Appearance
AI Models
AI Models are local model runtimes that Liatir can install, manage, and use on your machine. They are different from Plugins: Plugins are .lia extensions, while AI Models are model assets and runtime packages used by AI Tools.
If you are new to this area, start with Local AI for bioinformatics before choosing a model.
What AI Models do
An AI Model provides the local engine behind an AI Tool. For example, a single-cell annotation tool may use CellTypist, while a sequence embedding tool may use a Nucleotide Transformer or ESM-2 model.
You can use AI Models in two ways:
- run a model directly from the AI Models page when Liatir provides a direct runner;
- select a compatible model inside a pipeline AI Tool.
Choosing a model
| Goal | Start with | Why |
|---|---|---|
Annotate single-cell .h5ad data | CellTypist Local Annotation | CPU-friendly and practical for first-pass labels |
| Embed DNA/RNA sequence | Nucleotide Transformer v2 50M | Lightest genomic embedding model currently exposed |
| Embed protein sequence | ESM-2 8M Protein | Small protein language model for quick local tests |
| Score small VCF examples | Nucleotide Transformer v2 50M | Faster variant-effect smoke tests |
| Predict regulatory signal | Basenji2 or Enformer | Managed regulatory runtimes with BED outputs |
| Predict protein structure | Boltz-2 | Current primary local structure model |
| Explore future single-cell foundation models | scGPT, Geneformer, UCE, scFoundation | Preview entries documented but not installable yet |
Installation
AI Models are installed globally for the app, not per workspace. Once a model is installed, every workspace can use it.
Liatir manages the runtime dependency box for each model. Heavy dependencies are not bundled into the core app; they are installed only when the model needs them.
Preview models are different: they are visible in the catalog so you can see the roadmap and read the docs, but Liatir does not expose Install or Run until the runtime box and scientific output contract are validated.
Results and provenance
AI Model runs are recorded like other Liatir analysis runs. Outputs can include tables, JSON summaries, embeddings, structure files, genome tracks, and viewer artifacts depending on the model and tool.
Each output should carry provenance:
- model ID and version;
- runtime kind and runtime version;
- input files or sequences;
- user-selected parameters;
- generated output files and metrics.
Hardware warnings
Some models can run on CPU but may be slow. Others require CUDA GPUs or a specific operating system. Liatir shows compatibility warnings before install when the current machine cannot run a model.
Always check the model page before using a result for important scientific decisions.
Related AI Tools
- CellTypist Annotation
- Sequence Embedding
- Genomic Variant Effect
- Regulatory Prediction
- Protein Structure Prediction