We've got the news down to a science!

The Science Survey

We've got the news down to a science!

The Science Survey

We've got the news down to a science!

The Science Survey

AI’s New Role in Biochemistry

Scientists have begun to take a new approach to one of biochemistry’s biggest problems: using artificial intelligence to help predict the structure of biological molecules.
Stefan Bieniasz
Here is an image of one of the pellets/cartridges that are put into a cryo-EM microscope.

Whenever someone mentions artificial intelligence, the future of technology and its impact comes to mind. Some people imagine a generator used to steal jobs from artists and musicians. Others view it as an easy tool students use to cheat and plagiarize. Different people imagine some all-present sentient being right out of Terminator, wishing only to take over and rule the world.

The portrayal of artificial intelligence in the media has been similar to the portrayal of aliens. Both are often depicted as wildly advanced, malevolent beings whose only instincts are to expand and conquer – both are, after all, based on what humans would do if we had the kind of power that they have – demonstrated by a robot’s glowing red eyes or an alien’s sharp features.

However, our understanding has grown far from what we knew in the 1980s. While people were still fantasizing about the dystopia that artificial intelligence could bring, AI has begun to take on new forms. AI Programs like ChatGPT and Midjourney have demonstrated the capabilities AI can have. Bing, the search engine run by Microsoft, now has a built-in AI assistant that works as both a chatbot and image generator. Canva, a popular website used to make slideshows, posters, and presentations, now has several tools allowing users to edit images they insert into the website with AI, letting them manipulate objects, remove backgrounds, or expand images.

Artificial intelligence has been the subject of controversy for many years now. While some may be afraid of it becoming advanced enough to take over jobs, and others debate the ethics of using AI, scientists have started to use it to their advantage and potentially usher in a new era of discovery.

Using artificial intelligence programs, biologists have been able to predict the structure of proteins, a crucial step to advancing our understanding of microbiology.

Proteins are biological molecules that serve countless functions within our bodies. They make up the enzymes in our stomachs that break down food, protect us from infections, and determine which hormones our bodies produce. They can carry atoms or molecules around the body, and make up microscopic structures in our cells. Often described as the ‘building blocks of life,’ a single human cell could contain about 42 million different protein molecules.

To understand what function a protein serves, a scientist must first understand its structure. Proteins are made up of smaller molecules, called amino acids, which link together to form sequences. These ‘chains’ of amino acids fold and wrap around each other to form different types of proteins.

“One of the guiding principles of structural biology is that a protein’s structure determines its function,” said Jen Bohn, a Research Educator and Biochemist at the Rockefeller University. “You can think about this like giving a friend a high five. If you both enter into the high-five with your hands open, upright, flat, and facing one another then you will successfully high-five each other. If you are in a high-five position, but your friend is making a fist, a crisp sounding high-five is unlikely to happen.”

“Proteins are like hands in this analogy. In a given shape, two proteins can interact with one another to achieve a given output (a high-five),” Bohn elaborated. “If a protein takes on an alternate confirmation, or shape, then the desired output is not achieved.”

Understanding protein structures was of special interest for many scientists during the COVID-19 pandemic. A certain protein on the outside of the coronavirus particle, called the Spike protein, interacts with another protein on the surface of a cell, called the ACE2 receptor. “Scientists were interested in understanding the architecture of this molecular high-five so they could try to design molecules that might block the interaction and interfere with that very dangerous high-five between the COVID-19 Spike protein and the human ACE2 receptor protein,” said Bohn, who worked on researching protein binds that inhibited viral infection. Research labs studying the coronavirus Spike protein were crucial in developing a vaccine for COVID-19. Similar processes can also be applied to the development of a variety of vaccines and medicines for other diseases.


The process to figure out the structure of a particular molecule used to be complex and expensive, requiring massive machines and dedicated labs. 

Dr. Mark Ebrahim of the Rockefeller University specializes in cryogenic electron microscopy–or Cryo-EM–a process capable of examining a sample at a near-atomic level. In order to get such high resolutions, an electron microscope fires a beam of subatomic particles at a sample. “It is like a light microscope,” said Dr. Ebrahim, “but instead of photons [light], it uses electrons. Instead of your eye, the sensor is a camera.”

HEre is Dr. Mark Ebrahim of the Rockefeller University with a Cryo-EM microscope. (Stefan Bieniasz )

Such a microscope requires intense preparation in order to use. First, scientists have to prepare the sample. This involves flash-freezing it in liquid ethanol and nitrogen in order to quickly preserve it, preventing it from falling apart when it is put into the microscope. Ebrahim’s team then uses another lower-resolution electron microscope outfitted with an ion beam in order to cut down a given sample to the desired size – usually on the scale of nanometers (one of which is equal to a billionth of a meter), keeping it small enough to make sure the electron beam can actually interact with it.

In order to pick up such small signals, the camera needs to be incredibly sensitive – which also means it needs to be kept incredibly stable. “All of the microscopes are kept underground to protect them from vibrations,” Ebrahim said. Not only that, but a wire cage runs across the walls, ceiling and floor of the room that houses the microscope. The wires are connected to a sensor inside the machine, which measures the magnetic field, and runs a current in the opposite direction through the cage in order to negate it. “If you have two opposite magnetic fields, they cancel out,” he said. “But even something like the train passing by or a truck outside can disrupt it.”

