> For the complete documentation index, see [llms.txt](https://jwliaomath.gitbook.io/cocofold2/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://jwliaomath.gitbook.io/cocofold2/getting-started/7zdt_7zd5.md).

# Minimal example: 7ZDT/7ZD5

This map-derived example uses the 7ZDT sequence/MSA to generate a Protenix-v1 cache and the complete original 7ZD5 CIF to generate supervision. It does not trim the CIF to common residues or simulate heterogeneous particles. The author accepted smoke and refine in the existing server environment.

Run commands from the repository root after [installing dependencies](/cocofold2/getting-started/installation.md). Choose new output directories; prediction caches and training results refuse overwrite. The sequence/MSA JSON, weights and common resources are user inputs. The two original CIFs are under `examples/7zdt_7zd5/inputs/`.

## 1. Generate the map locally

Create a JSON configuration, replacing the paths:

```json
{
  "target_cif": "/absolute/path/to/CoCoFold2/examples/7zdt_7zd5/inputs/7ZD5.cif",
  "output_dir": "/absolute/path/to/new_map",
  "boxsize": 192,
  "apix": 1,
  "resolution": 3
}
```

Set `COCOFOLD2_MOLMAP_CONFIG` to this JSON and run `examples/7zdt_7zd5/make_molmap.py` inside ChimeraX:

```bash
export COCOFOLD2_MOLMAP_CONFIG=/absolute/path/to/molmap_config.json
ChimeraX --nogui --exit --script /absolute/path/to/CoCoFold2/examples/7zdt_7zd5/make_molmap.py
```

Use your installed ChimeraX executable path. The accepted map was generated with ChimeraX 1.7.1. The script uses all atoms loaded from the CIF, translates their electron-weighted center to zero, and creates a 192³ grid with origin (-96, -96, -96) Å. `molmap.json` records the exact translation, source/map hashes and atom count. It does not modify the input CIF. Transfer the generated `7zd5_full_3A.mrc` and `molmap.json` together if training on a server.

## 2. Simulate particles

```bash
python src/simulate_particles.py --map /path/to/new_map/7zd5_full_3A.mrc \
  --out-dir results/small_particles --n-particles 1000 --snr 1 --apix 1 --seed 42
```

The output contains a STAR file, noisy and clean CTF particle stacks, simulation parameters and readback checks. The two stacks together occupy about 281 MiB. STAR paths are relative to the simulation directory. This step needs no model.

## 3. Predict, then place the initial model

```bash
export PROTENIX_ROOT_DIR=/absolute/path/to/protenix_resources
python examples/7zdt_7zd5/run_case.py predict \
  --input-json /path/to/7zdt.json --resource-root "$PROTENIX_ROOT_DIR" \
  --output results/small_prediction --seed 42
```

The resources directory contains `common/` and `checkpoint/`. Prediction retains the upstream seed 101; `--seed 42` here selects the initial export/refinement diffusion seed. The stage ends with an initial CIF and a completion report. Use this same successful prediction directory in both training commands.

Fit `results/small_prediction/initial/7zdt_initial_prediction.cif` to the **generated map**, then save only the fitted atomic model as `/path/to/placed.cif`. Preserve all atom identities and order. Do not fit the prediction to a deposited truth structure instead. The deposited 7ZDT CIF is not automatically a compatible training reference for a newly generated cache.

An independently prepared reference with matching topology but a different conformation is allowed via `--reference-structure`; its map placement and atom identity/order remain the user's responsibility. Do not add this flag merely to bypass a failed rigid-movement check without understanding the difference.

## 4. Smoke, then refine

```bash
python examples/7zdt_7zd5/run_case.py train --kind smoke \
  --prediction results/small_prediction --particles results/small_particles \
  --placed-cif /path/to/placed.cif --output results/small_smoke
python examples/7zdt_7zd5/run_case.py train --kind refine \
  --prediction results/small_prediction --particles results/small_particles \
  --placed-cif /path/to/placed.cif --smoke-result results/small_smoke \
  --output results/small_refine
```

Smoke uses the first 32 particles for two epochs/two optimizer steps. Refine uses all 1000 for ten epochs/320 steps. Both use batch=32, mini-batch=16, seed=42, fixed stochasticity and explicitly frozen GMM amplitudes/widths. Keep these settings and the inputs identical between smoke and refine. The generic trainer uses fixed-frame projection for new runs. This example passes `--projection-origin 0 0 0` explicitly because its map was centered at physical `(0,0,0)` Å; do not reuse that value for a different map without checking its frame. See [choosing the origin](/cocofold2/reference/parameter_guide.md#choosing-the-projection-origin). Slurm is optional; submit the same commands on a suitable GPU compute node if required by your cluster.

Inspect `validation.json`/`validation.txt` in each output directory. Checks cover finite values, progress, latent updates, exact GMM freezing, output consistency and records without repeating diffusion decoding. The reported eight-particle FRC comparison uses training views, not held-out validation. If only the checker fails, inspect the error and recheck the existing results before retraining.

## 5. Inspect structural quality

Inspect the final CIF in the map. A Cα RMSD within 2 Å is a manual inspection target, not an automated condition. 7ZD5 and the modeled 7ZDT input can differ in sequence and atoms; when using an external alignment tool, report chain mapping, matched residues and coverage along with RMSD. Automated success alone does not establish structural accuracy. See [detailed test scope](https://github.com/jwliaomath/CoCoFold2/tree/main/examples/7zdt_7zd5/RUN_TESTS.md).


---

# Agent Instructions
This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com.

## Querying This Documentation
If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question.

Perform an HTTP GET request on the current page URL with the `ask` query parameter, and the optional `goal` query parameter:

```
GET https://jwliaomath.gitbook.io/cocofold2/getting-started/7zdt_7zd5.md?ask=<question>&goal=<endgoal>
```

`ask` is the immediate question: it should be specific, self-contained, and written in natural language.
`goal` is optional and describes the broader end goal you are ultimately trying to accomplish on behalf of the user. GitBook uses it to tailor the answer towards what is most useful for that goal.

The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.

Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.