Nach01

Nach01 is the production generative chemistry foundation model from Insilico Medicine, released on the AWS Marketplace as an Amazon SageMaker algorithm in June 2025. It fuses two published research lines into one deployable system: nach0, a multi-domain encoder–decoder language model over scientific text and SMILES strings developed with NVIDIA, and nach0-pc, which extended that backbone with a molecular point-cloud encoder so the model can read three-dimensional atomic arrangements directly. Insilico's own name for it, "natural and chemical languages + point cloud transformer," describes the architecture precisely.

It addresses a structural limitation of chemical language models. Models that operate purely on SMILES or SELFIES strings never see the spatial arrangement of atoms, which governs binding, conformation, and selectivity. The obvious workaround — serializing coordinates into text formats such as XYZ or PDB — costs dozens of tokens per atom, becomes impractical for pockets and proteins, and discards bond connectivity entirely. The point-cloud encoder sidesteps both problems by embedding each atom as a coordinate plus a few feature tokens, giving a compact, order-invariant representation that sits alongside text in the same sequence.

Nach01 has no standalone publication; the nach0 paper in Chemical Science and the nach0-pc preprint are the technical record, and Insilico's SDK notebook points readers to both. It is a commercial artifact rather than an open release, and is distinct from MMAI Liquid Foundation Model, the compact on-premise model Insilico built with Liquid AI on a different backbone.

#Key Features

• Two input channels in one model: Accepts text-based chemical representations (SMILES by default) and 3D molecular structures, covering small molecules, proteins, peptides, disease annotations, and gene interactions.
• Broad multi-task coverage: Trained across 100+ chemistry-specific tasks spanning molecular property prediction, generative design, retrosynthesis, hit identification, and lead optimization, so one checkpoint replaces a set of single-purpose predictors.
• Fine-tuning on proprietary data: A training job takes a plain CSV of molecule, input format, task description, target, and task type, with a switch between a classification and a regression head and optional molecular augmentation.
• Test-time SMILES augmentation: The TOP_K parameter generates predictions across multiple augmented SMILES writings of the same molecule and combines them by arithmetic mean or majority vote, stabilizing property estimates.
• Managed cloud deployment: Distributed as a SageMaker algorithm with fine-tuning and real-time or batch inference metered separately, driven through the SageMaker API or Insilico's MIT-licensed insilico-aws client.

#Technical Details

The language backbone descends from nach0, a T5 encoder–decoder trained in base (250 million parameters, 12 layers, 768-dimensional hidden state, 12 attention heads) and large (780 million parameters, 24 layers, 1024-dimensional hidden state, 16 heads) configurations. Its vocabulary extends FLAN-T5's sub-word tokens with per-atom SMILES tokens marked as <sm_{token}>, and pretraining used 13 million chemistry-filtered PubMed abstracts (355M tokens), 119,000 USPTO patent descriptions (2.9B tokens), and roughly 100 million ZINC molecules (4.7B tokens). The nach0-pc component adds a point-cloud encoder that applies sinusoidal spatial embeddings as relative attention biases, making self-attention invariant to point ordering, and uses farthest-point sampling to compress large ligand and pocket clouds. It is pretrained by reconstructing molecular fragments that have been dropped or blurred, and was evaluated on six spatial generation tasks — distribution learning, conformation generation, shape-conditioned generation, linker design, scaffold decoration, and pocket-conditioned generation — over GEOM-Drugs, ZINC, and CrossDocked2020, reaching sample quality comparable to single-task diffusion baselines such as DiffLinker and TargetDiff at lower training and inference cost. The shipped Nach01 artifact is a fixed checkpoint, version 20250522, generating up to 1024 tokens per request; fine-tuning runs on ml.g6e.2xlarge and endpoints on ml.p3.2xlarge and g4dn/g5 instances.

#Applications

The model is aimed at the design–make–test–analyze cycle in medicinal chemistry. A discovery team subscribes through AWS Marketplace, fine-tunes the checkpoint on internal assay data — ADMET endpoints, potency, solubility, blood–brain-barrier penetration — and serves the result from an endpoint inside its own AWS account, so proprietary structures never leave the customer's environment. Beyond property prediction, the same model supports generative design against a target pocket, linker and scaffold elaboration on existing series, and retrosynthetic planning — useful to biotechs lacking the data volume or engineering staff to train such models themselves.

#Impact

Nach01 is one of the few generative chemistry foundation models distributed through a mainstream cloud marketplace rather than as downloadable weights or a bespoke service contract, which lowers the procurement barrier for smaller drug-discovery groups. That distribution model is also its main limitation for the research community: the weights are not obtainable outside an EULA-gated commercial subscription, no paper describes the shipped checkpoint or reports its benchmark numbers, and the public code is a thin client library rather than training or inference source. The published results that support it belong to nach0 and nach0-pc, so independent evaluation of Nach01 as deployed rests on customers running it on their own data.