dpr/README.md

---
pretty_name: PhaseFlow-DPR
license: other
tags:
  - protein
  - disordered-protein-regions
  - phase-separation
  - region-localization
---

# PhaseFlow-DPR

PhaseFlow-DPR is the region-supervision package for learning driving phase-separation regions (DPRs) from protein sequences. It contains source protein records, deterministic base-training order, final region-refinement supervision, and a PhasePro-derived evaluation benchmark. The benchmark is physically separated from training data and must remain evaluation-only.

## Open access

All files in this package are openly available for research, benchmarking,
method development, and reproducible studies. We welcome community use and
redistribution; please cite the PhaseFlow work and retain relevant source
attribution.

## Intended Use

Use this package to reproduce or audit the PhaseFlow DPR data preparation and training protocol, to build compatible preprocessing pipelines, and to evaluate frozen models on the supplied benchmark. It is appropriate for research on sequence-based region scoring and localization. It is not a clinical annotation resource and does not establish biological causality for an individual sequence.

Do not train, select checkpoints, tune thresholds, or otherwise optimize on `benchmark/` when reporting evaluation on that benchmark. The published DPR training identifiers and PhasePro benchmark identifiers are disjoint.

## Data Structure

```text
dpr/
├── benchmark/                    # PhasePro-derived evaluation inputs only
├── configs/
│   ├── dataset_contract.yaml     # package-level data contract
│   └── training.yaml             # public training settings
├── data/
│   ├── proteins.parquet          # unique sequence records
│   ├── training_units.parquet    # base/refinement membership
│   ├── base_training_schedule.parquet
│   └── region_supervision.parquet
└── metadata/
    ├── data_dictionary.md
    ├── dataset_summary.json
    ├── feature_reproduction.md
    └── file_inventory.csv
```

`protein_id` and `sequence_sha256` form the required join keys. Always validate both keys when joining schedule or supervision tables to `data/proteins.parquet`. `training_units.parquet` labels the use of each protein as base training, region refinement, or both. `base_training_schedule.parquet` captures the deterministic update/rank/slot sequence. `region_supervision.parquet` contains the final region-level targets and supervision metadata.

`benchmark/` is documented in detail in [`benchmark/README.md`](benchmark/README.md). It uses independent protein IDs and includes explicit coordinate fields; do not infer coordinates from array indices alone.

## Quickstart

```python
from pathlib import Path
import pandas as pd

root = Path("dpr")
proteins = pd.read_parquet(root / "data/proteins.parquet")
units = pd.read_parquet(root / "data/training_units.parquet")
regions = pd.read_parquet(root / "data/region_supervision.parquet")

training = units.merge(
    proteins,
    on=["protein_id", "sequence_sha256"],
    validate="many_to_one",
)
assert (regions["region_start"] < regions["region_end"]).all()
```

## Reproducibility

Read `configs/dataset_contract.yaml`, `configs/training.yaml`, and `metadata/feature_reproduction.md` before training. This source-first package deliberately excludes derived embeddings, graph tensors, and packed feature caches. Regenerate them from `data/proteins.parquet` with the published PhaseFlow preprocessing code and retain caches outside this directory. Verify every distributed file against `metadata/file_inventory.csv` before use.

## Limitations

The available region labels reflect their source evidence, curation choices, and coordinate conversion rules. They do not describe every possible condensate-driving mechanism. Benchmark scores are not a substitute for experimental validation, and dataset composition may not represent all taxa, protein families, or experimental contexts. Cite the original PhaSePro resource when using or discussing benchmark-derived annotations.

## Citation and attribution

Please cite the accompanying PhaseFlow manuscript or public repository release. Cite the original PhaSePro resource when using or discussing the benchmark-derived annotations.