[Edaily Reporter KIM JI-WAN ] While an era is dawning in which artificial intelligence (AI) can rapidly design tens of thousands or even hundreds of thousands of antibodies, the designs produced by AI do not immediately become new drug candidates. It is necessary to verify whether the actual antibodies are produced, whether they bind properly to disease targets, whether they can be manufactured effectively, and whether they remain stable in environments with heat or acidity. PROTEINA CO., LTD. anticipates that the bottleneck in AI-driven new drug development will arise precisely at this “actual experimentation” stage.

Hwang Seong-taek, a director at PROTEINA CO., LTD., emphasized during a press briefing with Edaily on the 28th of last month, “Even if AI designs 100,000 antibodies in just one hour, its value is diminished if it takes several months to verify them in practice.” He added, “PROTEINA CO., LTD.’s competitive edge lies in converting the numerous AI-generated candidates into actual antibodies, testing them rapidly, and selecting the best ones.”

At the heart of this is SPID, which PROTEINA CO., LTD. has been developing for approximately 16 years. SPID is a platform that allows for the simultaneous comparison of numerous antibodies using only a small amount of sample. According to the company, it enables the evaluation not only of how well an antibody binds to its target but also of characteristics necessary for actual drug development, such as productivity, thermal stability, pH sensitivity, cellular internalization, immunogenicity, and aggregability. The company states that it currently generates data on approximately 10,000 antibodies per week and has even reduced the antibody optimization process—which typically takes 2 to 3 years—to as little as 3 to 6 months in some cases. The following is a Q&A with Hwang Seong-taek, Director of PROTEINA CO., LTD.

△What is PROTEINA CO., LTD.’s greatest technological competitive advantage?

-While the AI model itself is important, the bigger challenge going forward is how quickly we can actually test the candidates generated by AI. Even if AI designs 100,000 candidates, if we can only test a few hundred in actual experiments, the overall speed of new drug development won’t increase. The area where we are most confident is our ability to use SPID to actually produce a vast number of antibodies and validate them on a large scale.

△How would you explain SPID in simple terms to Saramin who is unfamiliar with biotechnology?

-You can think of it as a technology that actually tests an antibody once the AI suggests, “This antibody seems promising.” It’s not just about running calculations on a computer. We order DNA, produce the actual antibodies in cells, and then verify whether they bind well to the target. We also assess whether they can be produced efficiently, are heat-stable, and remain stable in the body’s environment. In other words, the core of SPID isn’t just about finding “antibodies that bind well”; it’s about quickly identifying, from among countless candidates, the “antibodies that will actually become drugs.”

△What equipment and technologies make up SPID?

- We developed Pi-Chip, Pi-View, and Pi-InSight in-house. We didn’t simply purchase external equipment and modify it. Our CEO majored in electrical engineering and, starting with chip technology, has spent about 16 years building our own platform. The key is that we can accurately observe the binding of antibodies to target proteins even with very small sample volumes.

△Why is it possible to analyze such small amounts?

-When antibodies are produced in cells, the culture medium contains not only the antibodies but also various proteins and impurities. If these are applied directly to a standard surface, these substances bind along with the antibodies, making accurate measurements difficult. Our Pi-Chip has a surface treatment designed to capture the signal from the desired antibody while minimizing the binding of unwanted substances as much as possible. This allows for analysis even without a complex purification process.

△What is the biggest difference from existing SPR and BLI methods?

-Conventional methods often involve multiple steps, such as measuring binding affinity, then using different equipment to verify productivity, and conducting separate stability tests. If a problem is discovered during this process, researchers must return to an earlier stage to test a new candidate. Additionally, there is a risk that some samples may be lost or their condition may change during the purification process—which involves mass-producing antibodies and filtering out impurities—or while undergoing various analytical steps. This can make it difficult to accurately assess the actual yield in its original state.

SPID allows for the direct analysis of small amounts of antibodies produced by cells and, using the same platform, enables the continuous assessment of not only binding affinity but also productivity, thermal stability, and pH sensitivity. The key difference is that it reduces the need to repeatedly switch between various instruments and stages, while also enabling the rapid early screening out of antibodies with a high likelihood of failure from among numerous candidates.

