GPU-accelerated conformer generation lets ZAO learn molecules in 4D at scale, and an AI agent will put it to work toward automating invention with AI.
TOKYO, June 26, 2026 — SyntheticGestalt today announced that it trains ZAO, its 4D molecular foundation model, using NVIDIA nvMolKit, and that it plans to adopt the NVIDIA BioNeMo Agent Toolkit to power an AI agent, advancing its goal of automating invention with AI.
Most molecular AI reads a molecule as a one-dimensional string or a two-dimensional graph. ZAO instead learns its three-dimensional shape, across the many conformations the molecule can adopt. The result is a 4D view that keeps the geometry flat representations cannot see, the geometry behind how a molecule binds its target, crosses a membrane, even packs into a crystal.
“A molecule’s three-dimensional form is what decides how it behaves, and flat 1D and 2D representations simply don’t carry it,” said Koki Shimada, CEO of SyntheticGestalt. “ZAO is built to capture that form, and being able to compute the 3D structure of enormous numbers of molecules at high speed is what lets us do it at scale. Building on this, we will bring molecular AI into everyday practice and move toward a future in which AI mass-produces inventions.”
Training a foundation model on 3D structure means generating many high-quality conformers for hundreds of millions of compounds, which was impractical on CPUs. By running conformer generation and optimization on GPUs with NVIDIA nvMolKit, SyntheticGestalt generated roughly 10 billion conformers to pre-train ZAO. In SyntheticGestalt’s benchmarks on an 8-GPU NVIDIA Hopper node, nvMolKit produced conformers about 20 times faster than the same node’s 224 CPU threads running RDKit, at roughly 18,000 conformers per second.
Given a molecule, ZAO generates and optimizes its multiple conformers, then turns the resulting 4D structure into a single embedding. Teams in pharmaceutical, agrochemical, and small-molecule materials research can use that embedding to predict properties like activity, ADMET, and binding affinity, often with a lightweight downstream model trained on very little labeled data.
SyntheticGestalt also plans to integrate ZAO with large language models and to develop an AI agent built on the NVIDIA BioNeMo Agent Toolkit, which includes nvMolKit. Instead of running each step by hand, a scientist would ask the agent to screen compounds, find molecules with a similar 3D shape, or train a predictor on their own measurements, and it would handle the underlying molecular work and feed each result into the next design cycle.
A member of NVIDIA Inception, SyntheticGestalt has worked closely with NVIDIA on nvMolKit. SyntheticGestalt put nvMolKit into production, benchmarked it at scale, and shared issues and feature requests backed by reproduction code; NVIDIA acted on that feedback, and requested capabilities have since shipped as nvMolKit features.
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. Learn more at https://www.syntheticgestalt.com.
