SRAgent

Agentic workflows for obtaining data from the Sequence Read Archive.

Manuscript

scBaseCount: An AI agent-curated, uniformly processed, and continually expanding single cell data repository. Nicholas D Youngblut, Christopher Carpenter, Jaanak Prashar, Chiara Ricci-Tam, Rajesh Ilango, Noam Teyssier, Silvana Konermann, Patrick Hsu, Alexander Dobin, David P Burke, Hani Goodarzi, Yusuf H Roohani. bioRxiv 2025.02.27.640494; doi: https://doi.org/10.1101/2025.02.27.640494

Install

Option 1: Via conda

Create a conda environment [optional]:

mamba create -n sragent-env -y python=3.12 sra-tools=3.1 \
  && conda activate sragent-env

Clone the repository:

git clone https://github.com/ArcInstitute/SRAgent.git \
  && cd SRAgent

Option 2: Via UV

Create a uv venv:

uv venv

Source the venv:

source .venv/bin/activate

Install the package

Install the package:

pip install .

For development, install the package with the dev optional dependencies:

pip install -e ".[dev]"

Environmental variables

• OPENAI_API_KEY = API key for using the OpenAI API
• required when using OpenAI models
• See Configuring models information on setting models
• ANTHROPIC_API_KEY = API key for using the Anthropic API
• required when using Claude models
• EMAIL = email for using the Entrez API
• optional, but HIGHLY recommended
• NCBI_API_KEY = API key for using the Entrez API
• optional, increases rate limits
• DYNACONF = switch between "test", "prod", and "claude" environments
• optional, default is "prod"
• this affects the SQL database used and models selected
• no database is used by default

Testing

pip install pytest

pytest tests/

Usage

SQL database

Components of SRAgent can use an SQL database to store the results.

This was crucial for the scBaseCount project, in order to: * track which datasets had been processed * quickly assess the progress of the project

However, for most users, the SQL database is not necessary. SRAgent does not use the SQL database by default.

Note: currently only a GCP Postgresql database is supported.

To set up the database, see Setting up the SQL Database.

Entrez Agent

The lowest-level agent in the SRAgent hiearchy. The agent can call various Entrez tools (esearch, efetch, esummary, and elink). Usually, the SRAgent agent will be more useful, since it includes more tools, including calling the Entrez agent.

Example accession conversion:

SRAgent entrez "Convert GSE121737 to SRX accessions"

Example of summarizing a dataset:

SRAgent entrez "Summarize SRX4967527"

SRAgent agent

A general tool for extracting data from the SRA database. The tools available:

• Entrez agent (see above)
• SRA BigQuery
• scraping NCBI webpage HTML
• sra-stat and fastq-dump (directly assessing sequence data)

Example of summarizing a dataset

using more tools than the Entrez agent

SRAgent sragent "Summarize SRX4967527"

Example of obtaining metadata for a specific SRX accession:

SRAgent sragent "Obtain any available publications for GSE196830"

Example of obtaining specific metadata fields for a dataset:

SRAgent sragent "Which 10X Genomics technology was used for ERX11887200?"

Papers agent

Locate publications linked to SRA accessions and download the corresponding manuscripts.

• Input formats: single accession (SRX/PRJ/ERX/ERP) or a CSV with an accession column (--accession-column to override the default accession).
• Workflow:
• converts accessions to Entrez IDs
• finds linked PubMed IDs
• extracts DOIs
• tries multiple sources to fetch manuscripts from the DOIs (preprint servers, CORE, Europe PMC, Unpaywall)
• Output: console summary of publications, DOIs, download status, and stored files under --output-dir/<accession>.
• CSV enrichment: when a CSV input is provided, an updated copy is written to --output-dir/<input-filename> with added pubmed_id, doi, and download_path columns.
• Helpful options:
• --core-api-key/CORE_API_KEY for CORE API
• --email/EMAIL for Unpaywall
• --max-concurrency for max concurrent tasks
• --recursion-limit for max recursion depth
• --write-graph for writing the workflow graph to a file

Examples

Download papers for a single SRX accession:

SRAgent papers SRX4967527

Use a study accession (SRP):

SRAgent papers SRP167700

Use a bioproject accession (PRJNA):

SRAgent papers PRJNA498286

Process a CSV and add DOIs and metadata to the CSV:

SRAgent papers accessions.csv

SRX-info agent

Obtain specific metadata for >=1 SRA dataset.