※ SPR (Surface Plasmon Resonance): A representative analytical technique that uses changes in light to measure how well an antibody binds to a target protein and how easily it detaches. It is widely used to precisely measure antibody binding affinity.

※ BLI (Bio-Layer Interferometry): This technique involves attaching antibodies or proteins to a sensor surface and measuring the changes in light that occur when they bind to other substances. It offers the advantage of facilitating the simultaneous analysis of multiple samples.

△What happens after AI designs 100,000 antibodies?

-For example, if 20% of the 100,000 antibodies are actually produced and bind to the target, that leaves 20,000. However, these 20,000 are not yet drug candidates. This is just the beginning. From the 20,000, the number is narrowed down to several hundred antibodies with strong binding affinity, and from those, the ones that can be produced efficiently are selected again. Subsequently, the list is further reduced to a few dozen while verifying thermal stability, pH stability, and other factors. Finally, after undergoing more detailed testing, a few final candidates are selected. You can think of SPID’s role as quickly narrowing down a large number of candidates to just a few that have a high likelihood of becoming actual drugs.

△Compared to traditional antibody development, how many compounds can be tested?

- Previously, researchers would often rely on their own experience or published papers to judge that “changing this amino acid might improve the compound,” and then synthesize and test about 20, or at most 50, variants. We actually generate hundreds or even thousands of variations from a single antibody and analyze the data. In one case, we started with about 1,000 variations and, through additional combinations, analyzed approximately 630 of them within a month, identifying compounds with binding affinity more than 10 times greater than the original antibody. In another project, we saw improvements of up to several dozen times.

△Does the development time actually decrease?

-Since there are differences between antibodies, I can’t say the timeframe is always the same. However, there have been cases where we’ve reduced a process that typically takes about 2 to 3 years to just 3 to 6 months. This is because we can screen out low-potential candidates early on by testing many candidates simultaneously.

△How many antibodies can you process currently?

-In our current projects, we are generating data on approximately 10,000 antibodies per week. By adjusting our equipment and project workload, we can scale this up to about 30,000–40,000 per month. Going forward, we plan to expand automation to reach 100,000 per month by the first quarter of 2027, increase to 500,000 thereafter, and ultimately process 1 million per month by 2028.

△How are AI and SPID connected?

-AI designs antibodies, and SPID verifies through actual experiments whether they are correct or incorrect. We then feed those results back into the AI for training. From the AI’s perspective, it’s not just the successful data that matters. Failure data—such as “the antibody wasn’t produced despite this design” or “it was produced but didn’t bind to the target”—is also crucial. We have been collaborating with researchers at Seoul National University on this work since 2024 and have been training the AI model based on approximately 500,000 data points.

△Does this ultimately mean that experimental data has become more important than the AI model?

-It doesn’t mean AI isn’t important; rather, it reflects the changes that will come as AI models reach a certain baseline level of quality. PROTEINA CO., LTD. is also developing AI technology in collaboration with Professor Baek Min-kyung’s team at Seoul National University, supported by a national research project from the Ministry of Science and ICT since April 2024. However, a great many high-quality AI models are emerging, and the availability of open-source technologies is also increasing. I believe that in the future, it will become more important to determine who can actually and quickly validate the substances created by AI and continue to accumulate high-quality experimental data, rather than who possesses a single AI model.

△What about the profitability of the antibody optimization service?

- We typically receive between 100 million and 200 million won per antibody optimization project. However, we determined that relying solely on repeating such projects dozens of times would limit the company’s significant growth. Therefore, we are shifting our focus from a service that simply improves antibodies for a fee to joint development. The structure involves creating promising candidates using our technology, developing them together with partner companies, and sharing greater value when technology transfer eventually takes place.

△Which is the company’s main focus: large-scale validation services or joint new drug development?

- The two businesses are interconnected. We generate cash through our large-scale screening service for the numerous antibodies created by external AI-driven drug development companies. We then reinvest that money into our own pipeline or into drugs developed through joint ventures. PROTEINA CO., LTD.’s ultimate goal is to become a company that rapidly generates high-quality drug candidates and consistently licenses them out. Our objective is to be a self-sustaining AI-driven drug development company that generates revenue through its own technology and reinvests those funds into creating new drugs, rather than a company that must raise tens of billions of won from external sources every time it develops a new drug.