---
size_categories:
- 10K<n<100K
task_categories:
- text-classification
dataset_info:
  features:
  - name: Canonicalized SMILES
    dtype: string
  - name: Standardized SMILES
    dtype: string
  - name: Canonicalized Taste
    dtype: string
  - name: Source
    dtype: string
  splits:
  - name: full
    num_bytes: 1865867
    num_examples: 15031
  - name: train
    num_bytes: 1304664
    num_examples: 10521
  - name: validation
    num_bytes: 278960
    num_examples: 2255
  - name: test
    num_bytes: 282243
    num_examples: 2255
  download_size: 1225561
  dataset_size: 3731734
configs:
- config_name: default
  data_files:
  - split: full
    path: data/full-*
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
tags:
- chemistry
---

# FartDB

<!-- Provide a quick summary of the dataset. -->

Composite dataset of 15,031 molecules and their taste (sweet, bitter, umami, sour, undefined). 

## Dataset Details

### Dataset Description

<!-- Provide a longer summary of what this dataset is. -->
FartDB is a curated dataset drawn from five data sources: FlavorDB, PlantMolecularTasteDB, ChemTastesDB, Tas2R Agonists DB and Scifinder.
A canonicalized SMILES is mapped to one of five flavor categories: sweet, bitter, umami, sour, undefined. Salty molecules are not considered as only a small number of compounds have this taste. 
The dataset was enriched with other descriptors from PubChem where available and when multiple datasets contained the same SMILES/Taste data point, these duplicates were removed. 

Note that a small number (< 5) of canonicalized SMILES in the dataset may not form valid molecules. These may need to be removed before use. 


- **Curated by:** Fart Labs
- **License:** MIT

### Dataset Sources

<!-- Provide the basic links for the dataset. -->

FlavorDB:
Neelansh Garg†, Apuroop Sethupathy†, Rudraksh Tuwani†, Rakhi NK†, Shubham Dokania†, Arvind Iyer†, Ayushi Gupta†, Shubhra Agrawal†, Navjot Singh†, Shubham Shukla†, Kriti Kathuria†, Rahul Badhwar, Rakesh Kanji, Anupam Jain, Avneet Kaur, Rashmi Nagpal, and Ganesh Bagler*, FlavorDB: A database of flavor molecules, Nucleic Acids Research, gkx957, (2017). †Equal contribution *Corresponding Author
<https://doi.org/10.1093/nar/gkx957>

PlantMolecularTasteDB:
Gradinaru Teodora-Cristiana, Madalina Petran, Dorin Dragos, and Marilena Gilca. "PlantMolecularTasteDB: A Database of Taste Active Phytochemicals." *Frontiers in Pharmacology* 2022; 12:3804. 
<https://doi.org/10.3389/fphar.2021.751712>

ChemTastesDB:
Rojas, C., Ballabio, D., Pacheco Sarmiento, K., Pacheco Jaramillo, E., Mendoza, M., & García, F. (2021). ChemTastesDB: A Curated Database of Molecular Tastants (1.0) [Data set]. Zenodo. 
<https://doi.org/10.5281/zenodo.5747393>

Tas2R Agonists DB:
Sebastian Bayer, Ariane Isabell Mayer, Gigliola Borgonovo, Gabriella Morini, Antonella Di Pizio, and Angela Bassoli, Journal of Agricultural and Food Chemistry 2021 69 (46), 13916-13924, 
<https://doi.org/10.1021/acs.jafc.1c05057>

IUPAC Digitilized Dissociation Constants:
Jonathan Zheng, Olivier Lafontant-Joseph
<https://doi.org/10.5281/zenodo.7236453>

Umami Compounds (manually added):
B. Suess, D. Festring, T. Hofmann, 15 - Umami compounds and taste enhancers, Editor(s): J.K. Parker, J.S. Elmore, L. Methven, In Woodhead Publishing Series in Food Science, Technology and Nutrition, Flavour Development, Analysis and Perception in Food and Beverages, Woodhead Publishing, 2015, Pages 331-351, ISBN 9781782421030,
<https://doi.org/10.1016/B978-1-78242-103-0.00015-1>

## Uses

<!-- Address questions around how the dataset is intended to be used. -->

This dataset is intended for the training of machine learning models, in particular transformer models trained on SMILES such as ChemBERTa. 

