AI Model Developed to Generate Anticancer Drugs by Analyzing Cancer Patient Genotypes

GIST Develops World's First AI Model
for Personalized Anticancer Drug Design
Based on Cancer Genotype Analysis

(From left) Nam Hojeong, Professor of Electrical, Electronics and Computer Engineering at GIST, Hyunho Kim, PhD.

Gwangju Institute of Science and Technology (GIST) announced on July 3 that a research team led by Professor Nam Hojeong of the Department of Electrical, Electronics and Computer Engineering has developed the world's first generative artificial intelligence (AI) model that analyzes the genotype of cancer patients to propose personalized anticancer drug candidates.

The AI model developed by the research team learns genotype information, which varies for each cancer cell, along with drug response data, to generate new anticancer drug candidates optimized for individual patients. This advancement not only enables patient-tailored precision medicine but also offers new solutions for intractable cancers that do not respond well to existing treatments.

Until now, research on generative AI-based anticancer drug development has faced several limitations. In complex diseases such as cancer, therapeutic targets are often unclear, which has limited the effectiveness of generated drugs. Additionally, many studies have relied on specialized data that are difficult to obtain in clinical settings, reducing their practical applicability.

To overcome these limitations, the research team developed a generative AI model called 'G2D-Diff' that was trained on approximately 1.5 million chemical structures and 1.2 million drug response data points. This model automatically designs anticancer drug candidates optimized for a given set of genetic information (such as mutations and copy number variations) and a target drug response level, both of which can be obtained in real clinical settings.

G2D-Diff operates in a manner similar to AI models that generate text-to-image outputs. When given the condition "a drug highly sensitive to a specific cancer genotype," it generates a molecular structure of an anticancer agent that meets this requirement.

This model consists of three main components: a 'Chemical Variational Autoencoder (Chemical VAE)' that numerically represents molecular structures; a 'Condition Encoder' that quantifies input conditions such as genotype and target drug response; and a 'Conditional Diffusion Model' that generates new molecular structures in accordance with the specified conditions.

Research on the Feasibility of Applying G2D-Diff

G2D-Diff demonstrated overwhelming performance across all metrics compared to existing generative AI models. In particular, when compared to IBM's 'PaccMannRL', known as the highest-performing model, G2D-Diff outperformed it in diversity, feasibility, and condition fitness.

In the 'condition fitness' category, which evaluates how well the generated compounds meet the input genotype conditions, G2D-Diff achieved an average error rate of about 1% in drug response prediction, whereas existing models showed an average error rate of about 51%.

The research team validated the practical applicability of the G2D-Diff model by applying it to triple-negative breast cancer, a representative case of intractable cancer. The candidate compounds generated by inputting patients' genetic mutation information accurately targeted key proteins involved in cancer cell proliferation, such as PI3K, HDAC, and CDK.

Furthermore, the chemical structures of these compounds were entirely different from those of existing therapies, yet they could achieve the same therapeutic effects. Computer docking simulations also confirmed that these compounds could indeed bind to the target proteins in cancer cells.

This is a groundbreaking achievement, demonstrating that AI can go beyond merely mimicking existing drugs to design entirely new therapies optimized for the unique genetic characteristics of each patient.

Another strength of G2D-Diff is its 'interpretability.' Utilizing an attention mechanism, the model can identify which genes or biological pathways are important for drug design in cancers with specific genotypes. This means the model not only generates new molecules but also provides scientific evidence supporting the validity of the therapy by explaining, at the genetic and biological pathway level, why a given molecule is effective.

Professor Nam Hojeong stated, "This research opens up new possibilities for personalized medicine, and we expect that AI technology will offer new hope to patients with intractable cancers."

Hyunho Kim, PhD (first author), emphasized, "G2D-Diff can dramatically improve the efficiency of the initial candidate discovery stage, which is the most challenging part of new drug development, and thus significantly shorten the time required for anticancer drug development."

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