SG4D100M

SG4D100M is a molecular embedding model from SyntheticGestalt that takes a SMILES string and returns a single numeric vector describing the compound. It is not trained end to end against any one endpoint; the embedding itself is the product, intended as an input feature for downstream predictors of properties such as solubility, toxicity, and binding affinity, and as a coordinate system for similarity search and compound clustering.

Its defining design choice is SE(3) invariance. The special Euclidean group in three dimensions covers rotations and translations but excludes reflections, so a structure presented in any rigid orientation encodes to the same vector while its mirror image does not. SyntheticGestalt states the consequence directly: enantiomers are embedded at different points in the space. Since stereochemistry routinely separates an active drug from an inactive or harmful counterpart, a representation that collapses mirror-image pairs throws away information that matters to the chemistry, and this encoder is built so that it does not.

The model belongs to SyntheticGestalt's SG4D line of molecular pretraining, which the company frames as four-dimensional: three-dimensional geometry plus the range of conformational states a molecule can adopt. ZAO, the company's conformer-ensemble embedding model, shares that framing but exposes a different interface and reports Therapeutics Data Commons results that SG4D100M does not. SG4D10B, trained on 10 billion compounds under Japan's GENIAC program, is the line's largest member. SG4D100M is distributed as an Amazon SageMaker model package on the AWS Marketplace at version v0.1, labelled an experimental release. No paper or preprint describes it.

#Key Features

• SE(3)-invariant encoding: Rigid rotations and translations of the same structure map to one representation, so the embedding tracks molecular geometry rather than the arbitrary frame the coordinates were written in.
• Chirality preserved: Reflections fall outside SE(3), so enantiomers occupy separate points in the embedding space instead of collapsing onto a shared vector.
• Conformation-aware representation: The model handles multiple conformational states rather than a single fixed pose, extending a static 3D picture with the flexibility a molecule shows in practice.
• Make-on-demand chemical space in pretraining: The corpus draws on Enamine REAL Space and ZINC, so coverage extends past catalog compounds into synthesizable virtual libraries.
• CPU-class managed inference: The published deployment runs on a single general-purpose ml.m5.xlarge instance and supports both real-time endpoints and S3 batch transform.

#Technical Details

SyntheticGestalt describes the architecture as an SE(3)-invariant transformation followed by pretrained transformers, trained through a multi-stage pipeline; the marketplace listing adds that the pipeline draws on persistent homology and spans 10 billion molecular data points. The documented training corpus is over 100 million molecules from Enamine REAL Space and ZINC. Inference runs over newline-delimited JSON under the application/jsonlines content type: each input record carries a smiles field, and each output record returns that field alongside an sg4d100m field holding a list of float64 values. Parameter count, embedding dimensionality, and layer configuration are not disclosed, and no benchmark results have been published for the model — the accuracy and generalization claims on the listing carry no accompanying numbers, leaving SG4D100M unevaluated in public.

#Applications

SG4D100M is aimed at teams that want pretrained molecular features without running their own pretraining. Computational and medicinal chemists can featurize a compound library once and fit lightweight downstream models for solubility, toxicity, or binding affinity on top of the resulting vectors, then reuse the same embedding space for nearest-neighbour searches and for clustering screening decks by structural and functional similarity. SyntheticGestalt targets the model beyond pharmaceuticals, at specialty chemical, agricultural chemical, and food ingredient development, where labelled data is scarce and a general-purpose representation carries more of the modelling load. Delivery as a SageMaker model package puts it inside existing AWS discovery pipelines, and the free software subscription leaves only AWS infrastructure charges.

#Impact

SG4D100M is an early, small-scale entry in SyntheticGestalt's molecular pretraining effort, and its significance is as much about distribution as architecture: a chirality-sensitive molecular featurizer that a chemistry team can call from a CPU endpoint without weights, GPUs, or a training budget. The constraints are real and worth stating plainly. It is a v0.1 experimental release; the weights are not downloadable and access runs only through the managed endpoint; the accompanying example repositories carry no license; and with no paper, DOI, or published evaluation, none of the vendor's accuracy claims can be independently checked or reproduced. That makes it a candidate for internal benchmarking against an in-house baseline rather than a settled tool, and it leaves the SG4D line's public track record resting on its siblings.