Syntekabio Registers US Patent for 'AI Deep Learning Model-Based Genomic Biomarker Detection Method'

AI new drug development specialist Syntekabio announced on the 16th that it has filed a U.S. patent for a new method that precisely selects biomarkers based on genomic mutation data using an artificial intelligence (AI) deep learning model, and has received approval for registration.

This patent is titled “Biomarker detection method using an artificial intelligence deep learning model for population genome base sequences and variant transformed data.” It involves a method that precisely selects biomarkers from genomic mutation data using an AI deep learning model called Convolutional Neural Networks (CNN). The company named this detection method Hilbert-CNN, as it uses a fractal curve created by the German mathematician Hilbert.

The patent also includes detailed information on Syntekabio’s unique method of representing big data collected using next-generation sequencing (NGS) data as 2D and 3D images.

Hilbert-CNN is a technology that converts one person’s whole human genome 1D sequence information into a 2D Hilbert image, representing thousands of genomic data as thousands of images and analyzing them precisely using an AI model. It was registered as a U.S. patent as a new method for discovering non-linear biomarkers that cannot be found through traditional linear disease association studies.

Hilbert-CNN is characterized as a non-linear model that shows high precision and reproducibility in biomarker detection even with fewer samples compared to conventional methods. It can be applied to patient stratification by classifying groups according to disease presence, disease type, diagnostic results, and patient condition, companion diagnostics in drug clinical trials, and drug development directly targeting biomarkers.

While biomarker elements used by general AI tools for patient stratification were included in a black box and difficult to verify, Hilbert-CNN enables identification of biomarkers used in stratification through research and development of black box interpretation methods.

Using big data accumulated in U.S. Alzheimer’s disease (ADNI database) and Parkinson’s disease (PPMI database) databases to train patient genomic features, the Hilbert-CNN AI demonstrated patient classification accuracy exceeding 90%. It was confirmed that high accuracy is maintained even with a very small number of biomarkers. Since high reproducibility is essential for industrial use of biomarkers, multiple reproducibility tests were conducted, confirming no significant differences in accuracy and performance between experiments.

Conventional methods are based on statistical significance within experimental groups and the presence or absence of specific biomarkers, which may show high accuracy in single experiments but can experience a sharp decline in performance when new blind samples are applied. However, Hilbert-CNN maintains performance even when new blind samples are applied because it identifies common features of each group based on relationships between biomarkers (Biomarker Networks).

Jongseon Jeong, CEO of Syntekabio, said, “We have continuously researched applying CNN deep learning methods to the bio and pharmaceutical fields. We are pleased that the Hilbert-CNN detection method, developed after long research, has been recognized for its excellent results and originality, culminating in U.S. patent registration. We believe this technology will greatly aid biomarker-based personalized drug development.”

He added, “Since the patent registration has been approved in the U.S. as well as Korea, we plan to publish papers on various disease models and validation data used for obtaining the patent.”