llps/README.md

---
pretty_name: PhaseFlow-LLPS
license: other
tags:
  - protein
  - liquid-liquid-phase-separation
  - biomolecular-condensates
  - protein-classification
---

# PhaseFlow-LLPS

PhaseFlow-LLPS is the protein-level liquid-liquid phase-separation (LLPS) learning package used by the PhaseFlow workflow. It provides canonical protein sequences, supervised training units, the fixed three-epoch batch plans, and evaluation inputs. Evaluation files are isolated under `benchmark/` and must never enter training or model selection.

## 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 public PhaseFlow LLPS data protocol, build compatible training inputs, and evaluate frozen models on the distributed benchmark inputs. It supports research on sequence-based LLPS prediction and data provenance. It is not a clinical, diagnostic, or experimentally validated annotation resource for individual proteins.

The supplied benchmark must not be used for fitting, feature selection, threshold optimization, checkpoint selection, or other model-development decisions when it is the reported evaluation target.

## Data Structure

```text
llps/
├── benchmark/                    # evaluation inputs only
│   ├── constructs.csv
│   ├── panel_membership.csv
│   └── proteins.csv
├── configs/
│   ├── dataset_contract.yaml
│   └── training.yaml
├── data/
│   ├── proteins.parquet
│   ├── training_units.parquet
│   └── training_plan/
│       └── batch_plan_epoch_*.parquet
└── metadata/
    ├── data_dictionary.md
    ├── dataset_summary.json
    ├── feature_reproduction.md
    └── file_inventory.csv
```

`data/proteins.parquet` contains canonical sequence records. Join `data/training_units.parquet` to it on both `protein_id` and `sequence_sha256`. The training-unit table provides label status, weights, evidence roles, teacher-derived covariates, and protocol metadata. Each row in `data/training_plan/batch_plan_epoch_*.parquet` must resolve by both `dataset_index` and `plan_dataset_index` to the same training-unit identifier and sequence hash.

`benchmark/proteins.csv` and `benchmark/constructs.csv` contain evaluation sequences and labels, while `benchmark/panel_membership.csv` provides membership in named evaluation panels. `constructs.csv` was losslessly transcoded from its legacy GB18030 source encoding to UTF-8 for portable CSV loading; its records and fields were not otherwise changed. See `metadata/data_dictionary.md` for field-level guidance.

## Quickstart

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

root = Path("llps")
proteins = pd.read_parquet(root / "data/proteins.parquet")
units = pd.read_parquet(root / "data/training_units.parquet")

training = units.merge(
    proteins,
    on=["protein_id", "sequence_sha256"],
    validate="one_to_one",
)
benchmark = pd.read_csv(root / "benchmark/proteins.csv")
```

## Reproducibility

Read the two configuration files and `metadata/feature_reproduction.md` before using the data. The release intentionally excludes all precomputed embeddings, graph tensors, and packed feature caches. Regenerate them from the distributed sequences with the published PhaseFlow preprocessing workflow, and keep any derived files outside this package. Use the fixed training-plan files only after checking their documented key identity. Verify transfers with `metadata/file_inventory.csv`.

## Limitations

LLPS labels, source provenance, teacher-derived covariates, and negative strata reflect the release’s curation and evidence policy. They are not exhaustive biological ground truth and may not generalize across species, assay conditions, mutations, or cellular contexts. Predictions built from this package require independent experimental validation. The benchmark files are evaluation inputs and must remain isolated from model development when they are the reported evaluation target.

## Citation and attribution

Please cite the accompanying PhaseFlow manuscript or public repository release. Preserve and cite the relevant upstream data sources when using benchmark or provenance fields.