Search for content and authors
 

QSPR for Ionic Liquids by Recursive Neural Networks

Riccardo Bini 1Cinzia Chiappe 1Celia Duce 2Alessio Micheli 3Antonina Starita 3Maria R. Tiné 2

1. Dipartimento di Chimica Bioorganica e Biofarmacia, via Bonanno 33, Pisa 56126, Italy
2. Dipartimento di Chimica, Via Risorgimento 35, Pisa 56126, Italy
3. Dipartimento di Informatica, Largo B. Pontecorvo, Pisa 56127, Italy

Abstract

Ionic liquids (ILs) are innovative solvents for organic synthesis. They are low-melting salts (m.p.<100°C) obtained by the combination of large organic cations with inorganic anions. The large number of conceivable ILs (est. >1014) doesn't make practically feasible to synthesize every of them and to investigate their properties, so, it would be really helpful if one could correlate the physical-chemical properties of already synthesized ILs with their molecular structure.

Recently, a new QSAR/QSPR method based on Neural Networks for structures, i. e. Recursive Neural Networks (RecNNs), has been introduced for the prediction of molecular properties. This model has been successfully applied to the prediction of the pharmacological activity of a series of substituted benzodiazepines [1] and of the physical-chemical properties of molecules [2] and polymers [3]. The RecNN deals with prediction tasks for compounds that can be represented in a structured domain. The network learn directly from the molecular structures, combining the flexibility and general advantages of neural network models with the representational power of structured domains. This approach overcomes the common difficulties and limitation deriving from the traditional representation with molecular descriptors. .

In the present work, a RecNN model has been applied to the analysis of the melting point of 126 substituted pyridinium bromides. The molecules have been represented as tree structures by selecting a limited set of constituent atomic groups and representation rules. Different representations are discussed. The descriptive and predictive abilities of our RecNN model have been tested and compared with those of traditional multiparameter descriptor approaches [4]. Preliminary results show a very good learning capacity of the networks and a promising prediction capacity.

[1] a) A. Micheli, A. Sperduti, A. Starita, A. M. Bianucci, J. Chem. Inf. Comput. Sci. 2001, 41, 202.

[2] L. Bernazzani, C. Duce, A. Micheli, V. Mollica, A. Sperduti, A. Starita, M. R. Tiné, TR-04-16, http://techrep.di.unipi.it/TR/files/TR-04-16.ps.gz, Dip. di Informatica, University of Pisa: Pisa 2004.

[3] C. Duce, A. Micheli, R. Solaro, A.Starita, M. R. Tiné, Macromol. Symp., in press.

[4] a) A. R. Katritzky et al., J. Chem. Inf. Comput. Sci. 2002, 42, 71. b) G. Carrera et al., Green Chem. 2005, 7, 20.

 

Legal notice
  • Legal notice:
 

Presentation: poster at 18th Conference on Physical Organic Chemistry, Posters, by Riccardo Bini
See On-line Journal of 18th Conference on Physical Organic Chemistry

Submitted: 2006-05-31 15:10
Revised:   2009-06-07 00:44