From SAR Diagnostics to Compound Design: Development Chronology of the Compound Optimization Monitor (COMO) Method

Abstract

In medicinal chemistry, compound optimization largely depends on chemical knowledge, experience, and intuition, and progress in hit-to-lead and lead optimization projects is difficult to estimate. Accordingly, approaches are sought after that aid in assessing the odds of success with an optimization project and making decisions whether to continue or discontinue work on an analog series at a given stage. However, currently there are only very few approaches available that are capable of providing decision support. We introduce a computational methodology designed to combine the assessment of chemical saturation of analog series and structure-activity relationship (SAR) progression. The current endpoint of these development efforts, the compound optimization monitor (COMO), further extends lead optimization diagnostics to compound design and activity prediction. Hence, COMO plays dual role in supporting lead optimization campaigns.

DeepCOMO: from structure-activity relationship diagnostics to generative molecular design using the compound optimization monitor methodology

“…For large ASs, we generally attempt to build global support vector machine regression (SVR) [ 33 ] and linear ridge regression [ 34 ] models to prioritize VAs. In addition, for all ASs, local FW predictions are attempted, which are supported by the generation of FW VAs for a given AS [ 5 , 6 ], as described above.…”

“…COMO combines different scoring schemes including chemical saturation, multi-property, SAR progression, and SAR heterogeneity scores [3][4][5][6]. For AS diagnostics, the chemical saturation (S score) and SAR progression (P score) are primary measures for assigning ASs to different development stages.…”

“…Thus, although COMO was originally devised as a diagnostic/descriptive tool it also had the intrinsic potential to bridge between LO analysis and compound design. Accordingly, different analog design strategies and activity prediction approaches have been implemented in COMO to design and prioritize VAs [ 5 , 6 ].…”

“…In this context, we also describe key components of the DeepCOMO methodology. Furthermore, we present an application of DeepCOMO on two exemplary ASs, illustrating its entire analysis and design spectrum, as it has evolved since its inception [ 6 ]. Here, emphasis is put on the compound design aspect applying the DeepCOMO framework.…”

Automatic Identification of Analogue Series from Large Compound Data Sets: Methods and Applications

“…An alternative scheme that uses SAR matrices for generating virtual compounds was explored in Reference [ 78 ]. This work resulted in the development of the COMO [ 79 , 80 ] and DeepCOMO [ 81 ] method, which augments the generation of virtual analogues utilizing transfer learning on recurrent neural networks.…”