Announcement: An artificial intelligence to predict chemicals markets

AlchemAI is a neural network developed and trained by R&D Mediation to predict market and uses of chemicals. It can find new markets for existing molecules or business opportunities for new chemicals. In August 2017, 98 markets and uses are covered by the AI.

From structure to application

Chemists and biochemists have imagined more than 8 million molecules. Main part of them has only been manufactured in small quantities and poorly tested. Finding chemical activity from molecular pattern is an important part of Research, especially in health industry. Often, the potential use is deduced from uses of similar chemicals or form traditional uses. Quantum mechanics codes, either semi-empirical like MOPAC [1] or ab initio like GAUSSIAN[2] allow discovering relationship between some properties (catalytic, biocide…) and the molecule. These calculations are long and complex. For some thermodynamical properties, Benson’s groups [3] make it possible to find good approximations, with the limit of the increments values available.

 

A Deep Learning approach

Exemple de réseau de neurone

AlchemAI was developed in another way. A multilayer neural network has been trained to analyse molecular structures and link them to the potential markets and uses. Based on Tensorflow [4], it can recognise 98 markets with a very good reliability on primary and secondary markets.

AlchemAI training

Applications for AlchemAI

AlchemAI is currently a research tool. The training framework may be adapted to the identification of further properties, like toxicity or cost. Beyond research, AlchemAI may also help marketing departments to analyse new markets. For example:

  • Discovering a new market for an existing manufactured molecule
  • Anticipating new market for new molecules from R&D departments
  • Designing molecules for a specific market
  • Screening old chemicals for new uses
  • Checking challengers, suppliers or customers markets in a due diligence approach

References

[1] J. J. P. Stewart, “MOPAC: A semiempirical molecular orbital program,” Journal of Computer-Aided Molecular Design, vol. 4, no. 1, pp. 1–103, Mar. 1990.
[2] P. J. Stephens, F. J. Devlin, C. F. Chabalowski, and M. J. Frisch, “Ab Initio Calculation of Vibrational Absorption and Circular Dichroism Spectra Using Density Functional Force Fields,” The Journal of Physical Chemistry, vol. 98, no. 45, pp. 11623–11627, Nov. 1994.
[3] S. Benson and J. Buss, « Additivity Rules for the Estimation of Molecular Properties. Thermodynamic Properties », The Journal of Chemical Physics, vol. 29, no. 3, pp. 546-572, 1958.
[4] M. Abadi et al., “TensorFlow: Large-Scale Machine Learning on Heterogeneous Distributed Systems,” arXiv:1603.04467 [cs], Mar. 2016.

 

Please cite this document by the DOI of the French version: 10.17601/RD_MEDIATION2017:2

Télécharger cet article au format PDF ou ePub

Laisser un commentaire

Votre adresse de messagerie ne sera pas publiée. Les champs obligatoires sont indiqués avec *