AI Tools

AI Tools are the pipeline-ready tasks that use local AI Models. They are not Plugins. Plugins are .lia extensions; AI Tools are built-in capabilities that run compatible AI Models through the shared Liatir I/O contract.

Available AI Tools

AI Tool	Compatible AI Models	Main output
CellTypist Annotation	CellTypist Local Annotation	cell labels and summary
Sequence Embedding	Nucleotide Transformer 50M/500M, ESM-2 8M	embedding CSV and JSON summary
Genomic Variant Effect	Nucleotide Transformer 50M/500M	variant scores and BED track
Regulatory Prediction	Enformer, Basenji2, Borzoi Mini	signal tracks and optional variant deltas
Protein Structure Prediction	Boltz-2	mmCIF/PDB structure and confidence metadata

Direct run or pipeline?

Direct model pages are useful for quick tests. Pipelines are better when you want repeatable workflows, connected outputs, and a full run history across multiple steps.

Use a pipeline when:

• one output file should feed another tool;
• you want to repeat the same workflow on more samples;
• you need Results to show the whole parent pipeline;
• you are testing model outputs together with viewers or downstream tools.

How model selection works

Each AI Tool exposes an AI Model dropdown. Liatir only shows installed models that match the tool contract.

For example, Protein Structure Prediction will not accept a Nucleotide Transformer model, because that model creates sequence embeddings and does not predict structures.

Common outputs

AI Tools usually create both human-readable Results sections and output files. Output files are important because later pipeline steps can consume them.

Output type	What it means
CSV	Table you can inspect or pass to other tools
JSON	Structured summary, metrics, warnings, and provenance
BED	Genome-browser track
PDB/mmCIF	Protein structure file
embeddings	Numeric vectors for comparison, clustering, or downstream analysis

Reading scientific confidence

Liatir helps you inspect output, but it does not turn model output into a final scientific conclusion. Always check:

• the model page;
• input preprocessing requirements;
• warnings;
• provenance;
• whether the model is meant for your organism, tissue, assay, or sequence type.