Predictive models for designing drugs – friend or foe?

by Tilly Bingham, Thursday, 3rd November 2022

AI (Artificial Intelligence) concept. Human brain and circuit board

As drug discoverers we are all trying to work out how to use machine learning and AI to our advantage.

What’s the history?

For medicinal chemists, who work at the earliest stage of the drug design process, there is a lot of history in using predictive models based on computationally calculated chemical properties. This goes way back to the 90s and the emergence of Lipinski’s rules1 and has been built on ever since. For example, the more recent MPO rules set out by Wager in 2016 for central nervous system (CNS) desirability2. Both these algorithms have at their core the “accessibility and ease of use” for human minds to engage with and are thus limited to a few simple key parameters which are intuitive for medicinal chemists.

Where are we now?

Nowadays we capture increasingly large experimental datasets on each of the compounds we make and can calculate an ever-larger number of calculated parameters in a short space of time. This has enabled even more complicated models to be developed which predict the “drug-likeness” of a prototype compound before it has even been made.

What’s left to do?

The challenge for medicinal chemists is now – what model to choose? There are so many options.  

Often, an experimentally led process of model selection in the early stages of a project is used. A favored model can arise in one program that does not necessarily translate to the next. However, there is nothing more satisfying than finding a fit. At that point a project can really take off and start to deliver improvements in each cycle of new prototype molecules as the predictive model fit pays off.

A great example of predictive models in action

A nice example of this was recently described by Koester et al at Novartis.3 They describe in the Journal of Medicinal Chemistry their “Discovery of Novel Quinoline-Based Proteasome Inhibitors for Human African Trypanosomiasis (HAT)”. In their project the CNS MPO models did not correlate well with experimental results, but they identified a useful qualitative correlation between the StarDrop model’s predicted LogBB and experimentally measured Kp which “helped rapid SAR generation of compounds in the desired property space”.

Find out more….

To choose the right model you need an awareness of what the options are, in a world where new possibilities are emerging all the time. If you are interested to learn about some of the new models which are available for medicinal chemists, then I encourage you to list to Dr. Tom Pesnot’s webinar on “15-minute Facts: Multi-parameter optimization – friend or foe?” on 30th November.

  1. Lipinski CA, Lombardo F, Dominy BW, Feeney PJ. Experimental and computational approaches to estimate solubility and permeability in drug discovery and development settings. Advanced Drug Delivery Reviews. 46 (1–3): 3–26. doi:10.1016/S0169-409X(00)00129-0
  2. Travis T. Wager*, Xinjun Hou, Patrick R. Verhoest, and Anabella Villalobos, Central Nervous System Multiparameter Optimization Desirability: Application in Drug Discovery, ACS Chem. Neurosci. 2016, 7, 6, 767–775, https://doi.org/10.1021/acschemneuro.6b00029
  3. Dennis C. Koester*et al, Discovery of Novel Quinoline-Based Proteasome Inhibitors for Human African Trypanosomiasis (HAT), Journal of Medicinal Chemistry  2022, 65, 17, 11776-11787, 10.1021/acs.jmedchem.2c00791
Register Now