Deepmind, the UK-based, Google-owned AI business, has announced a scientific breakthrough in the visualisation of protein structures that many of the world’s leading scientists didn’t believe possible in their lifetime. Described as a “once in a generation advance”, the achievement has been hailed as one that will lead to a revolution in biology and drugs discovery.

For 50 years scientists have been trying to unlock the key to understanding, and recreating, the shapes that proteins, the building blocks of all life, fold into. Deepmind has now developed AI able to do so.

Its algorithms have been shown to be able to predict the structures proteins form based on their chemical composition. The rate of success is better than the top human scientists are able to achieve in two thirds of cases. And it takes a fraction of the time required by humans.

At an atomic level, proteins are the building blocks that every process of life powered by. By forming into intricate and specialised shapes, they are able to perform the roles that support every biological process. From photosynthesis in plants to auto-immune reactions in the human body and the ability of viruses to hijack the cells of a host. Understanding the exact shapes sub-atomic proteins form is key to being able synthetically replicate biological processes, or manipulate them.

By extension, knowledge of protein shapes throws wide open the doors to a potential avalanche of new drugs being discovered. Some of those are now expected to play a central role in treating previously untreatable diseases, with certain forms of dementia high on the list.

Deepmind’s breakthrough was made while it participated in a competition established to encourage teams to work out protein structures computationally. Deepmind also won the biennial competition in 2018, but at the time the accuracy of its algorithmic predictions was still not reliable enough to be ‘biologically useful’.

It was a case of Deepmind’s scientists producing the best of what were many poor attempts. Two years later, however, and Deepmind has been credited with ‘solving’ one of biology’s greatest problems.

The significance of Deepmind’s breakthrough was summed up by Venki Ramakrishnan, the protein scientist who, alongside Thomas A. Steitz and Ada Yonath won the 2009 Nobel prize in Chemistry “for studies of the structure and function of the ribosome”.

He stated of the advance:

“It has occurred decades before many people in the field would have predicted. It will be exciting to see the many ways in which it will fundamentally change biological research.”

Who are Deepmind?

Deepmind are a London-based artificial intelligence company and research laboratory founded in 2010 Demis Hassabis, the current chief executive, Mustafa Suleyman and Shane Legg. The company was acquired by Google in 2014 and a year later become a wholly-owned subsidiary of Alphabet, Google’s holding company.

Deepmind is considered one of the most advanced developers of machine learning AI in the world and previously developed the first computer able to beat the world’s top human players of the board game Go.

While such developments demonstrated the potential of machine learning and attracted interest, they were largely abstract applications of the technology with little application to real world problems. Deepmind’s latest achievement is very much applicable. Mr Hassabis commented his companies work will result in “whole new avenues and branches of exploration for science and industry”.

What Deepmind’s protein structure computation breakthrough practically mean for science?

For those of us for whom the scientific details of Deepmind’s breakthrough are beyond our comprehension, which is the vast majority, its significance becomes most apparent from the unrestrained enthusiasm of the few for whom they are not. Ewan Birney, the European Molecular Biology Laboratory’s deputy director general, remarked his field was “in a different zone after this”. Followed by:

“I nearly fell off my chair when I saw these results. It’s hard to know what the impact will be because it’s been such a holy grail.”

But what might be the concrete developments that result over coming years?

Drugs discovery is the first big field of obvious application. Pharmaceuticals often work on the principal of molecules that block proteins. For a molecule to be able to block a protein, it usually has to latch onto it. That requires a inverted form to the protein, like two pieces of a jigsaw puzzle.

Until now, figuring out the exact shape of a protein was the kind of task that would often be the subject matter of a PhD thesis, and span its duration. Dr Hassabis says Deepmind’s machine learning approach will now be able to give scientists working on drugs discovery the structure of a protein within days, drastically speeding up their work.

Virology is another clear application. The team behind the breakthrough have already used their program to visualise the proteins on the surface of the Covid-19 coronavirus. In future, that could accelerate the production of new and better vaccines.

Custom-design of proteins for all sorts of applications, for example to biodegrade plastics, is another likely consequence. And in the bigger picture, we’ll have a far deeper understanding of biology. It’s difficult to forecast what new discoveries that might catalyse as a chain reaction, but the possibilities are exciting.

Source: The Times

As already discussed, proteins are fundamental to every biological process. They are also generally at the bottom of processes that go wrong, leading to illness or even, eventually, death. Which is why many drugs act on particular proteins.

In recent years we’ve gotten pretty good at figuring out the chemical formula of different proteins. But the real key to how proteins work is not the atoms they consist of but the structures they create. Drugs discovery relies on an understanding of what a protein looks like. The protein can be though of as a lock, and the active ingredients of a drug the corresponding key.

Chemistry now gives us a one-dimensional sequence of atoms but Deepmind’s machine learning program will now allow us to understand the 3D shapes those sequences fold into. Until now that’s taken years per protein. Which has been a serious bottleneck in the context of the 20,000 different proteins that power the biological processes of the human body, and 200 million known proteins found on planet Earth.

Dame Janet Thornton of the European Molecular Biology Laboratory, reflects that she thought she would never see the protein-folding problem solved:

“Proteins are the most beautiful, gorgeous structures and the ability to predict exactly how they fold up in three dimensions is really very, very challenging and has occupied many, many people over many years.”

What comes next?

Despite the scale of the Deepmind breakthrough, the research is far from complete. Some proteins seem harder for the machine learning programme to predict than others. Future research is also needed to model how different proteins interact with each other. That’s an even more intricate problem, and just as important.

The other hurdle to be overcome is that the model has been trained on crystallised proteins. However, in a living biological environment, proteins are not crystallised but are moving structures that bend and warp. Modelling that ‘real world’ behaviour is much harder. The data Deepmind can discover hints at how free-moving proteins might flex and contort, but doesn’t yet offer any definitive answer.

There’s a lot of knowledge still to be uncovered. But the fact that Deepmind’s AI was able to be successfully applied to solving how to compute to shape of proteins suggests that machine learning, still a relatively new science, will also prove key to making progress in solving the remaining problems. And if this discovery is any indicator, perhaps far earlier than anyone ever imagined until very recently.