ZAO

ZAO is a molecular foundation model from SyntheticGestalt that converts a compound's SMILES string into a fixed 2048-dimensional numerical embedding. Its defining choice is to encode each molecule as an ensemble of multiple three-dimensional conformations rather than a single flat representation. SyntheticGestalt describes ZAO as a "4D" model because it captures a molecule's 3D geometry together with its conformational flexibility — spatial and dynamic information that one-dimensional string representations and two-dimensional graph representations discard.

The problem ZAO addresses is the need for reusable, physically informed molecular representations in drug discovery. Rather than being trained end-to-end for a single endpoint, ZAO produces general-purpose embeddings that serve as drop-in input features for downstream machine-learning models — including absorption, distribution, metabolism, excretion and toxicity (ADMET) prediction, drug–target interaction modeling, and virtual screening. The model is scoped to drug-like chemical matter, accepting compounds with molecular weight between 100 and 1000 Da and returning null embeddings for molecules outside that range.

ZAO is developed by SyntheticGestalt, an AI life-science company founded in 2018 with operations in London and Tokyo. The model is offered as a managed inference endpoint through the AWS Marketplace (Amazon SageMaker) and Google Cloud's Vertex AI Model Garden, where users submit SMILES and receive embeddings from a frozen pretrained checkpoint. ZAO sits within a broader SyntheticGestalt effort in large-scale molecular pretraining that also includes SG4D10B, a related model trained on up to 10 billion compounds presented at NVIDIA GTC 2025 and through Japan's GENIAC program. ZAO has no peer-reviewed publication; its documentation is provided through the company's marketplace listings and a public examples repository.

#Key Features

• Conformation-aware representations: ZAO encodes multiple 3D conformations per molecule to capture spatial arrangement and conformational flexibility, distinguishing it from models that operate only on 1D SMILES strings or 2D molecular graphs.
• Fixed 2048-dimensional embeddings: Each SMILES input yields a 2048-dimensional float32 vector intended as a drop-in feature for any downstream predictor, decoupling representation learning from task-specific modeling.
• Deterministic outputs: Embeddings are reproducible for a given SMILES, with conformer generation using fixed random seeds so the same input always produces the same vector.
• Drug-like scope: The model targets compounds with molecular weight of 100–1000 Da and returns null embeddings for inputs outside this range, keeping it focused on small-molecule therapeutics.
• Managed cloud deployment: ZAO is served as a frozen inference endpoint on AWS SageMaker and GCP Vertex AI, so teams consume embeddings without hosting or fine-tuning the model themselves.

#Technical Details

ZAO takes a SMILES string, generates a set of 3D conformers with fixed seeds, and emits a single 2048-dimensional float32 embedding summarizing the molecule's conformational ensemble; the underlying architecture and parameter count are not publicly disclosed. The published inference endpoint requires a minimum of 24 GB of GPU VRAM and 32 GB of system memory. On the Therapeutics Data Commons (TDC) benchmarks, ZAO-derived predictors reached the top of the ADMET leaderboard on 8 of 22 datasets — spanning clearance, solubility, lipophilicity, CYP metabolism, and toxicity — and improved the TDC DTI BindingDB_Patent task by +0.099 Pearson correlation over the prior leaderboard leader. The model itself is distributed only as a proprietary marketplace endpoint; the MIT license applies to the accompanying examples repository, not to the model weights.

#Applications

ZAO is aimed at computational and medicinal chemistry teams building predictive models for small-molecule drug discovery. By featurizing compound libraries into fixed embeddings, users can train lightweight downstream models for ADMET and toxicity prediction, drug–target interaction, and prioritization of candidates in virtual screening campaigns, without maintaining their own pretraining pipeline. Because ZAO is delivered through the AWS and Google Cloud marketplaces, it integrates into existing cloud-based discovery workflows and is accessible to groups that lack the infrastructure to train large molecular foundation models from scratch.

#Impact

ZAO reflects a broader move toward conformation-aware molecular foundation models delivered as managed inference services rather than downloadable checkpoints. Its distribution through major cloud marketplaces lowers the barrier to adopting pretrained molecular representations in industrial discovery settings, and its reported TDC ADMET and DTI results position it as a competitive general-purpose featurizer. The model's reach is bounded by clear limitations: the weights are not openly available, there is no peer-reviewed publication or associated DOI, and its applicability is restricted to drug-like compounds in the 100–1000 Da range. These constraints make ZAO a practical tool for applied screening pipelines while limiting its use for independent reproduction or academic benchmarking.