Update README.md
Browse files
README.md
CHANGED
|
@@ -36,7 +36,17 @@ size_categories:
|
|
| 36 |
|
| 37 |
AptaBench is a benchmark for aptamer–small-molecule interaction prediction. It contains curated DNA/RNA aptamer–ligand pairs with standardized sequences, canonical SMILES, experimentally grounded active/inactive labels, quantitative affinity values where available, and fixed leakage-aware evaluation splits.
|
| 38 |
|
| 39 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 40 |
|
| 41 |
## Dataset summary
|
| 42 |
|
|
@@ -88,24 +98,83 @@ The three protocols are:
|
|
| 88 |
|
| 89 |
Random splits are not recommended because they may overestimate generalization by allowing recurring ligands or aptamer sequences across train and held-out data.
|
| 90 |
|
| 91 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 92 |
|
| 93 |
```python
|
| 94 |
import json
|
| 95 |
import pandas as pd
|
| 96 |
|
| 97 |
-
|
| 98 |
|
| 99 |
-
with open("stratified.json"
|
| 100 |
-
|
| 101 |
|
| 102 |
-
|
| 103 |
-
train_df =
|
| 104 |
-
test_df =
|
| 105 |
```
|
| 106 |
|
|
|
|
|
|
|
|
|
|
| 107 |
## Recommended metrics
|
| 108 |
|
| 109 |
For classification, report ROC-AUC, PR-AUC, accuracy, balanced accuracy, F1-score, precision, and recall.
|
| 110 |
|
| 111 |
-
For affinity regression, report R², RMSE, MAE, and Spearman correlation on entries with available `pKd_value`.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 36 |
|
| 37 |
AptaBench is a benchmark for aptamer–small-molecule interaction prediction. It contains curated DNA/RNA aptamer–ligand pairs with standardized sequences, canonical SMILES, experimentally grounded active/inactive labels, quantitative affinity values where available, and fixed leakage-aware evaluation splits.
|
| 38 |
|
| 39 |
+
|
| 40 |
+
**Authors:** Mariia Eremeyeva and Nikita Serov
|
| 41 |
+
**Affiliation:** Center for AI in Chemistry, ITMO University, St. Petersburg, Russia
|
| 42 |
+
**Publication:** NeurIPS 2026, Evaluations and Datasets Track
|
| 43 |
+
|
| 44 |
+
**Resources:**
|
| 45 |
+
- [Paper](https://neurips.cc/virtual/2026/poster/139036)
|
| 46 |
+
- [Code and reproducibility instructions](https://github.com/GenerativeMolMachines/AptaBench)
|
| 47 |
+
- [Dataset](https://huggingface.co/datasets/GenerativeMolMachines/AptaBench_dataset)
|
| 48 |
+
- [Croissant metadata](https://github.com/GenerativeMolMachines/AptaBench/blob/main/dataset/croissant.json)
|
| 49 |
+
|
| 50 |
|
| 51 |
## Dataset summary
|
| 52 |
|
|
|
|
| 98 |
|
| 99 |
Random splits are not recommended because they may overestimate generalization by allowing recurring ligands or aptamer sequences across train and held-out data.
|
| 100 |
|
| 101 |
+
|
| 102 |
+
## Loading and reproducibility
|
| 103 |
+
|
| 104 |
+
Load the dataset directly from Hugging Face:
|
| 105 |
+
|
| 106 |
+
```python
|
| 107 |
+
from datasets import load_dataset
|
| 108 |
+
|
| 109 |
+
ds = load_dataset(
|
| 110 |
+
"GenerativeMolMachines/AptaBench_dataset",
|
| 111 |
+
split="train"
|
| 112 |
+
)
|
| 113 |
+
df = ds.to_pandas()
|
| 114 |
+
```
|
| 115 |
+
|
| 116 |
+
The Hugging Face `train` split contains the full benchmark table. The benchmark evaluation folds are provided separately in the [GitHub repository](https://github.com/GenerativeMolMachines/AptaBench/tree/main/dataset/splits).
