Local AI for bioinformatics

This guide explains the AI part of Liatir without assuming that you already know single-cell analysis, genomic language models, regulatory prediction, or protein structure prediction.

Liatir uses AI locally. That means model runtimes are installed on your machine, your input files stay on your machine, and runs are recorded in Jobs, Results, and provenance just like other tools.

The three pieces

AI Models

An AI Model is the local runtime box: Python environment, packages, model cache, downloaded weights, hardware checks, and model metadata.

Examples:

• CellTypist Local Annotation.
• Nucleotide Transformer v2 50M or 500M.
• ESM-2 8M Protein.
• Enformer, Basenji2, or Borzoi Mini.
• Boltz-2.

AI Tools

An AI Tool is the actual task you run. It asks you for files and settings, then uses a compatible AI Model.

Examples:

• CellTypist Annotation uses the CellTypist AI Model.
• Sequence Embedding can use Nucleotide Transformer or ESM-2.
• Genomic Variant Effect uses Nucleotide Transformer models.
• Regulatory Prediction uses Enformer, Basenji2, or Borzoi Mini.
• Protein Structure Prediction uses Boltz-2.

Results

Results are the output of one run. They can include:

• readable tables and metrics;
• CSV and JSON files;
• embeddings;
• BED genome tracks;
• PDB or mmCIF protein structures;
• warnings;
• logs;
• provenance.

Provenance is important. It tells you which model, version, runtime, input files, parameters, and output files were used.

What each AI workflow is for

Workflow	Use it when	Typical input	Typical output
Single-cell annotation	You have cells and want likely cell-type labels	.h5ad AnnData	labels, label counts, summary
Sequence embedding	You want a numeric representation of DNA, RNA, or protein sequences	FASTA or pasted sequence	embedding table, dimensions, summary
Variant effect scoring	You want a first local signal for which variants change sequence representation	FASTA + VCF/VCF.GZ	scores, BED track, warnings
Regulatory prediction	You want predicted genomic signal tracks from DNA sequence windows	FASTA or pasted DNA, optional VCF	signal track, variant deltas
Protein structure prediction	You want a predicted 3D protein structure	protein FASTA or sequence	mmCIF/PDB, confidence, optional binding outputs

How to read results

Labels

Labels are categories predicted by a model. For example, CellTypist may label cells as T cells, B cells, monocytes, or other cell types.

High confidence does not mean the label is biologically final. It means the model found a strong match according to its reference. If your tissue, species, assay, or preprocessing differs from the model reference, labels can be wrong.

Embeddings

An embedding is a vector: a list of numbers that represents a biological input.

You usually do not read every number manually. You use embeddings to compare, cluster, visualize, or feed another tool. Similar embeddings often mean similar model representations, not guaranteed biological identity.

Variant effect scores

Liatir's current Nucleotide Transformer variant score compares the embedding of the reference sequence window with the embedding of the alternate sequence window.

The score is useful for prioritization and exploration. It is not a clinical pathogenicity label. A higher score means the model representation changed more, not automatically that the variant is harmful.

Regulatory tracks

Regulatory prediction models output signal across bins in a sequence window. Liatir writes those signals as CSV and BED so they can be inspected as genome tracks.

The targetIndex selects which model output track to inspect. Start with index 0 for a basic test. For serious interpretation, you need to know what the target represents.

Protein structures

Protein structure tools produce a 3D structure file. Confidence and binding values help you judge whether the run looks plausible, but predicted structures still need scientific review.

If a run completes without a structure file, Liatir treats that as a failed scientific output.

Good first tests

Start small:

1. Install CellTypist and run a small .h5ad demo.
2. Install Nucleotide Transformer 50M and run Sequence Embedding on a short FASTA.
3. Run Genomic Variant Effect on the demo FASTA and VCF.
4. Install one regulatory model and run Regulatory Prediction with targetIndex = 0 and a low maxVariants.
5. Run Boltz-2 on a short protein sequence and inspect the generated structure.

Avoid starting with large files, many variants, or high sample counts until you know the workflow is healthy.

Building useful pipelines

Single-cell

Use:

1. AnnData .h5ad input.
2. CellTypist Annotation.
3. Results table or single-cell preview.
4. Later, foundation-model embeddings and Vitessce viewers.

Genomic variants

Use:

1. Reference FASTA.
2. Variant VCF or VCF.GZ.
3. Genomic Variant Effect or Regulatory Prediction.
4. BED track output.
5. Genome viewer.

Protein structure

Use:

1. Protein FASTA or pasted sequence.
2. Protein Structure Prediction with Boltz-2.
3. 3D viewer.
4. Report or exported structure file.

Red flags

Be careful when:

• the input file format is not the one the tool expects;
• a single-cell matrix is raw counts when the model expects normalized data;
• VCF REF bases do not match the selected FASTA;
• a regulatory model uses an unknown targetIndex;
• CPU runtime takes much longer than expected;
• a result has warnings that you have not read;
• provenance does not match the model or input you intended to run.

Where to go next

• AI Models
• AI Tools
• Pipelines
• Data