https://bio.nikkeibp.co.jp/atcl/news/p1/25/10/22/13877/

SyntheticGestalt, a UK-based startup, develops and provides drug discovery support tools using AI specialized in molecular information. Founded in the UK by CEO Yuki Shimada in September 2018, the company subsequently established a Japanese subsidiary. The company owns a molecular-specific foundational AI model called “SG4D10B,” which has trained on information on 10 billion small molecular compounds.

The company creates generative AI using a method called “foundational modeling.” For example, when building a text-generation AI, a foundational model is a method of training a massive amount of text beforehand, and then building a specific AI on top of that, resulting in significantly improved performance. The company’s approach is to create a foundational model of 10 billion pieces of compound information and then develop a generative AI that designs small molecule compounds.

The unique feature of the molecular-specialized platform AI model SG4D10B is that it learns “four-dimensional compound information.” Molecules have three-dimensional structures, and this three-dimensional structure significantly affects activity and toxicity. Furthermore, there is no single set of three-dimensional structures; molecules vibrate and rotate, changing shape. The company trained the platform model by considering the multiple three-dimensional structures a single compound can adopt as a fourth dimension.

The company’s business model is to provide SG4D10B to pharmaceutical companies, chemical manufacturers, new materials research institutes, and others, earning usage fees and licensing revenue. A model for learning small molecule-protein interactions is also under development.