README.md

---
pretty_name: SKEMPI v2
license: other
tags:
- biology
- protein
- protein-protein-interaction
- binding-affinity
- mutation
- protein-engineering
- skempi
- skempi2
- jsonl
configs:
- config_name: default
  data_files:
  - split: train
    path: tables/labeled_skempi2_skempi_v2.csv.jsonl
---

# SKEMPI v2

SKEMPI v2 is a manually curated benchmark of experimentally measured changes in protein-protein binding affinity, kinetics, and thermodynamics upon mutation. It focuses on structurally resolved protein-protein interactions, linking mutations to PDB complexes and literature-derived binding measurements. The dataset is commonly used to evaluate models that predict mutation effects on protein interfaces, binding free energy changes, and protein complex stability.

This Hugging Face mirror stores the SKEMPI v2 CSV as normalized JSONL rows with provenance. Each record contains the original upstream row under `row`, plus `dataset_id`, `row_index`, and `source_file`. The upstream table contains 7,085 mutation records, including single and multiple mutations, wild-type and mutant affinities, optional kinetic measurements, optional enthalpy and entropy values, experiment temperature, literature references, protein names, and SKEMPI version labels.

## Splits

- `train`: 7,085 rows

There is no official train/test split in this mirror. Users should define task-specific splits carefully, for example by holding out PDB complexes, protein pairs, or interaction families, to avoid leakage between similar mutations from the same complex.

## Columns

Every JSONL record has these outer fields:

- `dataset_id`: dataset identifier, always `skempi2`
- `row`: raw upstream SKEMPI CSV row stored as a nested JSON object
- `row_index`: zero-based row index in the upstream source table
- `source_file`: original source path, `labeled/skempi2/skempi_v2.csv`

The upstream SKEMPI CSV fields are semicolon-separated and include:

- `#Pdb`: PDB complex and chain identifier
- `Mutation(s)_PDB`: mutation notation using PDB residue numbering
- `Mutation(s)_cleaned`: cleaned mutation notation
- `iMutation_Location(s)`: mutation location class, such as core, rim, support, or surface
- `Hold_out_type`, `Hold_out_proteins`: split/grouping metadata from SKEMPI
- `Affinity_mut (M)`, `Affinity_mut_parsed`: mutant binding affinity
- `Affinity_wt (M)`, `Affinity_wt_parsed`: wild-type binding affinity
- `Reference`: source publication identifier
- `Protein 1`, `Protein 2`: interacting protein names
- `Temperature`: experimental temperature
- `kon_mut`, `kon_wt`: mutant and wild-type association rates when available
- `koff_mut`, `koff_wt`: mutant and wild-type dissociation rates when available
- `dH_mut`, `dH_wt`: mutant and wild-type enthalpy values when available
- `dS_mut`, `dS_wt`: mutant and wild-type entropy values when available
- `Notes`: free-text notes
- `Method`: experimental method
- `SKEMPI version`: source SKEMPI version label

## Usage

Download the repository:

```bash
hf download LiteFold/SKEMPI2 --repo-type dataset --local-dir ./skempi2
```

Load the JSONL table with `datasets`:

```python
from datasets import load_dataset

ds = load_dataset("LiteFold/SKEMPI2", split="train")
print(ds[0])
```

The current mirror keeps the original semicolon-separated CSV row as the single value inside `row`. This helper expands it into a normal dictionary:

```python
from datasets import load_dataset

HEADER = (
    "#Pdb;Mutation(s)_PDB;Mutation(s)_cleaned;iMutation_Location(s);"
    "Hold_out_type;Hold_out_proteins;Affinity_mut (M);Affinity_mut_parsed;"
    "Affinity_wt (M);Affinity_wt_parsed;Reference;Protein 1;Protein 2;"
    "Temperature;kon_mut (M^(-1)s^(-1));kon_mut_parsed;"
    "kon_wt (M^(-1)s^(-1));kon_wt_parsed;koff_mut (s^(-1));"
    "koff_mut_parsed;koff_wt (s^(-1));koff_wt_parsed;"
    "dH_mut (kcal mol^(-1));dH_wt (kcal mol^(-1));"
    "dS_mut (cal mol^(-1) K^(-1));dS_wt (cal mol^(-1) K^(-1));"
    "Notes;Method;SKEMPI version"
)

fields = HEADER.split(";")

def expand_record(record):
    raw_value = record["row"][HEADER]
    values = raw_value.split(";")
    parsed = dict(zip(fields, values))
    parsed["dataset_id"] = record["dataset_id"]
    parsed["row_index"] = record["row_index"]
    parsed["source_file"] = record["source_file"]
    return parsed

ds = load_dataset("LiteFold/SKEMPI2", split="train")
row = expand_record(ds[0])

print(row["#Pdb"])
print(row["Mutation(s)_cleaned"])
print(row["Affinity_mut_parsed"], row["Affinity_wt_parsed"])
```

Stream and compute a simple binding-affinity change target:

```python
import math
from datasets import load_dataset

R = 1.9872036e-3  # kcal mol^-1 K^-1
DEFAULT_T = 298.15

ds = load_dataset("LiteFold/SKEMPI2", split="train", streaming=True)

for record in ds:
    row = expand_record(record)
    kd_mut = float(row["Affinity_mut_parsed"])
    kd_wt = float(row["Affinity_wt_parsed"])
    temperature = float(row["Temperature"] or DEFAULT_T)
    ddg = R * temperature * math.log(kd_mut / kd_wt)
    print(row["#Pdb"], row["Mutation(s)_cleaned"], ddg)
    break
```

Load directly from the JSONL path:

```python
from datasets import load_dataset

ds = load_dataset(
    "json",
    data_files="hf://datasets/LiteFold/SKEMPI2/tables/labeled_skempi2_skempi_v2.csv.jsonl",
    split="train",
)
```

## Data Notes

SKEMPI v2 contains experimentally curated values, but it is still a benchmark that needs careful splitting. Random row splits can leak information because many records share the same PDB complex, protein pair, or closely related mutations. For model evaluation, prefer complex-level, protein-pair-level, or interaction-family-level splits depending on the question.

Binding affinities are reported as dissociation constants in molar units. A common derived supervised target is `ddG = R * T * ln(Kd_mut / Kd_wt)`, but users should check units, missing temperatures, censored or non-detectable binding values, and multi-mutation rows before training or benchmarking.

The Hugging Face dataset viewer may fail to preview this mirror because the original CSV row is nested under a long header key. The local JSONL and `datasets` loading examples above are the most reliable way to consume the current upload.

## License

The dataset card metadata uses `license: other` because SKEMPI v2 should be used according to the upstream SKEMPI terms and citation requirements.

# Citation

```bibtex
@article{jankauskaite2019skempi2,
  title   = {{SKEMPI} 2.0: an updated benchmark of changes in protein-protein binding energy, kinetics and thermodynamics upon mutation},
  author  = {Jankauskait{\.e}, Justina and Jim{\'e}nez-Garc{\'i}a, Brian and Dapk{\=u}nas, Justas and Fern{\'a}ndez-Recio, Juan and Moal, Iain H.},
  journal = {Bioinformatics},
  volume  = {35},
  number  = {3},
  pages   = {462--469},
  year    = {2019},
  doi     = {10.1093/bioinformatics/bty635},
  url     = {https://doi.org/10.1093/bioinformatics/bty635}
}
```