
NEDO(New Energy and Industrial Technology Development Organization)
SyntheticGestalt KK
SyntheticGestalt Develops World's Largest Molecule-Specialized Foundation AI Model "SG4D10B" Through GENIAC Project
―Achieves World's Top Performance in Three Key Drug Discovery Benchmark―
SyntheticGestalt has developed the world's largest molecule-specialized foundation AI model "SG4D10B" through the "GENIAC" project, an initiative by NEDO and the Ministry of Economy, Trade and Industry aimed at strengthening generative AI development capabilities.
The development process for new molecules, including drug discovery and new material development, requires enormous costs and time to identify structures that may yield superior effects from complex combinations of molecular information.
The newly developed molecule-specialized foundation AI model "SG4D10B" has successfully learned vast structural patterns of compounds by utilizing 10 billion compound data points as training data, obtained through collaborative research between SyntheticGestalt and global compound suppliers. This foundation model has achieved world-leading performance in three critical indicators (toxicity, permeability, and stability) on global standard drug discovery benchmarks, and is expected to significantly contribute to streamlining development processes across a wide range of fields including not only drug discovery but also pesticides and new materials.

Figure 1: Characteristics of "SG4D10B"
1. Overview
The pharmaceutical industry faces urgent needs for efficiency improvements in new drug development processes. Current challenges include limited data volumes and molecular information complexity, resulting in insufficient accuracy for artificial intelligence (AI) systems developed specifically for molecular applications.

Figure 2: Challenges in AI Model Development within the Pharmaceutical Industry
SyntheticGestalt initiated research and development in October 2024, participating in GENIAC *1 Cycle 2 and receiving support from NEDO to successfully develop the world's largest molecule-specialized foundation AI model "SG4D10B *2." The utilization of SG4D10B is expected to reduce both costs and time required for experiments necessary in discovering new molecules.
2. Current Achievements
(1) Model Design and Construction
Molecules possess 3D structures from a physicochemical perspective and can adopt multiple conformations. The foundation AI model "SG4D10B" developed in this research employs innovative 4D technology. This 4D technology enables high-precision prediction of characteristics such as activity and toxicity while accounting for complex 3D molecular structures. Additionally, "SG4D10B" has successfully learned vast structural patterns of compounds by utilizing 10 billion compound data points as training data, obtained through collaborative research with Enamine, a global compound supplier. This foundation model achieved world-leading performance in three important indicators (toxicity, permeability, and stability) on global standard drug discovery benchmarks *3.
(2) Social Implementation
"SG4D100M *4," a compact model of "SG4D10B" pre-trained on 100 million compound data points, is publicly available free of charge on Google Cloud Marketplace and AWS Marketplace. Furthermore, activity data from 7,770 fragment compounds *5 obtained through actual wet laboratory testing has been released as evaluation data for AI models through SyntheticGestalt's public repository on GitHub. Currently, utilization is being promoted *6 with multiple domestic pharmaceutical companies, including Taiho Pharmaceutical Co., Ltd. These development technologies have received high international acclaim, with a presentation delivered at "NVIDIA GTC 2025," one of the world's largest AI conferences, held in San Jose, USA, in March 2025.
3. Future Plans
Building upon achievements from GENIAC, full-scale expansion into international markets is planned. Additionally, continued development of next-generation foundation models through further technological innovation will achieve dramatic improvements in accuracy. This will contribute to further streamlining development processes across a wide range of fields requiring cost and time reduction for experiments in new molecular discovery, including not only drug discovery but also cosmetics, pesticides, and new materials.
[Notes]
*1 GENIAC
A project aimed at strengthening domestic generative AI development capabilities, termed "GENIAC (Generative AI Accelerator Challenge)," which provides computational resource support for developing foundation models, the core technology of generative AI.
GENIAC Project: https://www.meti.go.jp/policy/mono_info_service/geniac/index.html
*2 SG4D10B
The world's largest molecule-specialized foundation AI model developed by SyntheticGestalt. It resolves issues of low prediction accuracy and generalization performance inherent in conventional molecular AI technologies, enabling high-precision model development even from small-scale data held by pharmaceutical companies and chemical manufacturers. It contributes to efficiency improvements in new drug and new material development through significant reductions in development periods and research costs.
*3 Drug discovery benchmarks
Therapeutics Data Commons (TDC) is standardly used in the AI drug discovery field as an indicator for evaluating AI model performance. TDC provides machine learning datasets, tasks, and leaderboards for performance evaluation of various indicators including toxicity, permeability, and stability.
*4 SG4D100M
SG4D100M is publicly available at the following links:
Google Cloud Marketplace: https://console.cloud.google.com/marketplace/product/syntheticgestalt-public/sg4d100m
AWS Marketplace: https://aws.amazon.com/marketplace/seller-profile?id=9af00667-4d52-4280-9f96-44124cf443cb
*5 Fragment compounds
Small, simple-structure compounds utilized in early-stage drug discovery. They are used for efficient candidate substance exploration.
*6 Promoting utilization with pharmaceutical companies
Examples include the following:
Taiho Pharmaceutical News Release (June 4, 2025): "Taiho Pharmaceutical and SyntheticGestalt Implements Technology Validation Towards Generative AI-Driven Expansion of Cysteinomix Drug Discovery Platform" https://www.taiho.co.jp/en/release/2025/20250604.html
Contact Information
NEDO: AI・Robot Department Generative AI team Team leader:Endo、Point person:Hagiwara TEL:044-520-5241
SyntheticGestalt KK: HR&PR Department TEL: 03-4570-8634
