README.md

---
task_categories:
- tabular-regression
- feature-extraction
tags:
- antibody
- antibody-developability
- assay-metadata
- multimodal-learning
---

# AbAssayBench

This dataset repository contains the processed data package for **AbAssayBench**,
a multi-endpoint antibody developability benchmark. The release combines the
FLAb2.0-derived antibody measurements used for model development with the
PROPHET-Ab measurements used for external validation.

The repository is intended to be used together with the MAP-Ab source code:
`https://github.com/gu-yaowen/MAP-Ab`.

## Package layout

| Path | Contents |
| --- | --- |
| `tables/` | Release tables with stable IDs and repository-relative file references. |
| `assay_metadata/` | One JSON file per referenced assay metadata record. |
| `structures/` | Antibody PDB files used by the processed structure features. |
| `features/` | Precomputed sequence, structure, and assay-metadata feature stores. |
| `splits/` | Long-format split assignments and split-size summaries. |
| `results/` | Model predictions, endpoint metrics, tables, and publication figures. |
| `schemas/` | Machine-readable description of the release contract. |
| `manifests/` | Dataset, assay, structure, metadata, and file manifests. |

## Main tables

`tables/flab2_measurements.csv` and `tables/prophet_ab_measurements.csv` retain
the measurement-level labels and canonical assay columns. Both tables add the
following release identifiers:

- `measurement_id`: stable identifier for one measurement row;
- `antibody_id`: stable identifier for a heavy/light-chain pair;
- `endpoint_id`: stable identifier for a property and endpoint definition;
- `structure_id`: stable identifier for the source structure reference;
- `assay_metadata_file`: path relative to this repository;
- `structure_file`: path relative to this repository;
- `property`: normalized broad property label;
- `source_row_index`: original row index in the source table.

The numeric measurement label is stored in the original `fitness` column.
`value_definition` identifies the endpoint definition, while `assay_id`
identifies the assay unit. The broad property labels are aggregation,
expression, immunogenicity, pharmacokinetics, polyreactivity, and
thermostability for the FLAb2.0-derived data; PROPHET-Ab retains its five
benchmark property labels in the same schema.

## Feature-store contract

Each feature directory contains `features.npy`, `metadata.csv`, and, when
available, `feature_info.json`. Row `i` in `features.npy` corresponds to row
`i` in `metadata.csv`; `measurement_id` is the preferred join key. Feature
arrays are stored as `float32` and are not re-normalized by this release.
Directory names identify the feature family, including `esmc_600m`, `ism_3b`,
`propermab_struct`, and the assay-metadata embedding stores.

## Splits and results

The split package is intentionally long-format: filter
`splits/split_assignments.csv` by `split_family` and `replicate`, then join on
`measurement_id`. The provided results are frozen outputs from the project
analysis and are not required to reproduce the feature stores.

## Downloading from Hugging Face
```bash
hf download yg3191/AbAssayBench \
  --repo-type dataset \
  --local-dir ./AbAssayBench
```

## Loading example

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

root = Path("AbAssayBench")
measurements = pd.read_parquet(root / "tables/flab2_measurements.parquet")
feature_meta = pd.read_csv(root / "features/flab2/esmc_600m/metadata.csv")
features = np.load(root / "features/flab2/esmc_600m/features.npy", mmap_mode="r")
assert len(feature_meta) == features.shape[0]
```

The checksum file at `manifests/checksums.sha256` covers the release files.
The source-attribution manifest records the component datasets and their roles;
source redistribution conditions should be checked before public reuse.