Molecular Flexibility Encodings for Virtual Screening and Machine Learning

Molecular similarity measures that are based on conformation of molecules are strongly affected by these conformations. Consequently, the quality of results of virtual screening or machine learning experiments is a function of the geometrical formation. Based on a fundamental principle of 3D-QSAR that only the descriptors, derived from a biological active conformation, are suitable descriptors, it is beneficial to know the biologically active conformations. The information about these active conformations is normally not available and in-silico generation of 3D coordinates is affected by the force field and parameters. Additionally, minimized conformations usually have different geometric formations than the active structures and, therefore, are doubtful. We developed different approaches that incorporate the conformational flexibility into the similarity calculation to implicit compare the conformational space of the molecules and to reduce the influence of given 3D conformations.
The flexibility extension of the Optimal Assignment Kernel (Jahn, Chemistry Central Journal, 2009; Fechner et al., Journal of Chemical Information and Modeling, 2009; Jahn et al., Journal of Cheminformatics, 2009) compares the local conformations of neihghboring atoms with respect to a core atom (Fig. 1). Therefore, the method decomposes the conformational space of the molecules and computes a similarity value based on the optimal assignment of these decomposed local conformational space environments.

Figure 1: Example of an approximation of the local conformational space environment.

Another approach uses the conformational space of molecules and calculates probabilistic models for the conformational space of the molecules. The conformational space is captured by atom-pair distance profiles of flexible atom-pairs (Fig. 2). A special similarity function compares the models and performs an implicit comparison of the conformational space of two molecules within one similarity calculation (Jahn et al., Molecular Informatics, 2010, Jahn et al., Molecular Informatics, 2011). The complete source code of this approach can be found in the software section on this site (4D FAP).

Figure 2: Example of a flexible atom-pair and the resulting distance profile of the atom-pair in the conformational space.


References

Boltzmann-Enhanced Flexible Atom-Pair Kernel with Dynamic Dimension Reduction
Andreas Jahn, Georg Hinselmann, Lars Rosenbaum, Nikolas Fechner, and Andreas Zell
Molecular Informatics, 2011, Accepted for publication
Probabilistic Modeling of Conformational Space for 3D Machine Learning Approaches
Andreas Jahn, Georg Hinselmann, Nikolas Fechner, Carsten Henneges, and Andreas Zell
Molecular Informatics, 2010, 29 (5), pp 441-455
Abstract, PDF, DOI: 10.1002/minf.201000036
Optimal assignment methods for ligand-based virtual screening
Andreas Jahn, Georg Hinselmann, Nikolas Fechner, and Andreas Zell
Journal of Cheminformatics, 2009, 1:14
Abstract, PDF, DOI: 10.1186/1758-2946-1-14
Incorporating molecular flexibility into three-dimensional structural kernels
Andreas Jahn
Chemistry Central Journal, 2009, 3(Suppl 1):O11.
Abstract, DOI: 10.1186/1752-153X-3-S1-O11
Atomic local neighborhood flexibility incorporation into a structured similarity measure for QSAR
Nikolas Fechner, Andreas Jahn, Georg Hinselmann, and Andreas Zell
in Journal of Chemical Information and Modeling, 2009, 49 (3), pp 549-560
DOI: 10.1021/ci800329r
Chronic Rat Toxicity Prediction of Chemical Compounds using Kernel Machines
Georg Hinselmann, Andreas Jahn, Nikolas Fechner, and Andreas Zell
in Lecture Notes in Computer Science (EvoBIO 2009), 2009, 5483, 25-36
Two-step hierarchical assignments on molecular graphs
Andreas Jahn, Nikolas Fechner, Georg Hinselmann, and Andreas Zell
Chemistry Central Journal, 2009, 3(Suppl 1):P13.
Abstract, DOI: 10.1186/1752-153X-3-S1-P13

Contact

Andreas Jahn, Tel.: (07071) 29-77174, andreas.jahn (at) onlinehome.de

Lars Rosenbaum, Tel.: (07071) 29-77174, lars.rosenbaum (at) uni-tuebingen.de