Proteins, artificial intelligence, and future of pandemic responses – Dailyuw

Posted: April 13, 2021 at 6:27 am

The Institute for Protein Design (IPD) at the UW announced March 31 a $5 million grant from Microsoft to collaborate on applying artificial intelligence to protein design.

Microsofts chief scientific officer Eric Horvitz and the IPDs director David Baker, in an article with GeekWire, said they believe that this collaboration will lead to major strides in medicine and technology, and accelerate the scientific response to future pandemics.

The IPD designs proteins molecules that carry out a wide range of functions from defending against pathogens to harnessing energy during photosynthesis from scratch, with the goal of making a whole new world of synthetic proteins to address modern challenges, according to the institutes website.

Researchers at the IPD have developed promising anti-viral and ultra-potent vaccine candidates against SARS-CoV-2, the virus that causes COVID-19, that are currently in human clinical trials.

And in protein design, form follows function.

We use 3D protein structures on the computer to design the protein sequences, Brian Coventry, a research scientist in the Baker Lab at the IPD, said. When we order the protein sequence, its function in real life should exactly mirror that on the computer.

But that does not always happen.

The problem with this method, which is based on the first principles of both physics and chemistry, is that it produces an abundance of possible proteins which must be tested, the majority of which do not have the exact desired form, Coventry said.

Coventry recently worked on a team that developed a SARS-CoV-2 antiviral medication candidate, and he stressed that for antivirals, it is important that the designed protein be precisely atomically correct.

In the context of a pandemic, the fast development of highly accurate therapeutic synthetic proteins is desirable. This is where deep learning, a subset of artificial intelligence modeled after the brains neural networks, comes into play.

There is a lot of room for improvement, Minkyung Baek, a postdoctoral scholar in the Baker Lab at the IPD, said about the first principles-based method of protein design. Baek believes that deep learning methods can be used to quickly discriminate between possible proteins and optimize design to produce proteins that are more stable and bind more tightly to targets.

Deep learning models are given a training data set, in this case experimental results of the structures of designed proteins, and then can learn based on real-world data. They use that information to predict and design protein structures, Baek said.

Microsoft has given the IPD access to their cloud computing service Azure, which will enable them to train and test deep learning models about 10 times faster, according to Baek.

Baek hopes that this will speed up the development of effective deep learning models, which will be helpful not only for designing proteins that match existing biological proteins, but also for discovering the structure of naturally occurring proteins.

There are many real-world situations where the structure of the target is not precisely known. In these situations, researchers must predict the shape of the metaphorical lock and design the key simultaneously.

Being able to better predict the structure of a protein when given its genetic code is important, with Baek using the variants of the COVID-19 virus as an example.

Using our deep learning base, we can predict the protein structure of the variant, and starting from there we may get some clue [about] why that variant may have been more severe or easy to spread, Baek said.

But these deep learning models have some limitations. They are limited by the available training data set, are not always generalizable to multiple situations, and do not explain the reasoning behind their decisions, Coventry said.

Despite these factors, Coventry and Baek are both optimistic about the potential for deep learning to improve the protein design process.

At the end of the day, Id like to see a 100% success rate, you know, Coventry said. Someday Im sure its possible.

Reach reporter Nuria Alina Chandra at news@dailyuw.com. Twitter: @AlinaChandra

Like what youre reading? Support high-quality student journalism by donatinghere.

Read more:

Proteins, artificial intelligence, and future of pandemic responses - Dailyuw

Related Posts