Fixing failed drugs: AI solutions for toxicity in drug discovery – part 2

In Part 2 of our conversation with Layla Hosseini-Gerami, Chief Data Science Officer at Ignota Labs, we explore the many forms of toxicity and how AI-powered tools like omics and cell painting are transforming early prediction in pre-clinical research.

The conversation revealed important distinctions between different types of toxicity and our current ability to predict them:

“For certain types of toxicity like liver and cardio, I think those are quite well understood in terms of the mechanisms underlying them,” says Hosseini-Gerami. So you can guess what you need to be measuring and testing early on – things like BSEP and mitochondrial toxicity and things that you know can manifest as liver toxicity when it comes to an animal.”

For these well-studied toxicities, Hosseini-Gerami explains that the data and knowledge exists to be able to make good predictions. However, many other types of toxicity remain challenging:

“There’s also a long tail of other toxicities like muscle toxicity, eye toxicity, dermal toxicity, kidney toxicity – there’s so many different ways in which something can go wrong. And to be able to get those early signs, you need some sort of proxy or thing that you’re measuring or representing. And I think for some of those, we don’t quite have that knowledge yet.”

Hosseini-Gerami noted the value of established resources for well-understood toxicities: “We have the FDA DILIRank, and now the DICTrank list, which is the cardiotox equivalent. Those datasets, I think, have been so valuable in increasing our understanding and doing our research on liver toxicity and the mechanisms behind it.”

The multiomics revolution in toxicity prediction

Recent advances in experimental technologies and computational methods are creating new opportunities to address these challenges. When asked what has enabled AI to tackle this problem now when it couldn’t 10-20 years ago, Hosseini-Gerami highlighted several key developments:

“What has really unlocked this ability to use AI and data analysis, especially when trying to understand things on a deeper level, is the advent of data, especially omics data. That explosion over the last decade or so has been amazing to see.”

She explained that earlier approaches were limited: “Drug discovery and the machine learning side of it was initially very target focused… understanding of the molecular target or the molecular off target, often based on chemical structure data.”

The critical missing element was understanding downstream effects: “The piece that was missing [was]: OK, we can predict that a drug is hitting a particular target or a particular off target, but what happens next? How can we understand what the consequences of that are?”

Hosseini-Gerami notes that modern multiomics approaches provide the following crucial context:

1. RNA sequencing: “We can measure with RNA sequencing the genes that have gone up or down in expression because of that compound interaction.”
2. Metabolomics: “Bringing in other paradigms of information like metabolomics data, measuring the different metabolites that are generated, again provides another piece of the puzzle.”
3. Cell painting: “A really cool data modality that’s also being used in this multimodal, multiomic picture, [enables us to] see how the cell organelles and compartments have changed in shape and size and look at things like cell death in a very automated way.”

Hosseini-Gerami provided a detailed explanation of how drugs affect cellular systems:

“When a drug interacts with a cell, it’s a cascade of events. The drug binds to the protein, and that will set off a signalling cascade, which will then lead to changes in activity of what are called effector proteins. These are proteins that actually do things – transcription factors that might change the regulation of different genes, proteins that can change the cell morphology.”

Advanced computational methods can also now integrate these diverse data types: “What we can do with that is then ask the question: ‘what signalling changes did that compound induce to lead to those changes in gene expression?’ says Hosseini-Gerami, “and ‘What pathways were modulated by that compound upstream that led to a specific set of transcription factors being modulated, which has led to differential expression of genes?”

Key research centres advancing the field

The conversation highlighted several research groups that are making significant contributions to toxicity prediction:

1. The Broad Institute (USA): “They do a lot of work with cell painting and analysis of omics data, looking at toxicity.” Hosseini-Gerami specifically mentioned the work of Anne Carpenter’s group and her former colleague Srijit, noting a recent presentation that combined “information from cell painting, from gene expression and from chemical structure, and encoding that within a representation of a molecule to use for machine learning.”
2. Saez-Rodriguez Lab (Heidelberg, Germany): “They do a lot of work with systems biology, [which is] really cool, and I think it is the key to getting those mechanistic insights.” She highlighted their work on “cell signalling modelling, looking at predicting protein–protein interactions, and they have a tool to do multiomics causal reasoning across metabolomics, proteomics and gene expression.”

She emphasised the collaborative nature of the field through open-source tools: “They make all the methods publicly available and open source for other people to use and benchmark. That’s always the first step I like to take – go and use someone else’s tool and then see how I can build on that or put different things together and try and get some synergy.”