### Direct Use

<!-- This section describes suitable use cases for the dataset. -->

This dataset has been used to finetune ChemBERTa to be able to predict the flavor from an arbitrary SMILES input. 

## Dataset Structure

<!-- This section provides a description of the dataset fields, and additional information about the dataset structure such as criteria used to create the splits, relationships between data points, etc. -->
The first two columns contain: "Canonicalized SMILES" as canonicalized by RDKit and the "Canonicalized Flavor " (sweet, sour, umami, bitter, undefined).
"Source": which database this data point derives from 

PubChem descriptors: Where available, the dataset was enriched with descriptors accessible through the PubChem API. 
"PubChemID", "IUPAC Name", "Molecular Formula", "Molecular Weight", "InChI", "InChI Key"

## Dataset Creation

### Curation Rationale

<!-- Motivation for the creation of this dataset. -->

This dataset contains the majority of all known SMILES to flavor mappings publicly available. In order to use this data for supervised machine learning, both the SMILES and the flavor categories had to be made consistent. 


### Source Data

<!-- This section describes the source data (e.g. news text and headlines, social media posts, translated sentences, ...). -->

All databases contained SMILES or SMILES data was obtained for the molecule from PubChem. The databases were previously curated dataset of tastants with three exceptions: Tas2R Agonists DB lists molecules which bind the human taste receptor; Scifinder was used to search for acidic molecules even if they had not been expressly tested for taste; some umami compounds were added manually from the literature. 

#### Data Collection and Processing

<!-- This section describes the data collection and processing process such as data selection criteria, filtering and normalization methods, tools and libraries used, etc. -->

SMILES were canonicalized with RDKit
Any rows with empty fields were removed. 
PubChem properties were enriched after duplicates were removed based on "Canonicalized SMILES" and "Canonicalized Flavor".

FlavorDB: FlavorDB has human generated labels (e.g. "honey", "sweet-like", "tangy") which then had to be mapped onto the 5 canonical flavor categories. 
ChemTastesDB: Human-generated labels were again canonicalized into the 5 flavor categories. 
PhytocompoundsDB: Human-generated labels were again canonicalized into the 5 flavor categories. 
Tas2R Agonists DB: This dataset contains molecules which bind the human bitter receptor. All datapoints were hence labelled as "bitter". 
Scifinder: A random subset of small molecules (< 500 Da) that are listed with a pKa between 2 and 7 which is a safe range for human tasting.
Umami DB: A small dataset of known umami compounds was manually curated from the literature.

## Citation

<!-- If there is a paper or blog post introducing the dataset, the APA and Bibtex information for that should go in this section. -->

```
Zimmermann Y, Sieben L, Seng H, Pestlin P, Görlich F. A Chemical Language Model for Multi-Class Molecular Taste Prediction. ChemRxiv. 2024; doi:10.26434/chemrxiv-2024-d6n15  This content is a preprint and has not been peer-reviewed.
```

```
@unpublished{Zimmermann2024chemical,
	doi = {10.26434/chemrxiv-2024-d6n15},
	publisher = {American Chemical Society (ACS)},
	title = {A Chemical Language Model for Multi-Class Molecular Taste Prediction},
	url = {http://dx.doi.org/10.26434/chemrxiv-2024-d6n15},
	author = {Zimmermann, Yoel and Sieben, Leif and Seng, Henrik and Pestlin, Philipp and G{\"o}rlich, Franz},
	date = {2024-11-19},
	year = {2024},
	month = {11},
	day = {19},
}
```