The disease and gene-analysis artificial intelligence (AI) model “Exaone PASS” version 2.5, unveiled late last year by LG AI Research, has been found to be at a top-tier level compared with major global AI models.

According to LG on the 10th, Exaone PASS 2.5 recorded the highest accuracy of 76.75% in a recent cancer diagnosis performance evaluation of major pathology AI models conducted by LG AI Research. LG AI Research compared Exaone PASS 2.5 with major open-source medical AI models. Using clinical data such as colorectal cancer and lung adenocarcinoma from hospitals in Korea and the United States, the institute assessed how accurately each model detects tumors and gene mutations.

The comparison AI models included TITAN and UNI2-h developed by Professor Faisal Mahmood’s research team at Harvard Medical School, Gigapath by Microsoft, and H-optimus-0 by French AI company Biooptimus. Following Exaone PASS 2.5, UNI2-h (76.16% accuracy), H-optimus-0 (75.78%), TITAN (73.20%), and Gigapath (71.43%) were ranked in that order.


LG also emphasized that when Exaone PASS 2.5 was evaluated based on the pathology AI performance metrics developed by the Mahmood research team, it demonstrated competitiveness by achieving an accuracy of 69.8%, second only to the top-ranked TITAN model (71.4%). An LG official explained, “Exaone PASS 2.5 is specifically designed so that AI can learn even when pathology images and multi-omics (genetic information) are not linked to each other, thereby maximizing learning efficiency.”

Typically, for AI to diagnose diseases, data are required in which microscope-based pathology images are matched with genetic information. However, Exaone PASS 2.5 can autonomously discover correlations and learn even without explicit linkage between the two types of information. As a result, LG stated, the model delivers high accuracy and performance despite being relatively small. In terms of parameters, which determine AI model size, Gigapath and UNI2-h have 1.221 billion and 681 million parameters, respectively, while Exaone PASS 2.5 has 50 million, a difference of roughly 10 to 24 times. This indicates that the model achieves high efficiency with a much lighter configuration.

LG expects Exaone PASS to play an important role in the development of personalized treatment methods going forward. Jang Jong-seong, Head of the Bio Intelligence Lab at LG AI Research, said, “Exaone PASS is an AI that can identify the exact locations where actual gene mutations are occurring, reducing the time required for genetic testing from more than two weeks to about one minute,” adding, “It will greatly help secure the golden time for treatment by enabling preemptive therapy before cancer spreads.”

Popular News