Chemistry Reference
In-Depth Information
Starting with peptide-mimetics, navigation in chemical space has evolved
considerably into procedures including Markov chains, product enumeration,
deterministic combinatorial approaches, stochastic sampling by evolutionary
algorithms, simulated annealing, particle swarms,
etc
….). The field has also
benefited from parallel developments in computer science and engineering
yielding sophisticated tools. These include visualization of multi-objective
compound optimization progress, online structure-activity landscape modeling
and observation of convergence and design runs.
One of the first structure-based
de novo
design studies was published in 1976
[81]. Early methods relied on static X-ray spectroscopy that imposed
pharmacophoric and structural constraints. Ligand and target flexibility, for
example, were not favored. In 1995, the computer designed carbonic anhydrase
inhibitor dorzolamide drug was first introduced leading a number of following
successful cases [41].
Current design tools allow for molecular flexibility with a corresponding increase
in CPU requirements. From a technical point of view GPU computing, cloud
computing and other distributed hardware solutions will sustain progress in this
research area. It is also noteworthy that chemical understanding is essential for
computer-based tools that should incorporate as much medicinal chemistry
knowledge as possible in the quest for drug discovery.
One of the great remaining challenges is activity and polypharmacology
predictions. We are also still far from being able to make reliable estimates of
entropic contributions to ligand-receptor complex formation. Innovative concept
and approaches are required. There has also been complementarity between
scoring concepts and machine learning models. Kernel-based regression models
and artificial neural networks have the adaptability and speed for calculation of
new data without need for explicit energy computation. The machine-learning
paradigm may offer temporary solutions,
i.e.
target-specific and knowledge-based
models. Chemical similarity is used already in scaffold-hopping. Bioisosteric
replacements are used in solving scoring problems in contrast to global energy
computation.
