PartiFlow / README.md
MegaGatlingPea's picture
Replace complete PepInter data with prepared subset
e549d60
|
Raw History Blame Contribute Delete
2.29 kB
---
tags:
- biology
- peptide-design
- flow-matching
library_name: pytorch
---
# PartiFlow
Model checkpoints and the original PepInter training structures for [PartiFlow](https://github.com/MegaGatlingPea/PartiFlow).
## Checkpoints
| File | Use |
| --- | --- |
| `ckpts/partiflow.ckpt` | General denovo and motif design |
| `ckpts/partiflow-aux.ckpt` | Docking, inverse folding, and side-chain packing |
| `ckpts/partiflow-dock.ckpt` | Optional docking specialist |
| `ckpts/partiflow-ss.ckpt` | Secondary-structure, scaffold-topology, and disulfide conditioning |
| `ckpts/autoencoder.ckpt` | Shared autoencoder required by all four models |
Download the checkpoints into a model directory:
```bash
partiflow download --model-dir models
```
The default set is `autoencoder`, `partiflow`, `partiflow-ss`, and `partiflow-aux`; a terminal preselects these four. The docking specialist is available separately with `partiflow download partiflow-dock --model-dir models`.
Use `--model-dir models` with PartiFlow. Select the auxiliary, docking, or scaffold checkpoint explicitly with `--ckpt` when needed. Installation and task examples are in the source repository.
## PepInter subset
[`data/pepinter-subset.tar.xz`](https://huggingface.co/MegaGatlingPea/PartiFlow/resolve/main/data/pepinter-subset.tar.xz) contains 64 training and 16 validation examples, with original PDB structures, prepared atom37 graphs, and a split index listing only those 80 samples. No test examples are included.
The examples were selected by sorted ID within the original train and validation splits. This small subset is intended for trying the VAE/FM training workflow; it is not a representative training corpus and cannot reproduce the paper's full-data training results. The complete PepInter training set is not distributed here.
Download and unpack the prepared subset with the latest PartiFlow:
```bash
partiflow download pepinter
```
The default destination is `train_data/pepinter_subset/`. The included graphs and index work with the bundled training configurations; no initial graph preprocessing is needed. Use `--data-dir DIR` to select another location. See the [training guide](https://github.com/MegaGatlingPea/PartiFlow/blob/main/docs/commands/train.md) for VAE and FM commands.