work.

**The problem:** Picking the wrong weights or the wrong route can break the terms you agreed to.

**The outcome:** A clear decision on which weights and which route fit your use.

### Steps

1. Read the Obtaining Model Parameters section and the linked WEIGHTS_TERMS_OF_USE.md
2. Note the SynthID Bio-structure version, which predicts watermarked structures, and the AF3 Leaving Atom variant fine-tuned for leaving atom handling
3. For non-commercial one-off work consider alphafoldserver.com, which the README says has a more limited set of ligands and covalent modifications
4. For commercial use follow the README pointer to Gemini Enterprise Agent Platform on Google Cloud

### Prompt to paste into your agent

```text
Using google-deepmind/alphafold3, compare the standard, SynthID Bio-structure and Leaving Atom weights plus the server and Google Cloud routes for <my project>, and tell me which fits.
```

### Caveat

This is a reading and decision task; it does not run a prediction.

Risk, credentials: All weight variants are subject to the AlphaFold 3 weights terms of use.

Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/

URL: https://dedupedai.com/use-cases/google-deepmind-alphafold3/choose-an-alphafold-3-weights-variant/