File size: 6,647 Bytes
915dc4a
 
 
 
e4a4623
 
 
 
 
 
 
915dc4a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3b14ca2
 
 
 
 
 
 
 
 
 
 
915dc4a
 
 
 
 
 
 
 
 
 
 
 
 
 
9f4a35f
 
 
 
 
 
 
915dc4a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3b14ca2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
915dc4a
 
 
 
 
3b14ca2
915dc4a
3b14ca2
 
915dc4a
3b14ca2
 
 
915dc4a
 
3b14ca2
 
 
915dc4a
 
 
 
3b14ca2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
---
license: cc-by-4.0
language:
- en

configs:
- config_name: default
  data_files:
  - split: train
    path: dataset/AptaBench_dataset.csv

pretty_name: AptaBench
task_categories:
- tabular-classification
- tabular-regression
tags:
- aptamers
- small-molecules
- dna
- rna
- molecular-machine-learning
- binding-prediction
- affinity-prediction
- benchmark
- leakage-aware-evaluation
size_categories:
- 1K<n<10K
---

<p align="center">
  <img src="logo.png" width="500">
</p>


# AptaBench

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.


**Authors:** Mariia Eremeyeva and Nikita Serov  
**Affiliation:** Center for AI in Chemistry, ITMO University, St. Petersburg, Russia  
**Publication:** NeurIPS 2026, Evaluations and Datasets Track

**Resources:**
- [Paper](https://neurips.cc/virtual/2026/poster/139036)
- [Code and reproducibility instructions](https://github.com/GenerativeMolMachines/AptaBench)
- [Dataset](https://huggingface.co/datasets/GenerativeMolMachines/AptaBench_dataset)
- [Croissant metadata](https://github.com/GenerativeMolMachines/AptaBench/blob/main/dataset/croissant.json)


## Dataset summary

The current release contains 6,289 aptamer–ligand records covering 1,610 unique aptamer sequences and 942 unique ligands from eight curated sources.

The benchmark supports two tasks:

- **Binding classification**: predict active vs inactive aptamer–ligand pairs.
- **Affinity regression**: predict pKd values for entries with quantitative affinity annotations.

Inactive labels are based on reported non-binding or low-affinity observations, not synthetic random cross-pairing.

## Files

- `dataset/AptaBench_dataset.csv`: main aptamer–ligand interaction table in CSV format.
- `dataset/AptaBench_dataset.parquet`: parquet version of the dataset for efficient loading and processing.
- `dataset/splits/`: directory containing fixed leakage-aware evaluation splits.
  - `stratified.json`: stratified in-distribution 5-fold split.
  - `disjoint_molecule.json`: molecule-disjoint 5-fold split.
  - `disjoint_aptamer.json`: aptamer-disjoint 5-fold split.
- `dataset/make_splits.py`: script used to generate evaluation splits.

## Data fields

- `type`: aptamer type, DNA or RNA.
- `sequence`: standardized aptamer sequence.
- `canonical_smiles`: canonical ligand SMILES.
- `pKd_value`: transformed dissociation constant value, where available.
- `label`: binary activity label; `1` = active, `0` = inactive or low-affinity.
- `buffer`: reported experimental buffer or assay condition.
- `origin`: source publication or database record.
- `source`: curated source name.

## Evaluation splits

Each split file contains five folds with:

- `fold`: fold identifier.
- `train_idx`: zero-based row indices for training.
- `val_idx`: zero-based held-out row indices used for validation or testing.

The indices refer to rows in `AptaBench_dataset.csv`.

The three protocols are:

- **Stratified**: in-distribution evaluation.
- **Molecule-disjoint**: held-out ligands do not appear in the corresponding training fold.
- **Aptamer-disjoint**: held-out aptamer sequences do not appear in the corresponding training fold.

Random splits are not recommended because they may overestimate generalization by allowing recurring ligands or aptamer sequences across train and held-out data.


## Loading and reproducibility

Load the dataset directly from Hugging Face:

```python
from datasets import load_dataset

ds = load_dataset(
    "GenerativeMolMachines/AptaBench_dataset",
    split="train"
)
df = ds.to_pandas()
```

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).

To reproduce the published evaluation, clone the repository and load the dataset and split indices from the same release:

```bash
git clone https://github.com/GenerativeMolMachines/AptaBench.git
cd AptaBench
```

```python
import json
import pandas as pd

df = pd.read_csv("dataset/AptaBench_dataset.csv")

with open("dataset/splits/stratified.json") as f:
    folds = json.load(f)

fold = folds[0]
train_df = df.iloc[fold["train_idx"]]
test_df = df.iloc[fold["val_idx"]]
```

Replace `stratified.json` with `disjoint_molecule.json` or `disjoint_aptamer.json` to evaluate the corresponding generalization protocol.


## Recommended metrics

For classification, report ROC-AUC, PR-AUC, accuracy, balanced accuracy, F1-score, precision, and recall.

For affinity regression, report R², RMSE, MAE, and Spearman correlation on entries with available `pKd_value`.


## Data provenance

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.

Records were harmonized through sequence standardization, ligand identification and SMILES canonicalization, affinity conversion to pKd, activity-label curation, and provenance tracking.

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.

## Intended use and limitations

AptaBench is intended for benchmarking aptamer–small-molecule binding classification and quantitative affinity prediction.

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.


## License

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.


## Citation

If you use AptaBench in your research, please cite:

```bibtex
@inproceedings{eremeyeva2026aptabench,
  title     = {AptaBench: A Benchmark for Aptamer-Small Molecule Binding},
  author    = {Eremeyeva, Mariia and Serov, Nikita},
  booktitle = {NeurIPS 2026 Evaluations and Datasets Track},
  year      = {2026}
}
```