LG AI Research 535

Recently, in the fields of chemistry and biology, research on structure prediction and molecular design using AI has been actively progressing. LG AI Research, through joint research with LG H&H, succeeded in developing cosmetic efficacy materials with improved solubility and safety.

The core of this achievement is EXAONE Discovery, an AI model specialized in new substance discovery. With the existing research method, selecting candidate substances took approximately 2 years, but using EXAONE Discovery, it can be performed in just one day.

The research that formed the basis of this study, "3D Denoisers are Good 2D Teachers," was selected as an Oral paper at AAAI-25. This research uses a Cross-Modal Distillation strategy that transfers knowledge from a Teacher model with three-dimensional (3D) information to a Student model in the form of a two-dimensional (2D) Graph.

The key features of the research are as follows:

• Utilization of 3D Information: The Teacher model is trained using rich 3D coordinate data.
• 2D Graph Efficiency: To solve the problem of physical limitations in acquiring large amounts of 3D data, a Student model was built that can effectively extract 3D information using only 2D structures.
• Versatility and Scalability: Prediction is possible immediately once the atomic composition of a molecule is known, and fine-tuning on small-scale property data is easy.

This technology aims to harmoniously combine AI's Inductive approach with the Deductive methodology of basic science, ensuring high predictive power even in environments with insufficient data.