With all of these factors considered, using an electron microscope requires a large amount of effort on the part of the scientists. The machine itself, housing the sensor, lenses, the electron beam and a variety of other equipment can cost anywhere from five to 15 million dollars. Running a single sample can potentially cost thousands of dollars.

A New Technology

Today, scientists are able to predict the structure of these proteins before they even put them into the microscope. An experimental program called AlphaFold uses artificial intelligence, data from previous experiments, and a global database of known structures in order to predict what new protein structures might look like.

The process to generate protein structure predictions and designs all use artificial intelligence. “There are specialized networks that perform each step,” said David Juergens, a Computational Science graduate student in the Baker Lab at the Institute of Protein Design (IPD). helps to develop machine-learning models for protein design. Baker’s lab designs proteins around the functions they want them to perform.  These proteins can be used for a variety of things, like enzymes, which accelerate chemical reactions, or binding proteins, like the Spike or ACE2 receptor protein.

He and the IPD developed a program called RoseTTAFold Diffusion, which uses the same kind of neural network as AI image generation software like DALL-E in order to design proteins.

“First you need a backbone, then an amino acid sequence, then once you’ve got your backbone and your sequence, you try to predict the structure of the sequence and see if the prediction matches the original design structure that you made,” Juergens said. By starting with only two or three amino acids grouped together, we can then ask the the structure generation model to make a backbone such that it positions all of these amino acids just as they are.” The locations of these amino acids determines how they interact with other biological molecules, like other proteins or certain compounds. By knowing the function that a certain amino acid combination will produce, then they can generate a protein around those molecules to perform that function.

Even though this technology is still new, it has already had an impact on the field. Skycovione, a COVID-19 vaccine that uses protein particles developed by the IPD was approved for use in South Korea and the United Kingdom in 2022. “We also have other molecules that we hope are on their way to clinical trials,” said Juergens “Our institute is constantly making and testing the proteins we make designs for.”

While scientists still do use the older methods in order to definitively confirm the structure of proteins, further developments in artificial intelligence could make programs like AlphaFold even more reliable. Additionally, AlphaFold’s database is publicly available, allowing researchers to share and access 3-D models of protein structures for free, all across the world.

Upon its release in 2021, AlphaFold’s database only had around 360,000 structures. Since then, that database has expanded to include over 200 million predicted structures. 

Predictive models like AlphaFold have gotten more and more accurate, and as artificial intelligence develops, they will likely continue to do so. 

“We’re super excited about it. It’s a really exciting time to be in the field. In the last four to five years, AI has made us a lot better at being able to make custom molecules that do custom functions. The pace of acceleration is really fast,” Juergens said. “The ability to design complex functions on the computer just means that we will be able to, say, if another pandemic comes, make drugs against that virus or that bacteria really fast. Or, if a chemical becomes super valuable, we can discover ways to synthesize that chemical as these methods get better.”

As technology advances, there is always the worry that machines will replace humans. The introduction of mechanical farming tools in 19th-century America drastically reduced the number of people needed to tend to a field. Now, people worry about whether they will lose their jobs to artificial intelligence.

However, in the academic world, certainty is paramount. “Biological molecules are dynamic and can change in shape,” said Bohn. “The thing about AI is that it is only as good as the information we provide it.”

“People might think AlphaFold will ‘kill’ cryo-EM – but it’s actually the opposite,” said Dr. Ebrahim. Even after years in development, an artificial intelligence could be 99% accurate, but in a field like biology, that 1% could change everything. If even a single part of a protein molecule is off, the whole structure could cease to function. Many disorders like Parkinson’s Disease, Sickle Cell Anemia, and Cystic Fibrosis are caused by malfunctioning proteins that haven’t folded properly. This mis-folding can be caused just by a single error in a molecule. Dr. Ebrahim’s lab now processes samples other researchers bring in to ‘check’ the program’s predictions.

“I do not necessarily consider AlphaFold to be an experimental method, but more of a prediction generator,” said Bohn. “While the predictions draw on real data, they are simply a hypothesis and must be tested experimentally. Many things are still unknown and so for now, I think that AI programs are not quite ready to replace experimental approaches.” 

“AI is useful in generating and synthesizing ideas, but we must check the work of AI and always follow-up ourselves, whether it be dividing into thoughts of ChatGPT or performing experiments to validate or debunk AlphaFold structure predictions. Use AI to ask any questions you might have and always, always ask yourself ‘does this make sense and how can I test/verify this?’” Bohn said.

While artificial intelligence programs like AlphaFold and others show promise, both in biochemistry and a variety of other fields, they still have a ways to go before they can – if ever – fully replace the tools we already use.

Artificial intelligence has been the subject of controversy for many years now. While some may be afraid of it becoming advanced enough to take over jobs, and others debate the ethics of using AI, scientists have started to use it to their advantage and potentially usher in a new era of discovery.

About the Contributor
Stefan Bieniasz, Staff Reporter
As a News Editor for ‘The Science Survey,’ Stefan Bieniasz likes to keep up with current and global events. He enjoys using journalistic writing as a way to tell the stories of other people and write about what he is interested in. Even when writing about broader events, Stefan believes one of the most important things is to include details about how they affect people. He also believes that images are some of the most important parts of an article, as they can be just as eye-catching as a good headline, and can be an effective method to convey emotion and detail. In his free time, Stefan enjoys listening to music and doing some creative writing. He is unsure what he would like to pursue in college, but is considering a wide variety of sciences and history. Although his future job may not involve writing, Stefan still believes it is a skill worth developing.