Approaches using AI in medicinal chemistry

The challenge of pharmacological effect prediction and its relation to analog design consists of the decision of which molecule to make next on the basis of the available data, medicinal chemistry knowledge, experience, and intuition. In the second half of the 1900s century, attempts were made to relate narcotics pharmacology to their physicochemical properties by specifically using distribution and partition coefficients. This was shortly followed by Paul Ehrlich's observation to attribute the pharmacological effect of a compound to a specific functional group. However, it was only in 1971 that this observation was called pharmacophore. About 30 years after Ehrlich, Hammett related the effect of changes in structure on reaction mechanisms, specifically based on resonance interaction of an aromatic ring. This model was later extended by Taft by separating the inductive effects from the steric properties of substituents. This concept was formalized with the work of Hansch, Free, and Wilson which reasoned that the biological activity for a set of analogs could be described by the contributions that substituents or structural elements make to the activity of a parent structure. This led to the analytical description of general quantitative-structure–activity relationship studies (QSAR). If the question “What to synthesize next?” is answered then “How to synthesize it?” follows up. The prediction of chemical reactions starting from educts or the product and educing input reactants and reaction conditions is a fundamental scientific problem. As QSAR computer-aided synthesis planning (CASP) has a long history starting in the 1960s with LHASA a rule-based approach to retrosynthesis planning.