Distributed Representation of Chemical Fragments

20 October 2017, Version 1
This content is a preprint and has not undergone peer review at the time of posting.


1. Approximately 3.7 million (3,700,103) PubChem compounds were used in training the fragment vectors. Five million compounds (from CID 1 to 5,000,000) were downloaded in SMILES format from PubChem Download Service1 for this purpose. Inorganic part of salts, charges on certain atoms were neutralized, components of mixtures were split and only one chemical from sets of duplicates were kept. Approximately 100,000 chemicals were held out for future needs.

2. Hansen2 and Bursi3 Ames mutagenicity benchmark datasets were used. The data was preprocessed in the same way as the PubChem chemicals, except in the case of duplicates, the compound with the highest mutagenicity value was retained, leaving 6771 compounds (3639 mutagenic and 3132 non-mutagenic).

3. A dataset of 575 compounds was taken from the publication of Ghose et al4 to compute molar refractivity of the fragments.

An in-house dataset5 of 7000 chemicals (Zhu Hao et al) with their experimentally observed LogP was also used to help in computing LogP contribution of the fragments.

1. PubChem Download Service, URL: https://pubchem.ncbi.nlm.nih.gov/pc_fetch/pc_fetch.cgi.

2. Hansen, K., Mika, S., Schroeter, T., Sutter, A., Laak, A. T., Steger-Hartmann, T., Heinrich, N. and Müller, K. R. (2009) Benchmark Data Set for in Silico Prediction of Ames Mutagenicity. J. Chem. Inf. Model. 49, 2077-2081.

3. Kazius, J., McGuire, R., Bursi, R. (2005) Derivation and validation of toxicophores for mutagenicity prediction. J. Med. Chem. 48, 312-320.

4. Ghose, A. K., Crippen, G. M. (1987) Atomic physicochemical parameters for three-dimensional-structure-directed quantitative structure-activity relationships. 2. Modeling dispersive and hydrophobic interactions. J. Chem. Inf. Comput. Sci. 27(1), 21-35.

Zhu, H., Sedykh, A., Chakravarti, S. K., Klopman, G. (2005) A new group contribution approach to the calculation of LogP. Current Computer-Aided Drug Design 1, 3-9.


Chemical fragments
Chemical similarity
Machine learning
Unsupervised training
Distributed representation
Artificial intelligence
Fragment vectors
Information And Computing Sciences

Supplementary weblinks


Comments are not moderated before they are posted, but they can be removed by the site moderators if they are found to be in contravention of our Commenting Policy [opens in a new tab] - please read this policy before you post. Comments should be used for scholarly discussion of the content in question. You can find more information about how to use the commenting feature here [opens in a new tab] .
This site is protected by reCAPTCHA and the Google Privacy Policy [opens in a new tab] and Terms of Service [opens in a new tab] apply.