|
| 117 |
+
|
| 118 |
+
To reproduce the published evaluation, clone the repository and load the dataset and split indices from the same release:
|
| 119 |
+
|
| 120 |
+
```bash
|
| 121 |
+
git clone https://github.com/GenerativeMolMachines/AptaBench.git
|
| 122 |
+
cd AptaBench
|
| 123 |
+
```
|
| 124 |
|
| 125 |
```python
|
| 126 |
import json
|
| 127 |
import pandas as pd
|
| 128 |
|
| 129 |
+
df = pd.read_csv("dataset/AptaBench_dataset.csv")
|
| 130 |
|
| 131 |
+
with open("dataset/splits/stratified.json") as f:
|
| 132 |
+
folds = json.load(f)
|
| 133 |
|
| 134 |
+
fold = folds[0]
|
| 135 |
+
train_df = df.iloc[fold["train_idx"]]
|
| 136 |
+
test_df = df.iloc[fold["val_idx"]]
|
| 137 |
```
|
| 138 |
|
| 139 |
+
Replace `stratified.json` with `disjoint_molecule.json` or `disjoint_aptamer.json` to evaluate the corresponding generalization protocol.
|
| 140 |
+
|
| 141 |
+
|
| 142 |
## Recommended metrics
|
| 143 |
|
| 144 |
For classification, report ROC-AUC, PR-AUC, accuracy, balanced accuracy, F1-score, precision, and recall.
|
| 145 |
|
| 146 |
+
For affinity regression, report R², RMSE, MAE, and Spearman correlation on entries with available `pKd_value`.
|
| 147 |
+
|
| 148 |
+
|
| 149 |
+
## Data provenance
|
| 150 |
+
|
| 151 |
+
AptaBench integrates records from eight sources: RSAPred, AptamerBase, AptaGen (Apta-Index), the UTexas Aptamer Database, RiboCentre-aptamer, AptaDB, manually curated primary literature, and the experimental specificity dataset.
|
| 152 |
+
|
| 153 |
+
Records were harmonized through sequence standardization, ligand identification and SMILES canonicalization, affinity conversion to pKd, activity-label curation, and provenance tracking.
|
| 154 |
+
|
| 155 |
+
The dataset retains distinct measurements, including conflicting observations where available. Detailed curation rules, source-level statistics, and overlap analyses are provided in the accompanying paper.
|
| 156 |
+
|
| 157 |
+
## Intended use and limitations
|
| 158 |
+
|
| 159 |
+
AptaBench is intended for benchmarking aptamer–small-molecule binding classification and quantitative affinity prediction.
|
| 160 |
+
|
| 161 |
+
Its main limitations include uneven coverage of ligands and aptamer families, heterogeneity across experimental assays, limited quantitative affinity annotations, and source-specific label imbalance. The benchmark's fixed disjoint protocols support evaluation under distribution shift, but do not eliminate every form of molecular or sequence similarity.
|
| 162 |
+
|
| 163 |
+
|
| 164 |
+
## License
|
| 165 |
+
|
| 166 |
+
The AptaBench dataset is released under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). The accompanying code is available under the MIT License. Original data sources should be cited and used in accordance with their applicable terms.
|
| 167 |
+
|
| 168 |
+
|
| 169 |
+
## Citation
|
| 170 |
+
|
| 171 |
+
If you use AptaBench in your research, please cite:
|
| 172 |
+
|
| 173 |
+
```bibtex
|
| 174 |
+
@inproceedings{eremeyeva2026aptabench,
|
| 175 |
+
title = {AptaBench: A Benchmark for Aptamer-Small Molecule Binding},
|
| 176 |
+
author = {Eremeyeva, Mariia and Serov, Nikita},
|
| 177 |
+
booktitle = {NeurIPS 2026 Evaluations and Datasets Track},
|
| 178 |
+
year = {2026}
|
| 179 |
+
}
|
| 180 |
+
```
|