How algorithms containing neural networks are helping scientists develop entirely novel compounds. Part three in our series on AI in Drug Discovery.

Gisbert Schneider is a professor of Computer-Assisted Drug Design at the Institute of Pharmaceutical Sciences in the Department of Chemistry and Applied Sciences, ETH Zurich. His focus has been the development and application of adaptive intelligent systems for molecular de novo design and drug discovery. Last year, he published results of a study and recently a perspective that described how his lab used “generative artificial intelligence”—that is generating new lead compounds in silico—to design drug-like chemical compounds with desired biological activities. Eureka connected with Dr. Schneider for its multi-part series on AI in drug discovery. Here are his emailed responses.

Eureka: Your lab has used “generative” AI to design drug-like compounds with desired biological activities. In laymen’s terms, how did you do it?

GS: Automated molecular design aims to make helpful suggestions about what to make next. We have developed software that assists a medicinal chemist (the “drug designer”) in this difficult task. The difficulties arise from the fact that there is an astronomical number of possibilities to choose from. Modern machine learning methods are very fast and can consider several design goals in parallel. So, our drug design software was trained to recognize important features and characteristics of known drugs. The obtained models were then used to automatically assemble new molecules with these learned desired properties from scratch.

Eureka: How long did it take your lab to get to this point? There must have been a lot of misses along the way.

GS: We started this whole field approximately 25 years ago. These were humble beginnings but the important concepts were established already in the 1990s. It goes without saying that there have been many failures and we had to endure and rebut harsh criticism. The most important breakthrough happened in the year 2000 with the first successful synthesis of an entirely de novo computer-generated, pharmacologically active new compound. The awareness for AI-based molecular design has steadily increased ever since, although we have avoided to use the term “AI” until recently. Today, de novo drug design is mainstream technology.

Eureka: So could we fairly call these AI tools nonhuman chemists? Or is that too much of a stretch?

GS: Synthesis planning and chemical reactivity prediction, and the de novo design of molecules with desired properties are examples of notoriously hard problems which might be addressable by a chemistry-savvy AI. Beyond a certain level, the human mind struggles to retain and process such complex networks of variables. However, these AI systems do not replace a chemist in this regard. Chemists can actually benefit from this collaborative human and machine intelligence. AI supports decision making and suggest molecules one might not have been considered otherwise.

Eureka: There is a lot of hype around what AI can do to accelerate drug discovery. In your estimation what can it do and what can’t it do?

GS: We should be aware of the limitations of AI when it comes to modelling human cognition and avoid repeating the mistakes of the past. Therefore, it would be wise not to place all one’s eggs in the AI basket, but to expect successful, creative solutions from the combined, collaborative efforts of human experts, process automation, and advanced computer-assisted decision making. The decisive factors for the success of AI in drug design will be the ethos, attitude, and willingness of chemists to apply these computational models in their own research projects.

Eureka: What do you think are the biggest challenges in using AI to help discover drugs? 

GS: There are several challenges we have to overcome. For example, advanced machine learning requires large well-annotated datasets that need to be compiled or generated. Also, the chemical structure of a drug alone rarely accounts for the observed pharmacological effect in a simple fashion. Most drugs have multiple biological targets and activities, and their relative importance is highly dependent on the individual genetic profile of patients, and a range of other factors. In certain areas of drug design we are confronted with inherently ill-posed problems owing to a multitude of often unknown contributing factors.

Eureka: What do you think will be the “Next Big Thing” in the application of AI in drug discovery?

GS: I expect fully autonomous laboratories iterate through the design-make-test-analyze cycle of drug discovery without direct human intervention. The result could be the delivery of “better starting points for drug discovery faster”.

Eureka: Robots are ubiquitous in entertainment. Who is your favorite robot?

GS: I have a clear favorite: “Robby the Robot” from the movie Forbidden Planet (1956).

Thanks for tuning in. Our next Q&A in this series on AI in Drug Discovery will be with Dr. Anne Carpenter, PhD, an Institute Scientist at the Broad Institute of Harvard and MIT. Dr. Carpenter is a pioneer in image-based profiling, the extraction of rich, unbiased information from images for a number of important applications in drug discovery and functional genomics. Her research group develops algorithms and strategies for large-scale experiments involving images.You can follow our series here.