A Leap Forward in Bioscience

Breakthroughs in the sciences have been powered by tools that advance our ability to see and understand. In biology, great inflections followed the advent of the optical microscope of Antonie van Leeuwenhoek, the electron microscope, x-ray crystallography, gene sequencing, and high-dimensional microarrays. I’m excited about directions forward with the rise of the computational microscope. A leap forward in bioscience, reported in Science this month, gives us a glimpse into the possibilities.

The late-breaking article by Humphreys et al. marks a breakthrough in our abilities to peer into the foundations of cellular functioning. Protein interactions and the complexes formed when proteins come together are the engines of biological systems. Despite intensive studies on protein interactions, much remains unknown about protein interactions and complexes.

Developing more comprehensive understandings about protein interactions would help to decode multiple mysteries of biology and accelerate the creation of novel therapeutics and cures. The new work shows how AI pipelines can illuminate the dark matter of the thousands of interactions among proteins and the complexes they form, shining light on the protein interactome of cells.?

The project on identifying protein complexes is based at the Institute for Protein Design (IPD) at the University of Washington (UW), and has been nurtured by an exciting collaboration between UW and Microsoft. Microsoft has provided IPD with large-scale computing resources on Azure and assistance with deep-learning engineering, and research teams across the groups have been sharing insights, directions, and data. We have been learning together about the challenging workloads and aspirations for computing in the biosciences. Our support to date has included efforts to evolve Rosetta, UW’s protein modeling software suite, into RoseTTAFold , and the more recent effort on identifying protein complexes, representing a shift of focus of attention from identifying protein structure to understanding protein function.

Breakthrough in identifying protein complexes

In the results reported in Science, hundreds of previously unknown protein complexes were identified in one of the simplest eukaryotes, the unicellular Saccharomyces cerevisiae (aka yeast). The newly discovered complexes have been linked to a breadth of fundamental processes in eukaryotic cells, including repairing damage to DNA, doing metabolism, translating RNA into proteins inside ribosomes, and transporting molecules through cell membranes. Some of the complexes are employed for tugging chromosomes apart during cell reproduction, playing a key role in the prominent dynamics of the alignment and motion of chromosomes during meiosis and mitosis.?

Decoding Protein Complexes. Deep learning methods have been used to identify likely protein complexes in eukaryotic cells. The complexes have been linked to processes of transcription, translation, DNA repair, mitosis and meiosis, metabolism, and protein transport within cells and across membranes. The dark blue lines indicate likely points of contact predicted between the proteins. The function of some of the identified complexes are mysteries (complexes drawn from Humphreys et al. (2021)).

I selected several examples of the inferred protein complexes to share in the figure above. The inferences include the structure and configuration of interacting proteins, as well as predictions about the contact points (dark blue lines) between proteins participating in the complexes. Beyond pairs, the team identified complexes of three and four interacting proteins. It is not hard to imagine how the views of the protein complexes enabled by AI-powered optics can provide insights about potential drug targets and supercharge drug discovery efforts.

One of the selected protein complexes has an intriguing label: unknown function. The AI methods have helped us to discover protein complexes with roles that we do not yet understand! Using AI methods to identify mysterious actors on the cellular stage foreshadows exciting discoveries ahead enabled by our new computational microscopes.

Leveraging evolutionary signals

The leap forward with inferring protein complexes builds on the exciting breakthrough reported last year on using deep neural networks to predict the 3D structure of proteins from their underlying amino acid sequences. DeepMind’s AlphaFold system demonstrated how a deep learning pipeline could predict protein structure as well as x-ray crystallography. The methods harnessed to identify the previously unknown protein complexes and their configurations leveraged the capabilities of both AlphaFold and RoseTTAFold.?

The analysis of protein complexes in yeast started with a consideration of evolutionary signals about the co-evolution of proteins. The intuition is that proteins that interact in critical ways in a cell would likely have had to evolve together across animal species to maintain their key functions. The team compared the amino acid sequences of 6000 known yeast proteins to their orthologs, the different variants of the proteins appearing in thousands of plants and animals. That analysis led to the identification of 8.3 million pairs of proteins that appear to have changed in synchrony over the millennia.

In a second step, the likelihood that each of the candidate pairs that form protein complexes were explored via a modified version of RoseTTAFold and AlphaFold, yielding 3D structures and potential points of interactions between proteins. In the end, the computing pipeline identified 1506 proteins that are likely to interact and the 3D structures of 712 were computed.

New understandings of our cells

The new results are deeply relevant to human biology because the machinery of our cells is remarkably similar to the machinery of yeast. We, and all the animals and plants around us, are constructed from eukaryotic cell building blocks. The eukaryotic cell has served as a resilient “platform” for the evolution of complex life. The core structure and processes of eukaryotic cells have largely been conserved; they have changed surprisingly little over millions of years of evolution across thousands of plants and animals. So, understanding the interactions, dalliances, and couplings of proteins within the unicellular yeast is a leap forward on understanding the functioning of our own cells--and an important steppingstone on the aspirational path of decoding the full human protein interactome.

There’s much work ahead on the path to developing a more complete understanding of the human protein interactome. I’m optimistic that we’re well on our way and I’m excited about the next steps in our collaborations and learnings.

Eric Horvitz serves as Microsoft’s Chief Scientific Officer, and previously served as director of Microsoft Research, overseeing research labs in Canada, India, U.K., and the U.S. He received Ph.D. and M.D. degrees at Stanford University.?