KAIST: "AI Automatically Designs Optimal Drug Candidates... Set to Change Drug Discovery Paradigm"

An artificial intelligence model capable of automatically designing optimal drug candidates has been developed in South Korea. This model is expected to enable faster and more accurate drug discovery.

KAIST announced on August 10 that a research team led by Professor Woo-yeon Kim from the Department of Chemistry has developed an AI model with these capabilities, named 'BInD'.

(From left) Wonho Jung, Integrated MS-PhD Program, Joongwon Lee, Integrated MS-PhD Program, Woo-yeon Kim, Professor, Jisoo Seo, Integrated MS-PhD Program. Provided by KAIST

The core feature of BInD is 'simultaneous design'. Without any prior information about the drug candidate molecule, BInD can optimize both the molecular structure of the drug candidate and its binding mode (non-covalent interactions) with the protein, using only the protein's structure.

Whereas conventional AI models could either generate molecules or separately evaluate whether a generated molecule binds to a protein, BInD allows for the simultaneous design of both the molecule and its binding mode with the protein in a single process.

In addition, BInD incorporates key factors important for binding to proteins in advance, increasing the likelihood of generating effective and stable molecules. This generative process visually represents how, in a single step, the types and positions of atoms, covalent bonds, and interactions are simultaneously created to match the protein's target site.

Another advantage of BInD is that it is designed to satisfy multiple essential factors for drug design at once, such as molecular stability, physicochemical properties, and structural naturalness. Previously, models often focused on one or two objectives at the expense of others, but BInD balances a variety of conditions, greatly enhancing its practical utility.

BInD operates based on a 'diffusion model', which gradually draws more refined structures from a random state. Diffusion models have proven highly efficient, as demonstrated in the protein-drug structure generation of 'AlphaFold3', which was awarded the Nobel Prize in Chemistry last year.

Furthermore, the research team incorporated knowledge-based guides into BInD, such as 'bond length' and 'protein-molecule distance', to ensure that the generated structures adhere to real chemical principles. This allows BInD to produce more realistic results, distinguishing it from AlphaFold3, which focuses on predicting the spatial coordinates of atoms.

A schematic diagram of the diffusion model generating molecular structures and non-covalent interactions based on protein structures developed by the research team. Provided by KAIST

The research team also applied an optimization strategy that reuses the best binding patterns found among the generated results. As a result, they were able to produce outstanding drug candidates without additional training, and notably succeeded in generating molecules that selectively act on mutations of the cancer-related target protein EGFR.

This study is also significant in that it advances previous AI models, which required input conditions specifying 'which molecule binds to which protein and how', by allowing for more autonomous and flexible design.

Professor Woo-yeon Kim stated, "The AI 'BInD' developed by our research team can learn and understand the key elements required for binding to target proteins, enabling it to design optimal drug candidates (molecules) that interact without any prior information. We expect this will redefine the paradigm of drug discovery and contribute to developing new drugs more quickly and accurately."

This research was conducted with support from the National Research Foundation of Korea and the Ministry of Health and Welfare. Joongwon Lee and Wonho Jung, doctoral students in the Department of Chemistry at KAIST, participated as co-first authors. The research results (paper) were also introduced in the international journal 'Advanced Science' on July 11.

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