KAIST

< (From left) Ph.D candidate Wonho Zhung, Ph.D cadidate Joongwon Lee , Prof. Woo Young Kim , Ph.D candidate Jisu Seo >

Traditional drug development methods involve identifying a target protin (e.g., a cancer cell receptor) that causes disease, and then searching through countless molecular candidates (potential drugs) that could bind to that protein and block its function. This process is costly, time-consuming, and has a low success rate. KAIST researchers have developed an AI model that, using only information about the target protein, can design optimal drug candidates without any prior molecular data—opening up new possibilities for drug discovery.

KAIST (President Kwang Hyung Lee) announced on the 10th that a research team led by Professor Woo Youn Kim in the Department of Chemistry has developed an AI model named BInD (Bond and Interaction-generating Diffusion model), which can design and optimize drug candidate molecules tailored to a protein’s structure alone—without needing prior information about binding molecules. The model also predicts the binding mechanism (non-covalent interactions) between the drug and the target protein.

The core innovation of this technology lies in its “simultaneous design” approach. Previous AI models either focused on generating molecules or separately evaluating whether the generated molecule could bind to the target protein. In contrast, this new model considers the binding mechanism between the molecule and the protein during the generation process, enabling comprehensive design in one step. Since it pre-accounts for critical factors in protein-ligand binding, it has a much higher likelihood of generating effective and stable molecules. The generation process visually demonstrates how types and positions of atoms, covalent bonds, and interactions are created simultaneously to fit the protein’s binding site.

<Figure 1. Schematic of the diffusion model developed by the research team, which generates molecular structures and non-covalent interactions based on protein structures. Starting from a noise distribution, the model gradually removes noise (via reverse diffusion) to restore the atom positions, types, covalent bond types, and interaction types, thereby generating molecules. Interacting patterns are extracted from prior knowledge of known binding molecules or proteins, and through an inpainting technique, these patterns are kept fixed during the reverse diffusion process to guide the molecular generation.>

Moreover, this model is designed to meet multiple essential drug design criteria simultaneously—such as target binding affinity, drug-like properties, and structural stability. Traditional models often optimized for only one or two goals at the expense of others, but this new model balances various objectives, significantly enhancing its practical applicability.

The research team explained that the AI operates based on a “diffusion model”—a generative approach where a structure becomes increasingly refined from a random state. This is the same type of model used in AlphaFold 3, the 2024 Nobel Chemistry Prize-winning tool for protein-ligand structure generation, which has already demonstrated high efficiency.

Unlike AlphaFold 3, which provides spatial coordinates for atom positions, this study introduced a knowledge-based guide grounded in actual chemical laws—such as bond lengths and protein-ligand distances—enabling more chemically realistic structure generation.

<Figure 2. (Left) Target protein and the original bound molecule; (Right) Examples of molecules designed using the model developed in this study. The values for protein binding affinity (Vina), drug-likeness (QED), and synthetic accessibility (SA) are shown at the bottom.>

Additionally, the team applied an optimization strategy where outstanding binding patterns from prior results are reused. This allowed the model to generate even better drug candidates without additional training. Notably, the AI successfully produced molecules that selectively bind to the mutated residues of EGFR, a cancer-related target protein.

This study is also meaningful because it advances beyond the team’s previous research, which required prior input about the molecular conditions for the interaction pattern of protein binding.

Professor Woo Youn Kim commented that “the newly developed AI can learn and understand the key features required for strong binding to a target protein, and design optimal drug candidate molecules—even without any prior input. This could significantly shift the paradigm of drug development.” He added, “Since this technology generates molecular structures based on principles of chemical interactions, it is expected to enable faster and more reliable drug development.”

Joongwon Lee and Wonho Zhung, PhD students in the Department of Chemistry, participated as co-first authors of this study. The research results were published in the international journal Advanced Science (IF = 14.1) on July 11.

●        Paper Title: BInD: Bond and Interaction-Generating Diffusion Model for Multi-Objective Structure-Based Drug Design

●        DOI: 10.1002/advs.202502702

This research was supported by the National Research Foundation of Korea and the Ministry of Health and Welfare.