Johnny Gasteiger: Self-organizing neural networks in chemistry

Johnny GasteigerJohnny Gasteiger of the University of Erlangen-Nürnberg is skeptical about deep neural networks: they are good for getting funding, but they are yet to be proven. Johnny illustrated some of the useful applications of shallow neural networks. Much like the human brain generates two-dimensional sensory maps of the environment, a Kohonen network (a self-organizing map) can generate two-dimensional maps of high-dimensional chemical data. Crucial for the success of the study of chemical problems by a self-organizing neural network is the representation of the chemical data.

The shape and surface of molecules are very important: the entire electrostatic potential can be seen in a colored 3D model. Johnny has projected the 3D Cartesian coordinates of, for example, 2-chloro-4-hydroxy-2-methylbutane onto a Kohonen net to get a 2D map:


The neurotransmitter acetylcholine binds to two types of receptors, the muscarinic and the nicotinic receptor. Kohonen maps of the van der Waals surface of muscarinic agonists (muscarine, atropine, scopolamine, pilocarpine) and nicotinic agonists (nicotine, (+)-anatoxin a, mecamylamine, pempidine) have also been produced by projecting points of the 3D surface on a 2D space.29 Such maps allowed the total molecular electrostatic potential (MEP) of a compound to be represented in a single picture, instead of requiring a series of pictures as formerly. Johnny showed the maps of the MEPs of the eight compounds with muscarinic agonists in the top row and nicotinic agonists below.


The results showed that the MEP is important for the binding of these compounds to their receptors. The Kohonen maps reflect significant characteristics of the MEPs and can therefore be used in the search for biologically active compounds.

In analytical chemistry, neural networks have been used in the classification of Italian olive oils.30 The classification was performed on a set of 572 Italian olive oils, from nine different regions, on the basis of an analysis of eight fatty acids. Kohonen learning was superior to a network using the back-propagation of errors. There were 250 oils in the training set and 322 in the test set; 312 of the 322 were correctly predicted. The nine Italian regions were nicely differentiated in the Kohonen map. What is, however, even more interesting is that the Kohonen map is reflecting the map of Italy. This emphasizes the power of unsupervised learning, discovering information that is hidden in the data. In this case, clearly, the different climates and the different soils are responsible for the separation of the regions of Italy in the self-organizing map:

Kohonen Map

Kohonen networks use unsupervised learning. Johnny next discussed examples of supervised learning. In one experiment the electronic properties located on the atoms of a molecule such as partial atomic charge, and electronegativity and polarizability values were encoded by an autocorrelation vector accounting for the constitution of a molecule.31 Using the 49-dimensional vector of seven properties and seven distances, it is possible to distinguish between 112 dopamine agonists and 60 benzodiazepine receptor agonists even after projection into a Kohonen map. The two types of compounds can still be distinguished if they are buried in a dataset of 8,323 compounds of a chemical supplier catalog comprising a wide structural variety. The method can be used for searching for structural similarity, and, in particular, for finding new lead structures with biological activity.

Gasteiger’s team has also worked on simulation of infrared spectra.32 They developed an empirical approach to the modeling of the relationships between the 3D structure of a molecule and its IR spectrum based on a novel 3D structure representation, and a counterpropagation (CPG) neural network. The 3D coordinates of the atoms of a molecule are transformed into a structure code that has a fixed number of descriptors irrespective of the size of a molecule. The structure coding technique is referred to as radial distribution function (RDF) code.33 3D structures were transformed into radial codes (128 values) and put into a CPG network. IR spectra (128 absorbance values) were also input, and the network was trained. When IR spectra are simulated the fingerprint region is predicted well because of the representation of the 3D structure. A CPG network can be operated in reverse mode,33 enabling the prediction of a structure code. The input of a query infrared spectrum into a trained CPG network provides a structure code vector, which represents the radial distribution function with 128 discrete values. This RDF code is then decoded to provide the Cartesian coordinates of a 3D structure.

Johnny concluded by mentioning his recent collaboration with David Winkler on dye solubility in carbon dioxide.34 David has also worked on melting points of ionic liquids, fibrinogen adsorption to polymeric surfaces, and normalized metabolic activity of polymeric biomaterials. Johnny encouraged David to continue to do good science.


  1. Gasteiger, J.; Li, X. Representation of the electrostatic potentials of muscarinic and nicotinic agonists with artificial neuronal nets. Angew. Chem., Int. Ed. Engl. 1994, 33 (6), 643-646.
  2. Zupan, J.; Novic, M.; Li, X.; Gasteiger, J. Classification of multicomponent analytical data of olive oils using different neural networks. Anal. Chim. Acta 1994, 292 (3), 219-34.
  3. Bauknecht, H.; Zell, A.; Bayer, H.; Levi, P.; Wagener, M.; Sadowski, J.; Gasteiger, J. Locating Biologically Active Compounds in Medium-Sized Heterogeneous Datasets by Topological Autocorrelation Vectors: Dopamine and Benzodiazepine Agonists. J. Chem. Inf. Comput. Sci. 1996, 36 (6), 1205-1213.Schuur, J.; Gasteiger, J. Infrared Spectra Simulation of Substituted Benzene Derivatives on the Basis of a 3D Structure Representation. Anal. Chem. 1997, 69 (13), 2398-2405.
  4. (Hemmer, M. C.; Steinhauer, V.; Gasteiger, J. Deriving the 3D structure of organic molecules from their infrared spectra. Vib. Spectrosc. 1999, 19 (1), 151-164.
  5. Tarasova, A.; Burden, F.; Gasteiger, J.; Winkler, D. A. Robust modelling of solubility in supercritical carbon dioxide using Bayesian methods. J. Mol. Graphics Modell. 2010, 28 (7), 593-597.