Lunit and the Korea Advanced Institute of Science and Technology (KAIST), a first-round consortium in the specialised AI foundation model programme, are speeding up development ahead of a final evaluation on Sept. 9.

A specialised foundation model is an AI model optimised for a specific industry by training on data from sectors such as healthcare, bio, manufacturing, finance and defence. Unlike an independent AI foundation model, which aims to develop a general-purpose AI, it focuses on deployment in industrial settings.

The Ministry of Science and ICT last year first recruited participants in the specialised foundation model programme for the healthcare and bio sector. Eighteen consortia applied, and the Lunit-KAIST consortium was selected. The two consortia began development on Nov. 1 last year.

An interim evaluation was held on March 31 this year. Both consortia scored at least 80 points, above the 70-point threshold for phase-two support, and received an extension of GPU support through September.

Lunit expands pilot tests to 9 hospitals, targets medical science AI platform

The Lunit consortium aims to develop a full-cycle medical science AI foundation model. The core is to build a base model spanning basic science research, new drug development, clinical trials and clinical decision support.

In the interim evaluation, Lunit’s 16B-class model recorded strong performance on some benchmarks, including medical Q&A accuracy, consistency between AI answers and supporting evidence, and parts of scientific code writing and analysis assessments, outperforming very large models in the 100B to 1T range such as Anthropic’s Claude 3.5 Sonnet.

Phase-two development is under way to confirm, in real clinical settings, the usefulness of the foundation model validated in phase one. The consortium has expanded proof-of-concept testing to nine hospitals nationwide to verify that the model works in multiple hospitals.

Lunit said it is also upgrading the model by using data accumulated through the pilot process for additional training. Because AI algorithms become more accurate and efficient as data builds up, Lunit’s task is to ensure the model operates reliably in real hospitals.

Pilot tests are being carried out separately with public and private partners and medical institutions. It created favourable conditions to upgrade the model around the National Health Insurance Service, which it described as top-tier in Korea in data volume and quality, and Severance Hospital.

Lunit also plans to expand existing services for diagnosing individual diseases into a platform similar to an operating system for medical information systems. Its goal is to build a base model used by medical institutions nationwide and create an ecosystem that can continuously expand medical AI services.

A Lunit official said it developed existing disease-diagnosis apps such as finding lung cancer in X-ray images or detecting breast cancer through mammography. The official said it would expand into a platform that covers other diseases beyond lung and breast cancer, spanning diagnosis, treatment decisions and post-treatment patient management.

Building trust among medical staff is cited as a remaining challenge. Confidence and trust in AI must accumulate among clinicians for it to translate into actual use. A Lunit official said that until 1 to 2 years ago some clinicians believed they were more accurate than AI, but recently the atmosphere has shifted to acceptance without resistance. The official said healthcare is relatively conservative, but adoption could spread faster once perceptions change. The official added that scientifically proving accuracy and efficiency is key.

K-Fold scales up to prepare for commercial rollout

The KAIST consortium is developing a bio-specialised foundation model called K-Fold. It aims to predict the structures of proteins and molecular complexes quickly and accurately. In drug development, it can be used to screen in advance by computer whether a specific compound binds well to a target protein, reducing the time and cost needed to find drug candidates.

While existing models such as Google DeepMind’s AlphaFold3 and Boltz2 rely on statistical patterns in accumulated data, K-Fold adopts an approach in which AI learns on its own the principles of physical and chemical interactions that occur inside proteins.

In the interim evaluation, the consortium said it secured accuracy close to Google DeepMind’s AlphaFold3 and cut inference time by up to more than 30 times.

In phase two, it will scale its existing 2B-sized model to 7B. It is also pushing to build a platform-as-a-service applicable to drug development. The consortium said phase one verified the model’s accuracy and speed, while phase two expands scale and turns it into a service that can be used in real drug development settings.

It has also prepared commercialisation plans, taking into account the nature of state-led AI model development projects that can easily remain at the laboratory stage.

KAIST spin-off HITS plans to provide K-Fold as software-as-a-service through a web-based platform called HyperLab that can be used immediately without installation. For institutions where security is important, KAIST startup Atolab will build and provide HyperLab on dedicated in-house servers or on-premise systems.

Merck Life Science plans to apply K-Fold to its digital lab tools platform to support use by more than 30,000 laboratories worldwide. The KAIST consortium said it is significant that a Korean model developed under the specialised foundation model programme has been deployed on an overseas company’s platform.

A consortium official said it improved performance in predicting important protein structural changes in drug development by raising prediction accuracy even for new drug complexes where data is limited. The official said it is meaningful in that it confirmed the potential of a service-type drug development platform based on K-Fold.