Cell & Systems Biology (CSB) Professor Alan Moses is using a breakthrough innovation in AI, AlphaFold2, for new directions in both teaching and research.

AlphaFold2 is an astonishing development in deep learning that can predict the structure of almost any protein based on its sequence alone. This monumental task that has been pursued by scientists for over five decades was finally made possible using an AI developed by Google’s DeepMind in 2021.

University of Toronto undergraduates have witnessed this breakthrough occurring in real time and were excited to explore its uses through Moses’ advanced genome biology and bioinformatics seminar CSB471. “Students were apparently so interested in the topic that for the first few weeks, we didn’t have enough chairs in the room,” says Moses. “Students had to grab extra chairs from other nearby classes.”

The AI revolution in protein structure prediction is having unexpected consequences outside of the classroom too. Research in Moses’s group investigates protein fragments called intrinsically disordered regions (IDRs), which connect and terminate well-defined, folded protein structures but can have variable shapes themselves.

IDRs play important roles in cellular processes such as transcription, translation, and signaling pathways, but most methods used for determining protein structure experimentally, such as X-ray crystallography, are unsuitable for IDRs due to their dynamic nature.

Eager to use computational approaches to answer basic biological questions, Moses and colleagues found that, unexpectedly, AlphaFold2 could identify large numbers of IDRs that appear to take on stable structures in certain situations, which they refer to as “conditionally folded IDRs”. These IDRs can adopt structures conditionally (upon binding or interacting with another protein, for instance)

Their findings were recently published in a paper in the journal Proceedings of the National Academy of Sciences titled “Systematic identification of conditionally folded intrinsically disordered regions by AlphaFold2“.

Since AlphaFold2 has predicted structures for hundreds of millions of proteins, Moses’s team were able to compare IDRs in different lifeforms across the tree of life. They made the fascinating observation that human and animal IDRs are less likely to “conditionally fold” than bacterial IDRs.

Mutations in intrinsically disordered regions have been linked to conditions including autism, cancer and amyotrophic lateral sclerosis (ALS). Using AlphaFold2, Moses’s team found that mutations in the “conditionally folded” IDRs are nearly five times more likely to cause disease than mutations in structureless IDRs.

AlphaFold and other related AI approaches to understanding the effects of mutations in proteins was also one of the topics covered in the seminar course on AlphaFold and its implications.

The potential uses of AI to treat disease (as well as other practical applications) were explored by Moses’ undergraduate students by proposing business plans, inspired by Toronto’s biotechnology companies centered around AI and deep-learning technologies.

According to Moses “many of the students’ start-up ideas could be turned into real companies. It’s hard to overstate the impact AI is having on the biotech industry right now.” Next year’s students can anticipate exploration of more foundational discoveries.