README.md

---
pretty_name: PhaseFlow
license: apache-2.0
library_name: pytorch
tags:
  - protein
  - peptide
  - liquid-liquid-phase-separation
  - disordered-protein-regions
  - pytorch
---

# PhaseFlow

![PhaseFlow: sequence data to phase-behaviour learning](figures/PhaseFlow.png)

This repository hosts the unified PhaseFlow runtime checkpoint used for peptide
phase prediction, full-length protein LLPS prediction, and droplet-promoting
region (DPR) localization.

| Component | Task | Embedded weights |
| --- | --- | --- |
| Peptide PhaseFlow | Short-peptide 4×4 PSSI phase-diagram prediction and phase-conditioned sequence generation | Peptide tokenizer, shared Transformer, and Flow Matching stack |
| CondenGT / LLPS | Protein-level liquid-liquid phase-separation prediction | Full-protein sequence-context encoder and LLPS readout |
| DPR stack | Driving phase-separation region localization | Region-scoring head and post-processing state for DPR scanning |

Download the checkpoint with:

```bash
hf download GENTEL-Lab/PhaseFlow PhaseFlow.pt --local-dir artifacts/models
```

## Related resources

| Resource | Link | Role |
| --- | --- | --- |
| PhaseFlow code | [github.com/GENTEL-lab/PhaseFlow](https://github.com/GENTEL-lab/PhaseFlow) | Installation, inference scripts, configs, and documentation |
| OpenPhase training data | [GENTEL-Lab/OpenPhase](https://huggingface.co/datasets/GENTEL-Lab/OpenPhase) | Public peptide, LLPS, and DPR research-data packages |
| Online demo | [phaseflow.bio](http://phaseflow.bio/) | Interactive PhaseFlow usage |

## Open access

The checkpoint and accompanying model card in this repository are fully open
source and freely available to the community. We warmly welcome their use in
research, benchmarking, method development, and reproducible studies. Please cite
the relevant PhaseFlow work and retain appropriate source attribution when
using or redistributing the weights.

## Integrity

This repository contains `metadata/file_inventory.csv`. Verify all listed files
before use; the inventory intentionally excludes itself to avoid a
self-referential checksum.

```python
import csv
import hashlib
from pathlib import Path

repo = Path(".")
with (repo / "metadata/file_inventory.csv").open(newline="") as handle:
    for item in csv.DictReader(handle):
        digest = hashlib.sha256((repo / item["path"]).read_bytes()).hexdigest()
        assert digest == item["sha256"], item["path"]
```

| File | Description | SHA256 |
| --- | --- | --- |
| `PhaseFlow.pt` | Combined PhaseFlow runtime checkpoint (peptide, full-protein, and DPR weights) | `7fb0091e6dd5a85bd3a6be7a0b606501700c4b8f28ff9b6e309267835a2fdff0` |

## Responsible use

These weights are research resources, not clinical or diagnostic tools. Model
predictions require independent validation and must be interpreted under the
documented evidence and coordinate conventions in the PhaseFlow codebase.