Local AI for bioinformatics
Liatir runs supported scientific AI Models locally through signed Runtime Boxes. A Runtime Box contains the model weights, Python interpreter, packages, and Liatir runner that were tested together. Liatir does not install model packages into your system Python.
AI Models and AI Tools
An AI Model is the packaged scientific model. An AI Tool is the task that sends an input to a compatible installed model and turns its output into Jobs, Results, files, and provenance.
The current product catalog contains:
| AI Model | Current published targets | Main use |
|---|---|---|
| Geneformer V1 10M | macOS arm64 Metal; Linux CUDA 12.9; Windows through WSL2 | human single-cell embeddings |
| scGPT Whole-human | macOS arm64 Metal; Linux CPU; Linux CUDA 12.9; Windows through WSL2 | human single-cell embeddings |
| UCE 4-layer | macOS arm64 Metal | multi-species single-cell embeddings |
| Boltz-2 | Linux CUDA 12.9; Windows through WSL2 | structure prediction and protein–ligand affinity |
| Protenix base v1.0.0 | Linux CUDA 12.6; Windows through WSL2 | structure prediction |
The AI Tools are Single-cell Embedding, and Structure Prediction and Protein–Ligand Affinity under Tools: the first runs Boltz-2 or Protenix, the second Boltz-2. No other AI Model or AI Tool is available until it has its own reviewed, published Runtime Box path.
Installation
- Open AI Models.
- Choose a model compatible with the detected computer.
- Select Install.
- Liatir downloads the target-specific Runtime Box, verifies its signature and hashes, and activates it atomically.
- The installed model becomes available to direct runs and pipelines.
AI Models are installed once per device and are available to all workspaces. Removing a model removes its activated Runtime Box and reclaims its disk space.
Input requirements
All three models accept AnnData .h5ad files with raw scRNA-seq counts in .X, but gene identifiers differ:
- Geneformer is human-only and expects Ensembl IDs in
var["ensembl_id"]orvar_names. - scGPT is human-only and expects gene symbols in
var["gene_name"],var["feature_name"], orvar_names. - UCE supports the species listed in its run form and expects gene symbols in
var_names.
Outputs
A successful run creates:
- an embedded AnnData file;
- a lightweight CSV preview;
- a JSON summary;
- model-specific intermediate artifacts when required;
- Runtime Box provenance identifying the exact box release and native target.
The embedded AnnData file is the main scientific artifact. An embedding is a numeric representation for comparison, clustering, visualization, or later analysis; it is not a final biological conclusion.
Hardware support
Liatir installs only a published target that exactly matches the operating system, architecture, accelerator, memory, and driver constraints.
On Windows, AI Models run through WSL2, the Linux environment Liatir already requires on Windows: the model you install is the Linux one, on your own NVIDIA GPU. There are no native Windows model packages, and none are inferred from the Linux ones.