Scroll for more
blog Apr 3, 2018

Love drug discovery? You just need the right chemistry!

Author: Ian Churcher, VP Drug Discovery at BenevolentAI

There is an increasing number of varieties of drug agents in clinical use ranging from antibodies and proteins to nucleic acids and, increasingly, cellular and genetic therapies but the majority of drugs on the market and in development today are still small, synthetic molecules made in a chemistry laboratory.

Small molecules have a number of advantages including being able to access drug targets all around the body and within cells as well as being cheaper to produce, stable on long term storage and being generally amenable to a wide range of dosing routes including, in contrast to many other agents, oral delivery.

Small molecule drugs are not perfect of course and the uncertainties of developing them have been widely documented including a risk that some may show unexpected effects or toxicity, some drug targets stubbornly resist allowing a good molecule against them to be designed and all molecules have a finite patent lifetime in which to use them.

To make sure we have the best small molecule to give the best chance of a positive clinical outcome, chemists need to design a molecule with the right properties but, once they’ve done this, they need to go into the chemistry lab and actually make it. Making it is frequently not as easy as we’d like it to be and some molecules can be extremely difficult to make.

We’re now using AI trained on huge volumes of chemical reaction data much more to suggest the best synthetic route to make a molecule (see BenevolentAI’s senior machine learning researcher Marwin Segler’s Nature paper here) but, in fast moving drug discovery projects, if a molecule can’t be made quickly, it is often passed over in favour of other molecules.  Consequently, some excellent molecules which could have made great drugs go unmade and untested. To overcome this problem, we need new chemical reactions that provide new ways to make molecules and this is an area where academia and drug discovery scientists can work together to define the most pressing areas of need and then really focus on solving them.

A group of senior drug discovery scientists, including BenevolentAI’s Head of Research, Ian Churcher, recently published a Perspective article in Nature Chemistry (link) discussing the topic. Some of the reasons why chemistry to make drug molecules can be so tricky is that these molecules tend to have a wide range of different chemical functional groups which are assembled to give just the right combination of desired properties including high activity and selectivity for the desired protein target as well as the right pharmacokinetic profile (i.e. they are designed to be absorbed, reach the desired site of action in the body and then are removed again).

Putting all these different chemical groups together in a single molecule then creates real challenges in making them. To get chemical reactions to work on the right part of the molecule, and not on other parts in unwanted ways (the ‘chemoselectivity problem’) is often tough, frequently stretching current synthetic reactions to breaking point and meaning some molecules just cannot be made in a suitable process.

This is an area that BenevolentAI is looking at carefully. We use AI tools to tell us the best biological drug targets to work on and then we use other AI tools for helping design the best molecules as potential drugs. We then use a different set of tools again to suggest how to use modern synthetic chemistry to make these molecules but, in some cases, the state of the art chemical reactions can only provide slow or inefficient ways to make them which has the potential to slow down our drug discovery projects so we’ll be looking at ways to help find the new reactions which will help all drug discovery chemists make the molecules they want to faster and more reliably every time.

 

Ian Churcher, VP Drug Discovery at BenevolentAI


 

More researches

Exploring deep recurrent models with reinforcement learning for molecule design

Planning chemical syntheses with deep neural networks and symbolic AI