453
13 Continuous Molecular Fields Approach Applied to Structure-Activity Modeling
13.4.6 Further Extension of the Approach to Encompass
Biological Macromolecules and Their Interactions
with Both Macro- and Small Molecules
One of the most promising ways to develop further the CMF approach is its further
extension to the description of the binding sites of biomolecules and their interactions with ligands. This seems to be feasible, because molecular fields of biological
targets are identical by their nature to those of small ligands. Because of this, many
of the approaches and methods originally developed for working with small molecules can be transferred to biological macromolecules.
One can suggest several ways of conducting research in this direction. First,
kernels for comparing molecular fields of biological targets (or their binding sites)
could be constructed in exactly the same way as it has been done for small molecules and described in this paper. By combining them with various kernel machine
learning methods, various regression, classification, novelty detection, ranking and
dimensionality reductions tasks could be formulated and solved for biological macromolecules and their binding sites. This could lead to promising applications in
biology-related sciences.
Second, kernels for protein-ligand pairs can be constructed by combining kernels
for small molecules (ligands) and kernels for macromolecules (proteins), as it was
done by Erhan et al. [68], Faulon et al. [69], Jacob and Vert [70], and Bajorath et al.
[71]. The main advantage of using continuous molecular fields in this case is that
this approach can be applied consistently to construct kernels for both small organic and big biological macromolecules using the same typed of molecular fields.
Thanks to this, the same data analysis methods could be used to describe also protein–peptide, protein-protein, protein-DNA, protein-RNA interactions, as well as
properties of peptides and proteins with non-standard residues.
Third, kernels for protein-ligand interactions can be constructed in the frame
of the CMF approach by encapsulating products of molecular fields of ligand and
protein into kernels. Such combined kernels could easily be used in conjunction
with various kernel machine learning methods to solve different tasks relating to
protein-ligand interactions.
It is expected that the results obtained in this direction might be useful in the
field of chemogenomics for target profiling, in bioinformatics and proteomics for
classifying and annotating protein macromolecules and their binding sites by considering and comparing their molecular fields, and in system biology for predicting
interaction graphs for biomolecules, and in drug design.
13.4.7 Integration of Approaches to Prevent the Model Selection
Bias
As it has already been discussed, one of the drawbacks of the CMF approach in its
present state is the danger of over-fitting because of the model selection bias [72].
13 Continuous Molecular Fields Approach Applied to Structure-Activity Modeling
13.4.6 Further Extension of the Approach to Encompass
Biological Macromolecules and Their Interactions
with Both Macro- and Small Molecules
One of the most promising ways to develop further the CMF approach is its further
extension to the description of the binding sites of biomolecules and their interactions with ligands. This seems to be feasible, because molecular fields of biological
targets are identical by their nature to those of small ligands. Because of this, many
of the approaches and methods originally developed for working with small molecules can be transferred to biological macromolecules.
One can suggest several ways of conducting research in this direction. First,
kernels for comparing molecular fields of biological targets (or their binding sites)
could be constructed in exactly the same way as it has been done for small molecules and described in this paper. By combining them with various kernel machine
learning methods, various regression, classification, novelty detection, ranking and
dimensionality reductions tasks could be formulated and solved for biological macromolecules and their binding sites. This could lead to promising applications in
biology-related sciences.
Second, kernels for protein-ligand pairs can be constructed by combining kernels
for small molecules (ligands) and kernels for macromolecules (proteins), as it was
done by Erhan et al. [68], Faulon et al. [69], Jacob and Vert [70], and Bajorath et al.
[71]. The main advantage of using continuous molecular fields in this case is that
this approach can be applied consistently to construct kernels for both small organic and big biological macromolecules using the same typed of molecular fields.
Thanks to this, the same data analysis methods could be used to describe also protein–peptide, protein-protein, protein-DNA, protein-RNA interactions, as well as
properties of peptides and proteins with non-standard residues.
Third, kernels for protein-ligand interactions can be constructed in the frame
of the CMF approach by encapsulating products of molecular fields of ligand and
protein into kernels. Such combined kernels could easily be used in conjunction
with various kernel machine learning methods to solve different tasks relating to
protein-ligand interactions.
It is expected that the results obtained in this direction might be useful in the
field of chemogenomics for target profiling, in bioinformatics and proteomics for
classifying and annotating protein macromolecules and their binding sites by considering and comparing their molecular fields, and in system biology for predicting
interaction graphs for biomolecules, and in drug design.
13.4.7 Integration of Approaches to Prevent the Model Selection
Bias
As it has already been discussed, one of the drawbacks of the CMF approach in its
present state is the danger of over-fitting because of the model selection bias [72].