Scientists have developed artificial intelligence software that can generate proteins that could be useful as vaccines, cancer treatments or even tools to pull carbon pollution out in the air.
An artist’s conception of a hallucinogenic deep machine learning program of new ideas for vaccine molecules. Image credit: Protein Design Institute
The study, reported in Science today, was led by the University of Washington School of Medicine and Harvard University. Article titled “Building functional sites of proteins using deep learning”.
“The proteins we find in nature are amazing molecules, but engineered proteins can do it,” said senior author David Baker, HHMI Investigator and professor of biochemistry at UW Medicine. more than that”. “In this work, we show that machine learning can be used to design proteins with a variety of functions.”
For decades, scientists have used computers to try to design proteins. Several proteins, such as antibodies and fusion-binding proteins, have been adapted into drugs to fight COVID-19. Others, such as enzymes, aid in industrial production. But a single protein molecule usually contains thousands of bonded atoms; even with specialized scientific software, they are difficult to research and manufacture.
Inspired by how machine learning algorithms can generate stories or even images from prompts, the team has started building similar software to design new proteins. “The idea is the same: neural networks can be trained to see patterns in the data. Once trained, you can give a prompt and see if it can produce a good solution. Lead author Joseph Watson, a postdoctoral scholar at UW Medicine, says the results are often compelling — or even beautiful.
The team trained multiple neural networks using information from the Protein Data Bank, a public repository of hundreds of thousands of protein structures from across the kingdoms of life. The resulting neural networks have surprised even the scientists who created them.
The team has developed two approaches to designing proteins with new functions. The first tool is called “hallucination” similar to DALL-E or other popular AI tools that generate new results based on simple prompts. The second feature, called “inpainting,” is similar to the autocomplete feature found in search bars and modern email clients.
“Most people can come up with new pictures of cats or write a paragraph from the prompt if asked, but with protein design, the human brain can’t do what today’s computers can do. doable,” said lead author Jue Wang, a postdoctoral scholar at UW Medicine. “Humans can’t imagine what the solution might look like, but we’ve built machines that do.”
To explain how the neural network ‘hallucinates’ a new protein, the team compares it to how it might be to write a book: “You start with a bunch of random words – overall meaningless. Then you make a request like in the opening paragraph that it should be a dark and stormy night. Then the computer changes word by word and wonders ‘Does this make my story more meaningful?’ If it does, it will hold the changes until a complete story is written,” explains Wang.
Both books and proteins can be interpreted as long strings of letters. In the case of proteins, each letter corresponds to a chemical building block called an amino acid. Starting with a random sequence of amino acids, the software changes the sequence over and over until it produces the final sequence that encodes the desired function. These final amino acid sequences encode proteins that can then be manufactured and studied in the laboratory.
The team also showed that the neural network can fill in the missing parts of a protein structure in just a few seconds. Such software could aid in the development of new drugs.
“With autocomplete aka “Inpainting Protein,” we start with the key features we want to see in a new protein, then let the software do the rest. Those traits could be binding motifs or even enzyme active sites, Watson explains.
Lab testing shows that many proteins created through hallucinations and printed paint work as intended. This includes novel proteins that can bind metals as well as those that bind to the anti-cancer receptor PD-1.
The new neural networks can make several different proteins within a second. Some include potential vaccines for the deadly respiratory syncytial virus, or RSV.
All vaccines work by introducing a portion of the pathogen into the immune system. Scientists usually know which fragments will work best, but creating a vaccine that achieves the desired molecular shape can be challenging. Using new neural networks, the team prompted a computer to generate new proteins that include the necessary pathogen fragment as part of their final structure. Free software to generate any support structure around the primary segment, generating several potential vaccines with diverse molecular shapes.
When tested in the lab, the team found that known antibodies against RSV were stuck to three of their hallucinogenic proteins. This confirms that the new proteins have the expected shape and suggests they may be possible vaccine candidates that could prompt the body to make its own highly specific antibodies. Additional testing, including on animals, is still needed.
“I started working on the vaccine tool just as a way to test our new methods, but while working on the project, my two-year-old son contracted RSV and had to spend an evening in the emergency room for a lung transplant. deleted. It made me realize that even the ‘testing’ issues we’re working on really make sense,” Wang said.
“These are exciting new approaches, but there is still room for improvement,” said Baker, a recipient of the 2021 Life Sciences Breakthrough Prize. “For example, the design of highly active enzymes is still very challenging. But every month, our method keeps getting better! Protein structure prediction has transformed in deep learning over the past two years, we are now in the midst of a similar protein design transition. “
Source: University of Washington
