Knee Deep in the Hoopla
David Clark, PhD

Knee Deep in the Hoopla

An update on the use of artificial intelligence in drug discovery

I have a confession to make: amongst my Christmas presents was a CD containing the greatest hits of the ‘80s US rock band, Starship. And so, I have been listening to “We Built this City on Rock and Roll”, which has been described as “the worst song of all time”. And enjoying it. The song contains the line “Knee deep in the hoopla” which is defined by the Urban Dictionary as “being very involved, mixed up in the excitement or fervor of the moment”.

Hold that thought.

It’s been almost impossible to avoid the hype around artificial intelligence in drug discovery over the last year or two (indeed, it was one of Eureka’s hot topics for 2017). In fact, at the recent JP Morgan Healthcare conference, the topic was so front and center that it became almost nauseating, according to one reporter: “If I hear ‘AI’ again, I’m going to barf”!

Putting these two things together, it’s not surprising that my eye was caught by the title of a recent article, “Sifting Nuggets of Truth out of the AI Hooplah”, which does a nice job of trying to assess where we are really at in terms of the useful application of artificial intelligence to drug discovery and healthcare more generally. Some related discussion has cropped up on Derek Lowe’s popular “In the Pipeline” blog in the last week or so.

Moreover, just last week, a scientific paper was published that does provide a concrete instance of the successful application of AI to a drug discovery project. Prof. Gisbert Schneider and his co-workers at the Swiss Federal Institute of Technology in Zurich describe how they used what they term “generative artificial intelligence” to design drug-like chemical compounds with desired biological activities.

First, a generic computational model of “drug-likeness” was built by a neural network learning from the chemical structures of over half a million bioactive compounds extracted from the ChEMBL database. Then, this model was fine-tuned by using the structures of 25 compounds that exhibit the particular biological activities of interest – in this case, agonistic activity at two nuclear receptors: retinoid X receptors (RXRs) and/or peroxisome proliferator-activated receptors (PPARs).

Starting from a “privileged fragment” (in this case, a carboxylic acid, which is a common feature of known RXR/PPAR agonists), the neural network was used to generate 1,000 candidate compounds, which were demonstrated to lie in the same chemical space as the 25 training compounds, but to have unique chemical structures. Using a series of computational filters, the set of 1000 was then whittled down to just five compounds to be synthesized in the lab.

The five compounds proved to be readily synthesizable using between two and four chemical reaction steps and, when tested in functional assays measuring RXR (α, β and γ sub-types) and PPAR (α, γ and δ sub-types) agonism, four of the five compounds demonstrated compounds demonstrated interesting levels of activity and selectivity profiles (see Table below, IA=inactive.)



Such activity levels are well in line with what might be anticipated from a virtual or high-throughput screen, for instance, and certainly suggest that the neural network has promise as a tool for hit identification.

What is particularly noteworthy, in the context of de novo molecular design, is that the neural network has produced compounds that are synthetically tractable without being specifically trained to do so. This suggests that it has learned something about this important facet implicitly from the training structures with which it was provided. Historically, lack of synthetic tractability has been an Achilles’ heel of many de novo design approaches.

Of course, the presented example is only one instance and it will be important for the researchers to prove that the network can be trained to generate readily synthesized hits against other protein classes of interest to drug discoverers, e.g., kinases, G-protein-coupled receptors, proteases and epigenetic targets. Nonetheless, this is certainly a promising start and it’s good to see some hope emerging from the “hoopla” surrounding AI in drug discovery.