Improving synthetic efficiency using the computational prediction of biological activity

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Drug Discovery

Background:

• Lead optimization in drug discovery requires efficient synthesis and screening of compound libraries.
• Traditional methods can be time-consuming and resource-intensive.
• Computational tools offer potential for streamlining this process.

Purpose of the Study:

• To develop a computationally driven process for efficient synthesis and screening of compound libraries for lead optimization.
• To improve the efficiency of identifying active compounds by prioritizing synthesis candidates.

Main Methods:

• Generation of virtual compound libraries based on a lead chemical structure.
• Clustering compounds to select a representative training set (less than 1/3 of the library).
• Development of a predictive model using Partial Least Squares (PLS) regression with 1D/2D descriptors.
• Application of the model to predict activities and rank-order compounds in the test set.

Main Results:

• A predictive model was built using 52 out of 169 PDE4 inhibitors.
• Applying the model to the remaining 117 compounds allowed for effective rank ordering.
• Synthesizing only 50% of the library (including the training set) yielded 78% of active compounds (hits).
• Synthesizing 67% of the library yielded 97% of the hits.

Conclusions:

• The developed process enhances efficiency in lead optimization by prioritizing synthesis candidates.
• The method overcomes limitations of traditional 2D descriptors, offering better interpretation and extrapolation.
• Key Quantitative Structure-Activity Relationship (QSAR) assumptions, such as descriptor redundancy impact and the necessity of a high predictive r2 for rank-ordering, were shown to be unnecessary.