Pre-trained on 11 billion conformers, the most reported for any molecular model, ZAO is now available for researchers to test immediately in their own environments with their own data
Tokyo, 10 November 2026 — SyntheticGestalt KK (headquarters: Minato-ku, Tokyo; CEO: Koki Shimada) has released ZAO, its flagship molecular foundation model, on Hugging Face. ZAO was pre-trained on 11 billion conformers (the three-dimensional shapes a compound can take), the largest number used by any molecular model reported to date.*1 Researchers who accept the licence can download ZAO immediately and evaluate its performance in their own computing environments, using their own data.*2
Hugging Face is a platform for sharing AI models and datasets, hosting more than 3 million models.*3 SyntheticGestalt has also released scikit-learn-compatible tools that make ZAO easy to try. ZAO is designed for research and development across a broad range of small molecules, including pharmaceuticals, and SyntheticGestalt provides it to researchers at companies and research institutions.
In a computational evaluation the company conducted for drug discovery across 31 target proteins, including GPCRs and protein–protein interactions (PPIs), ZAO was used to pick out, from a vast pool of compounds, candidates with scaffolds different from those of known active compounds. The hit rate reached a median of 26.4% across the target proteins, compared with 16.0% for an existing combination of methods using Boltz-2.*4 The hit rate is the share of selected compounds that calculations predicted would bind strongly to the target protein.
Underpinning these results is ZAO's core technology, the "4D Feature." To work with compounds on a computer, researchers describe them numerically as "features." The widely used 2D fingerprints carry no information about a compound's conformers, which strongly influence its properties. As a result, fingerprint-based searches have difficulty finding compounds that bind to the same target protein as a known active compound but have different scaffolds. Approaches that do use conformers face a further limitation: a compound can adopt many conformers, and no single one represents it adequately. ZAO overcomes this limitation by consolidating the multiple conformers a compound can adopt into a single feature. Because it brings together multiple three-dimensional conformers, SyntheticGestalt calls it a "4D Feature."
By combining the features ZAO outputs with in-house experimental data, users can also build their own models that predict binding affinity (how strongly a compound binds to its target protein), as well as solubility, membrane permeability, whether a compound is toxic, and other properties.*2
"Molecular foundation models are still rare in the industry, and AI that truly understands small molecules is still a work in progress," said Koki Shimada, CEO and CTO of SyntheticGestalt. "Through this release, we hope to accelerate the development of molecular foundation models across the industry. I look forward to seeing researchers test ZAO on their own data and to discussing the results with them, good and bad alike."
About ZAO
ZAO is a molecular foundation model offered by SyntheticGestalt. Its weights are available on Hugging Face under the SyntheticGestalt Technical Verification and Academic Research License (SG-TVAR-1.0).
Hugging Face: https://huggingface.co/syntheticgestalt/zao
Product page: https://syntheticgestalt.com/zao
Enquiries: info@syntheticgestalt.com
About SyntheticGestalt
SyntheticGestalt is a Tokyo-based molecular AI company. It develops and deploys molecular AI across the pharmaceutical, chemical, materials, agrochemical, cosmetics, and food industries, and since 2018 has pursued the goal of building systems that mass-produce inventions with AI. In its research projects the company has discovered drug lead compounds and low-environmental-impact materials while cutting discovery cost and time by up to 90% versus conventional approaches.*5 Learn more at https://syntheticgestalt.com.
*1 12.4 times the 884 million conformers used to pre-train Uni-Mol2, the largest reported by others (SyntheticGestalt survey, October 2026).
*2 Free for technical verification and for academic research at universities and public research institutions. Companies may use it for technical verification only by comparing outputs with existing data. Outputs, and models built with them, are subject to the same limits.
*3 Models listed at https://huggingface.co/models, 30 October 2026.
*4 Boltz-2 (AI binding prediction) plus ESPSim (3D similarity). Hits: −10 kcal/mol or lower in absolute binding free energy calculations. Selection procedures were compared, searching different numbers of compounds. ZAO-based selection led on 22 of the 30 target proteins compared.
*5 Past SyntheticGestalt projects; results vary by target and conditions.
Media Enquiries
SyntheticGestalt KK
Sou Minakami, CEO Office
s.minakami@syntheticgestalt.com