que paso
At the 2026 Global Bio Conference in Seoul, KAIST research vice president Lee Sang-yeop argued that AI, bioengineering and medicine must be developed together. ChosunBiz reports that his team used DeepEC to analyze the functions of 34 million proteins, reducing an estimated conventional workload of 150 years to 10 days. The team also applied AI to adverse-drug-reaction prediction and engineered microorganisms to produce biodegradable plastic and spider silk-related materials.
ChosunBiz reports that Lee Sang-yeop presented the work during a keynote address at the 2026 Global Bio Conference at the Grand InterContinental Seoul Parnas on August 26. Lee is described as a KAIST professor and research vice president whose work combines metabolic engineering, systems biology, synthetic biology and evolutionary engineering. The report frames AI as part of a broader effort to connect biological research with medicine and engineering, rather than as a standalone software product launch.
According to ChosunBiz, Lee began using AI more intensively in 2016 and assembled gaming-oriented graphics processing units to run deep-learning systems. His team developed DeepEC, a system for predicting enzyme functions from protein sequences. The report says the system analyzed 34 million proteins in 10 days, while the conventional standard method would have taken 150 years. ChosunBiz also reports that applying transformer technology improved accuracy and enabled predictions involving substrate-binding and cofactor-binding sites that the system had not been explicitly taught.
The report connects those predictions to biological production. ChosunBiz says Lee’s team redesigned Escherichia coli genes so the cells would accumulate plastic, producing about 180 grams of biodegradable plastic per liter. It also includes a photograph caption describing spider silk produced by microorganisms used by the research team. The article does not specify the production conditions, material properties, experimental controls, scale of the work or whether these results were independently reproduced.
ChosunBiz further reports that the team built an AI system to assess possible adverse reactions among approved drugs. Lee said that pairing 2,159 approved drugs would create roughly 2.3 million combinations, while about 460,000 adverse reactions were known at the time described. The system reportedly predicts 113 types of adverse reactions and can suggest alternative drugs with similar effects but different structures. The article gives Paxlovid as an example, saying analysis against 2,248 approved drugs identified potential interactions with 1,600. These figures and capabilities are attributed to Lee through ChosunBiz and are not independently confirmed by the supplied source.
Lea la fuente principal: biz.chosun.com ↗
Por qué es importante
The report illustrates a practical use of AI in biology: narrowing the search space for experiments involving proteins, enzymes, medicines and engineered cells. If the reported performance and production results hold up under independent evaluation, such systems could help researchers identify promising biological designs more quickly. They do not remove the need for laboratory validation, clinical testing or manufacturing controls.
The central significance is the use of AI to make biological search more tractable. Protein sequences encode information that is difficult to interpret experimentally at large scale. A prediction system can prioritize candidate enzymes or binding sites for laboratory testing, potentially reducing the number of experiments needed before researchers reach a useful design. The reported 34-million-protein analysis is therefore relevant as an example of computational triage, although the article does not provide enough methodological detail to assess its accuracy.
The reported materials work points to a second practical pathway: using predictions about enzymes and metabolism to redesign microorganisms as production systems. That approach could support biological manufacturing of chemicals or materials that otherwise depend on petroleum-based processes. ChosunBiz reports that the motivation included South Korea’s limited crude-oil resources and the need to find alternatives to petrochemical production. The reported plastic output is a concrete result, but output volume alone does not establish durability, cost competitiveness, energy efficiency or environmental benefit.
The drug-interaction system addresses a different bottleneck. Large numbers of possible drug combinations make exhaustive experimental testing difficult, and an AI system could help identify combinations that deserve closer pharmacological review. That could be useful for clinicians and drug developers, particularly when patients take multiple medicines. But a predicted interaction is not the same as a confirmed adverse reaction, and the article does not say whether the system has been validated prospectively in patients or incorporated into clinical decision-making.
Lee’s broader argument about combining AI with bioengineering also has institutional implications. Research programs may increasingly need expertise spanning machine learning, molecular biology, chemical engineering, medicine and data governance. The potential public benefit is faster discovery of treatments and biological materials; the corresponding risk is that confident predictions could be mistaken for established biological facts. The report supports interest in the direction, but not a conclusion that the systems have already changed clinical care or industrial production at scale.
Qué ver a continuación
The key unanswered questions concern validation, reproducibility and deployment. ChosunBiz does not provide the underlying paper, datasets, model details, error rates, independent replication or regulatory status for the systems described. Future reporting should establish how accurately DeepEC predicts functions outside its training data, whether the drug-interaction predictions improve clinical decisions, and whether microorganism-based production can be scaled safely and economically.
The first verification priority is DeepEC’s evaluation design. ChosunBiz reports improved accuracy after transformers were applied, but does not state the baseline accuracy, test-set composition, false-positive and false-negative rates, or how the model handled proteins unlike those in its training data. Independent testing on held-out proteins and laboratory confirmation of predicted enzyme functions would help determine whether the reported speedup translates into reliable research value.
The drug-safety claims require similarly careful scrutiny. Future evidence should distinguish predictions from confirmed interactions, identify the data used for training and validation, and show whether the system performs better than existing pharmacovigilance methods. The article does not identify the 113 reaction categories in detail or explain how the proposed alternative drugs were assessed. Clinical use would also require review of patient characteristics, dosage, timing and other factors that can affect adverse reactions.
The microorganism-production results should be followed through the full manufacturing chain. Important questions include whether the 180-gram-per-liter figure was achieved repeatedly, how much energy and feedstock were required, whether the plastic meets relevant biodegradability standards, and whether the engineered organisms remain contained. For spider silk and other materials, readers would need data on strength, consistency, purification and cost before treating the work as an industrial breakthrough.
Finally, the report should be placed against publicly available primary evidence. ChosunBiz does not cite a research paper, dataset, code repository, regulatory filing or independent collaborator in the supplied text. That does not invalidate the account, but it limits what can be confirmed from the source alone. Further reporting should seek the underlying studies and clarify which claims describe completed experiments, which describe prediction systems, and which represent Lee’s expectations for future applications.
