Skip to content

Foundation Models

Pre-trained models for specialized omics domains.

Overview

Foundation models are large pre-trained models specialized for specific data types. Agentomics-ML can enable selected model families so the agent can use their catalog and local/downloaded weights during a run.

Available Types

Type Domain Use Case
dna Genomics DNA sequences, variants
rna Transcriptomics RNA sequences, expression
protein Proteomics Protein sequences, structure
molecule Chemistry Small molecules, drugs

Enabling Models

Enable foundation models before running:

./run.sh --foundation-models-type dna

This enables the selected family for the run and makes its catalog available to the agent. In Docker mode, the launcher pulls or builds an image tagged for that foundation-model type. In local mode, models are downloaded into the local workspace cache.

You can also use --foundation-models-type all to include every type.

Multiple Types

Enable a different type per run, or use all when the agent should see every configured type:

./run.sh --foundation-models-type dna
./run.sh --foundation-models-type protein

DNA Models

For genomic sequence data:

  • Variant effect prediction
  • Regulatory element detection
  • Sequence classification

Example datasets: - Gene expression from DNA features - SNP effect prediction - Promoter classification

RNA Models

For transcriptomic data:

  • RNA sequence analysis
  • Secondary structure prediction
  • Expression-based classification

Example datasets: - RNA-seq classification - Splice site prediction - RNA modification detection

Protein Models

For protein sequence data:

  • Protein function prediction
  • Structure-based classification
  • Interaction prediction

Example datasets: - Protein family classification - Enzyme activity prediction - Binding site detection

Molecule Models

For small molecule/chemical data:

  • Drug property prediction
  • Molecular classification
  • Activity prediction

Example datasets: - Drug-target interaction - Toxicity prediction - ADMET properties

How the Agent Uses Foundation Models

  1. Discovery - Agent queries available foundation models
  2. Selection - Agent chooses appropriate model for the data
  3. Embedding - Features extracted using the model
  4. Training - Embeddings used as input to ML model

Configuration

Foundation model configurations are in:

foundation_models/models.yaml

The YAML file lists each model family with: - Model names and sources - Domain type (dna, rna, protein, or molecule) - Brief summaries - Usage instructions for the agent

Without Foundation Models

If --foundation-models-type is omitted on a fresh run, no foundation models are made available to the agent. Forked runs inherit the source run's foundation-model type when the flag is omitted.

Storage Requirements

Foundation models can be large:

Type Approximate Size
DNA 1-5 GB
RNA 1-5 GB
Protein 2-10 GB
Molecule 0.5-2 GB

Ensure sufficient disk space in the Docker volume or local environment.

GPU Acceleration

Foundation models benefit significantly from GPU:

  • With GPU: Fast embedding generation
  • CPU only: Much slower, but functional

Use --cpu-only if GPU unavailable, but expect longer run times for foundation model-based approaches.

Custom Foundation Models

To add custom foundation models:

  1. Add the family to foundation_models/models.yaml
  2. Add a companion usage guide under foundation_models/
  3. Ensure the listed model names can be downloaded by src/utils/download_foundation_models.py