The Great Debate: The Utilisation of True AI in All Aspects of Drug Discovery

As a part of Proventa’s Medicinal Chemistry and Biology Strategy Meeting May 2021, we organised a number of interesting keynote sessions pondering the latest innovations, challenges and strategies for the future. The morning keynote, hosted by Gilead’s Executive Director for Structural Chemistry Uli Schmitz, was a particularly fascinating panel discussion focused on the utilisation of AI across the many aspects of drug discovery.

Why can’t we design drugs like we design airplanes?

“Everybody knows an airplane or a car when they see one. Staring at some molecules and saying this is a drug is a whole different ball game”. 

This was an interesting discussion highlighting the complexity of drug discovery. One speaker emphasised how a drug is not an intrinsic property of matter, but a ‘quality’ that humans bestow upon a molecule whether it’s biologic or a small molecule. Building an aeroplane requires basic criteria – “we need wings, an engine” – components that cannot be identified as anything else. With drug discovery, false positives are a common challenge, hence searching for a molecule with desired characteristics is not as easy as finding the components for building a plane.

AI and machine learning

While AI has been utilised across the drug development process, its role in drug discovery specifically remains under scrutiny. One of the questions raised about AI in drug discovery was “how do we evaluate it?”. In other words, how do we evaluate the success of AI applications across all aspects of drug discovery? The answer: Trust and verify. Machine learning has shown significant innovations in drug discovery, target identification and validation, for example. The Generative Adversarial Network (GAN) is one recent innovation in deep learning for drug discovery. Trust the system and then verify it is correct, appeared to be the attitude regarding AI in drug discovery. 

One of the main challenges expressed with regards to drug discovery was finding data that informs us what is truly relevant. “If you’re going to work on neurodegeneration, and most of your datasets come from cancer, you might be going in the wrong direction even if your data sets are correct”. Therefore, it was suggested that there needs to be a method of modulation for the hypothesis and the given datasets. There has to be a way to guide the experiment to give true or false answers with regards to your hypothesis.

It was suggested that computers may overcome the emotional attachment of humans to their projects, which can sometimes cause persistence beyond reason and multiple failures e.g. the amyloid hypothesis of Alzheimer’s. 

Using AI to predict molecules was also suggested as a method of understanding if our calculated methods are working. In other words, use AI in the form of active learning to make sure we’re using the correct methods. “We need a period of refinement to work forward so in other words, use AI to predict molecules and then see if the underlying assumptions are correct in terms of the calculated methods that we use. So I don’t see a direct benefit of AI, more just an introspective refining of how it could be used”.

The two main advantages of AI were highlighted as the accurate tracking of what hasn’t been made and what’s very similar or contiguous to things which have been made.

It also appears there is a misconception in the community about the volume of data input for AI and the desired outcome. “AI works when you feed it high quality data, (but) there’s a misconception in the community, that if you feed the computer more data, it will be able to sift through the data and find what’s relevant to the question.” This of course is not always the case – big data does not necessarily equal true data, and a computer or mathematical model cannot distinguish truth from falsehood.

“It’s still difficult to select targets that are disease relevant” – this is the main Achilles heel in the drug discovery process with the possibility that new therapeutic modalities are coming about which might require a completely different set of computational tools. “So whether we’re talking about small RNAs for gene therapy, CRISPR, vaccines (etc)”. 

ML could be used to create a model that can criticise drug discovery experiments i.e. does this method suggest this molecule? Using AI to gain insight into drug discovery as well as validating and improving it appears to be the direction for the future. 

Change in expectations

In comparison to research decades ago, medicinal chemists are expecting a lot more from the area of AI. “We now routinely take predictions from various computer models without question. There is an acceptance that computers are here to stay and that they are a partner in drug discovery”. This is a positive trend that has improved the hybrid model: this refers to the act of humans guiding a computer and taking feedback and interacting with the system in a direct manner.

Alexa for example demonstrates how our interaction with computers has changed and improved. “Self driving cars can use the information from popular travel routes, signs, traffic lights etc… to create a framework. In drug discovery however, you’re missing a lot of data and it’s very fragmented”. This interesting point questioned to what extent AI could innovate drug discovery, leading to a number of question to be answered:

“Why don’t we think more about building physics-based methods into AI?

“Why don’t we think about putting chemistry knowledge into AI?

“Why don’t we take the expert knowledge that we have such as Tesla learning from people driving, so why don’t we take the (molecular) dockings, the alignments and then figure out a way to get that into the eye – to me that is the way to move forward.”

Charlotte Di Salvo, Former Editor & Chief Medical Writer
PharmaFEATURES