Korean Female Scientist Who Challenged Google's Pride... "Changing the Biotechnology Landscape by Combining AI" [Reading Science]

Exclusive Interview - Baek Min-kyung, Former Postdoctoral Researcher at Washington University
Developed RosettaFold Surpassing DeepMind's 'AlphaFold2'
Protein Structure Analysis Taking Years Now Completed in Minutes
"Bold and Active Investment Needed in Convergent Science Research"
Also Revealed Appointment as Assistant Professor at Seoul National University Department of Life Sciences in September

[Asia Economy Reporter Kim Bong-su] "It has changed the landscape of scientific discovery."

This is an evaluation of the protein structure prediction artificial intelligence (AI) program 'RoseTTAFold' developed by Min-Kyung Baek (32), a former postdoctoral researcher at the University of Washington in the U.S., who will join the Department of Life Sciences at Seoul National University as an assistant professor in September. Baek, together with Professor David Baker, created a masterpiece that instantly surpassed the performance of AlphaFold2, an AI of the same kind developed by DeepMind, a Google subsidiary, which mobilized enormous computing resources, personnel, and capital. It was such an important achievement that the world-renowned scientific journal Science selected it as the top innovative research achievement of the year through a reader vote in December last year.

In a written interview with Asia Economy, Baek said, "By understanding protein structures through AI, we have laid the foundational groundwork for advancing humanity's scientific and technological research to the next level." She also emphasized, "Just as AlphaFold2 achieved groundbreaking progress by integrating AI and biotechnology, interdisciplinary fusion research is currently becoming the foundation for significant scientific and technological advancements, so bold investment and stable support are necessary."

Below is a Q&A with former researcher Baek.

- What is the significance of protein structure prediction?

△ Proteins are very important biomolecules involved in almost all life phenomena. For example, various proteins are involved in processes such as sensing external stimuli like vision, taste, and smell, or obtaining energy for cells through food. Antibodies, which have become more familiar recently due to COVID-19, are also very important protein molecules involved in immune responses. They are essential elements for understanding various life phenomena such as external stimulus detection, signal transduction, energy supply, and immune response. Proteins have such diverse functions because they have different structures depending on their sequences (combinations of amino acids), which determine their functions. In other words, if we can easily determine the structure from the protein sequence, it greatly helps us understand how the protein is involved in life phenomena. Moreover, based on this, it can be applied to various fields such as developing therapeutics, vaccines, plastic-degrading enzymes, and biosensors.

- Was protein structure prediction not possible before?

△ Of course it was. Many experimental scientists have devoted great efforts to experimentally determine protein structures. However, it required significant costs and time ranging from several months to several years. By using the AI-based protein structure prediction method developed this time, the cost and time required for experimental structure determination can be reduced. This allows more focus on more important biological problems or applied fields such as drug development.

- What are the differences between RoseTTAFold and AlphaFold2?

△ There has been a question of how to improve the performance and efficiency of traditional computational methods developed so far by utilizing AI. Among traditional computational methods before AlphaFold2 and RoseTTAFold, the best-performing method used evolutionary information of proteins. Then came the idea that if patterns related to protein structures hidden within the given data of protein sequence alignments could be found, protein structures could be well predicted. Pattern recognition is exactly what AI excels at. At this point, AI was applied to protein structure prediction. Both AlphaFold2 and RoseTTAFold predict protein structures by utilizing evolutionary information accumulated in nature, so they are generally similar. However, the detailed algorithms differ at the implementation stage. RoseTTAFold frequently creates 3D structures and receives feedback during the process of obtaining interaction information between amino acids from the sequence alignments containing evolutionary information. In contrast, AlphaFold2 generates the 3D structure only at the end, so it receives feedback on the generated structure much less frequently. Because RoseTTAFold is designed to frequently communicate between protein sequence evolutionary information and 3D structure, it can predict structures more efficiently than AlphaFold2.

- What impact and challenges will this have on future science and technology such as medical and life sciences?

△ Take new drug development as an example. One commonly used method in developing new drugs is structure-based drug development. At this time, it is essential to predict the binding structure of the target protein and the drug candidate. Previously, if the experimentally determined structure of the protein was not available, it was difficult to use structure-based drug development methods. Protein structure prediction technologies like AlphaFold2 provide high-accuracy models for proteins whose structures were previously unknown, enabling structure-based drug development even for target proteins without experimental structures.

However, this does not solve all problems in the drug development process. Target proteins often have multiple structural states depending on their activity, and drugs must bind to specific states to regulate protein activity and act as medicines. Current protein structure prediction technologies do not predict the various structural states that proteins can have. Proteins may undergo slight structural changes when drugs bind, but predicting these changes is also currently impossible. In other words, simply knowing the protein structure is not enough; high-accuracy predictions considering structural changes of proteins when binding with various drug candidates, as well as predictions of drug binding affinity and activity, are additionally required. If protein structure prediction technology can predict interactions between proteins and other molecules in the future, it may become more directly relevant to our daily lives.

**Who is Dr. Min-Kyung Baek?

Born in 1990 in Gwangju. Graduated from Gwangju Science High School and Seoul National University Department of Chemistry (Ph.D.). Worked as a postdoctoral researcher at the University of Washington from May 2019 to July 2022. In July 2021, developed the protein structure decoding AI program 'RoseTTAFold' with Professor David Baker. In December of the same year, selected as 'Top Innovative Research of the Year' by the international journal Science (first Korean to receive this honor). Scheduled to join the Department of Life Sciences at Seoul National University as an assistant professor in September 2022.