[Exclusive] AI Drug Development Breakthrough: Achieving Both Efficacy and Productivity

Joint Research Team Led by Professors Baek Min-kyung and Yoon Tae-young at Seoul National University

Application of AlphaFold and ProteinMPNN on the SPID Platform

Enhancing Both Productivity and Binding Affinity


Development of Ada

Domestic researchers have made significant progress toward developing artificial intelligence (AI)-based antibody drugs that simultaneously achieve both productivity and binding affinity. By overcoming the traditional limitations of antibody engineering—where an increase in binding affinity often results in reduced productivity—they have demonstrated the practical effectiveness of AI-driven drug design.

According to the pharmaceutical, biotech industry, and academia on April 27, a collaborative research team led by Professors Baek Min-kyung and Yoon Tae-young from the Department of Biological Sciences at Seoul National University published these findings on the international preprint server 'bioRxiv'. The paper, titled "Structural Logic of AI-Based Antibody Rescue and Therapeutic Optimization Through Large-Scale Fitness Landscape Exploration," was released on bioRxiv as a preprint ahead of publication in a peer-reviewed international journal.

While AI antibody drugs are recognized as a new leap forward for the pharmaceutical and biotechnology industries, the dilemma between binding affinity and productivity in the development process has been a chronic challenge. Even if an antibody with excellent binding to a disease-causing antigen is discovered, large-scale cultivation and production in pharmaceutical factories often result in a dramatic drop in productivity. This is due to a phenomenon known as "negative epistasis," where even amino acids with outstanding individual characteristics, when combined in a single antibody, can actually compromise binding affinity or severely damage productivity. As a result, drug development experts have pointed out that AI was previously considered unsuitable for antibody drug development.

The root cause of this issue is that the amino acid sequence of the "complementarity-determining region (CDR)"—the core site where antigens and antibodies bind—often fails to harmonize with the overall scaffold of the antibody, leading to instability in the antibody protein itself. When the CDR sequence is modified to enhance binding affinity, physical clashes with the rest of the antibody's three-dimensional structure can occur, preventing the protein from folding correctly and resulting in structural defects. Cells then dispose of these poorly folded proteins, making it impossible to mass-produce the antibody in factories.

The research team utilized the single-molecule protein interaction detection (SPID) platform developed by the Korean biotech company Proteina, which can precisely analyze tens of thousands of antibody variants per week, to obtain data on 9,517 adalimumab (an autoimmune disease treatment, product name Humira) variants. They measured both binding affinity and productivity, discovering that these two properties are intricately intertwined—much like a rugged and irregular mountain range—making it impossible to create the optimal antibody simply by combining favorable mutations. To address this complexity, the team employed AI technologies. Specifically, they used 'AlphaFold3', an AI for protein structure prediction, and 'ProteinMPNN', a reverse protein design AI model, to uncover hidden rules linking the three-dimensional structure of antibodies with their productivity.

Through this approach, the AI analyzed the structures of antibodies that exhibited excellent binding affinity but low productivity—candidates at risk of being dropped from the drug pipeline—and proposed a tailored solution of replacing just a single amino acid. When the research team applied this structure-based "rescue strategy" to actual antibodies, they succeeded in restoring cell culture productivity to normal levels while maintaining the original binding affinity to the target.

The final next-generation adalimumab variants identified through this process (clones 1207 and 1208) demonstrated up to a 100-fold improvement in in vivo efficacy compared to existing therapies in animal experiments. When the research team administered these new antibodies to mouse models artificially induced with psoriasis, they found that even an extremely small dose—only one-twentieth to one-hundredth of the amount used with conventional treatments—was sufficient to reduce severe skin inflammation and thickness to normal levels. The results showed that the disease was effectively suppressed with 100 times the therapeutic potency, without side effects or weight loss.

This near-miraculous improvement over conventional drugs was found to be due not only to the strengthened binding force of the antibody to its target but also to a dramatic increase in "complex lifetime"—meaning, once bound, the antibody remained attached and active for a much longer period. Measurements revealed that while existing adalimumab maintains binding for about 1.8 hours, the newly engineered antibody variants sustained binding—and thus therapeutic efficacy—for up to 50 hours.

This content was produced with the assistance of AI translation services.