Predicting cell responses to disease and drugs with AI

Listen to this blog story:

Listen to "Predicting cell responses to disease and drugs with AI" on Spreaker.

Early in Mo's academic career, during his Master’s in Computer Engineering and Artificial Intelligence (AI) at Sharif University of Technology, Iran, he developed an interest in bioinformatics and the biological applications of AI. The infographic below shows the different forms of AI in use today.

“A lot of machine learning (ML) research isn’t really transferable to a real-life setting. During my Master’s, I found it very interesting to design and apply ML methods to real datasets obtained from living organisms, and use these to predict things that happen in the body. It motivated me to do something that would benefit other people – you see a problem and maybe in 20 years you can cure a disease.”

Mo Lotfollahi,
Group Leader, Cellular Genetics programme, Wellcome Sanger Institute

He then embarked on a PhD in Computational Biology at the Technical University of Munich where he developed ML methods to facilitate the analysis of single-cell data. After working at various AI and Biotech companies, Mo has returned to academia as a Group Leader at the Sanger Institute, developing generative AI models for predicting cellular responses to disease, drug treatment and genetic changes. However, as scientific co-founder of AI VIVO, a pharmacology start-up in Cambridge, Mo is still connected to industry.

“Being a part of two different research communities – academia and industry – has given me lots of insight into solving interdisciplinary problems. Industry is very results-driven, focused on translation and real-world application. What academia gives me is the opportunity to explore and try ideas that wouldn’t be tested in industry because they are too risky. I’m excited that the Sanger Institute provides the opportunity to bring these two perspectives together, to do exploratory research with a real-world impact.”

Many diseases do not have clear causes, and drug targets are often hard to predict. This means that researchers need to screen many drugs and test various ways to modify cells to change them from a diseased to a healthy state, which can be costly and time-consuming. Generative AI models that accurately predict how cells may respond to drugs and genetic changes offer an exciting way to address these complications.

One of Mo’s research directions involves ‘foundation models’ – everyday examples of this type of generative AI include the Large Language Models (LLMs) such as ChatGPT, Gemini and Claude. Foundation models are trained on huge datasets across many contexts, enabling them to detect intricate associations and relationships between different data types, and make predictions about data that are missing. The Lotfollahi Group is currently training its foundation models on large-scale data across different cells, tissues, diseases and individuals, which would be too much to process manually.

“We already have proof of concept that we can generate cells in silico (in a computer simulation) under unseen conditions and accurately predict expression profiles, i.e. how each RNA transcript is expressed in a certain cell in the context of a specific disease or genetic intervention. There is a lot of room for improvement, but in a few years these in silico experiments could replace a lot of wet lab work. This would allow experimental scientists to explore more efficient experiments with narrowed hypotheses.”

Mo’s group also uses a technology called ‘Geometric Machine Learning’ to integrate multiple large-scale datasets extracted from patient samples, including single-cell RNA sequencing – examining which genes are expressed by an individual cell, chromatin accessibility – how easily DNA in our cells can be accessed and read, and structural variations in DNA, including gene insertions and deletions. This allows scientists to identify areas of tissue that differ between healthy and diseased tissue.

“The tissue samples used in these experiments are very heterogeneous – they come from different patients, different clinics and have various stages of disease. But AI helps to maintain data integrity, which is the reliability and trustworthiness of the data, ensuring specific disease signals are preserved while accounting for the effects of differences in sample collection.”

With enough data, scientists can create spatial atlases for each disease, which are reference maps showing the spatial distribution and organisation of various biological features. These atlases increase scientific understanding of disease and allow clinicians to make more informed decisions about therapeutics and diagnosis.

Mo's team includes experts from various fields, such as machine learning, computer science, and wet lab sciences. The interdisciplinary nature of the team enables a comprehensive approach to solving complex biological questions.

“My team includes machine learning experts but we also need biologists who want to transition into the field. In order to design the right model, you need a lot of information about the context of the disease and the types of data you will have – biologists bring that wealth of knowledge with them.

“All our models and predictions have to come with experimental validation, which is why I’m excited to join Sanger. These models won’t replace experiments: they work hand in hand.”

Mo has established his lab at an exciting time for AI: awareness and usage have skyrocketed in recent years, and many people are integrating it into their everyday lives.

“These days, even my mum and grandma know what AI is! It’s a very good time to be involved in AI, especially for biology. There has been an increase in investment from government and funding agencies, which provides a lot of momentum for our research. This means we can do larger, more ambitious projects that wouldn’t have been possible a few years ago. I think this is a win-win for everyone and I’m very hopeful for what AI could achieve and bring to society.”

The Lotfollahi Group is using generative AI to investigate if inflammation caused by a wide range of neurodegenerative diseases, including Alzheimer's disease, Parkinson's disease, Multiple sclerosis and others, can be reduced. If so, this could alleviate the symptoms of these diseases, such as disorientation and tremors. The team also works on kidney cancer and liver disease, but Mo stresses that the technologies can be applied widely.

One potential impact of Mo’s work is reducing or even removing the need for animal research. If AI models continue to increase in complexity, they might one day accurately model the whole human body in various states of health and disease. This could be used to test drugs or genetic treatments quickly and effectively, reducing the timescale for the drug approval process, which currently takes an average of 10 years.

As AI continues to increase in power and scale, it has great potential to transform the speed at which new medicines are developed and even how biological research is conducted. Mo and his team’s research into AI leverages the large-scale data generation at the Sanger Institute and complements other collaborators working in the field of generative and synthetic genomics.

Footnote:

For a simple guide to key concepts in AI, see our article: Using artificial intelligence for genomic research on the YourGenome website.