• Input: >=1 Entrez ID
• Output metadata fields:
• SRX accession for the Entrez ID
• SRR accessions for the SRX accession
• Is the dataset Illumina sequence data?
• Is the dataset single cell RNA-seq data?
• Is the dataset paired-end sequencing data?
• Which scRNA-seq library preparation technology?
• If 10X Genomics, which particular 10X technologies?
• Single nucleus or single cell RNA sequencing?
• Which organism was sequenced?
• Which tissue(s) were sequenced?
• Corresponding tissue ontology ID(s)
• Any disease information?
• Any treatment/perturbation information?
• Any cell line information?
• Workflow
• The agent converts the Entrez IDs to SRX accessions
• For each SRX accession, the agent obtains metadata
• The agent consolidates the metadata into a single report

As of now, the metadata fields are hard-coded into the agent. If you need alternative metadata fields, you will have to modify metadata.py.

Examples

A single SRA dataset:

SRAgent srx-info 25576380

Multiple SRA datasets:

SRAgent srx-info 36106630 32664033 27694586

Use the SQL database to filter out already-processed datasets:

SRAgent srx-info --use-database 18060880 27454880 27454942 27694586

Metadata agent

Similar to the SRX-info agent, but you can provide SRX accessions directly, instead of Entrez IDs. This saves compute time, since the agent does not need to convert the Entrez IDs to SRX accessions.

Provide a CSV of Entrez IDs and their associated SRX accessions to obtain metadata. Useful for when you already have the SRX accessions, instead of providing the Entrez IDs to SRAgent srx-info.

The CSV should have the header: entrez_id,srx_accession.

The metadata fields are the same as the SRX-info agent.

Examples

SRAgent metadata "entrez-id_srx-accession.csv"

Tissue-ontology agent

An agent for categorizing tissue descriptions using the Uberon ontology. The agent helps identify the most suitable Uberon ontology term for a given tissue description.

• Input: Free text description of one or more tissues
• Output: Uberon ontology IDs (UBERON:XXXXXXX) for each tissue description
• Workflow
• The agent processes each tissue description separately
• For each description, it identifies the most suitable Uberon ontology term
• The agent returns the corresponding Uberon ID for each tissue

Examples

Categorize a single tissue:

SRAgent tissue-ontology "Categorize the following tissue: brain"

Categorize multiple tissues:

SRAgent tissue-ontology "Tissues: lung, heart, liver"

Finding ontology terms for complex tissue descriptions:

SRAgent tissue-ontology "Find the ontology term for the thin layer of epithelial cells lining the alveoli in lungs"

Disease-ontology agent

An agent for categorizing disease descriptions using the MONDO/PATO ontology. The agent helps identify the most suitable MONDO or PATO ontology term for a given disease description.

• Input: Free text description of one or more diseases
• Output: MONDO/PATO ontology IDs for each disease description
• Workflow
• The agent processes each disease description separately
• For each description, it identifies the most suitable MONDO or PATO ontology term
• The agent returns the corresponding ontology ID for each disease

Examples

Categorize a single disease:

SRAgent disease-ontology "Categorize the following disease: heart disorder"

Finding ontology terms for specific diseases:

SRAgent disease-ontology "What is the MONDO ontology ID for congestive heart failure?"

Categorize multiple diseases:

SRAgent disease-ontology "Diseases: heart neoplasm, bursitis"

find-datasets agent

A high-level agent for finding datasets in the SRA via esearch and then processing them with the SRX-info agent.

• Input: a search query
• Output: metadata fields for the datasets found (same as SRX-info agent)
• Workflow
• The agent uses esearch to find datasets
• The agent processes the datasets with the SRX-info agent
• The agent consolidates the metadata into a single report

Examples

SRAgent find-datasets --max-datasets 2 \
  "Obtain recent single cell RNA-seq datasets in the SRA database"

Target specific organisms

SRAgent --no-summaries \
  find-datasets --max-datasets 1 --organisms pig -- \
    "Obtain recent single cell RNA-seq datasets in the SRA database"

Available organisms

• Mammals
• Human (Homo sapiens)
• Mouse (Mus musculus)
• Rat (Rattus norvegicus)
• Macaque (Macaca mulatta)
• Marmoset (Callithrix jacchus)
• Horse (Equus caballus)
• Dog (Canis lupus)
• Bovine (Bos taurus)
• Sheep (Ovis aries)
• Pig (Sus scrofa)
• Rabbit (Oryctolagus cuniculus)
• Naked mole-rat (Heterocephalus glaber)
• Chimpanzee (Pan troglodytes)
• Gorilla (Gorilla gorilla)
• Cat (Felis catus)
• Bonobo (Pan paniscus)
• Green monkey (Chlorocebus aethiops)
• Gray short-tailed opposum (Monodelphis domestica)
• Goat (Capra hircus)
• Alpaca (Vicugna pacos)
• Chinchilla (Chinchilla lanigera)
• Domestic guinea pig (Cavia porcellus)
• Golden hamster (Mesocricetus auratus)
• Eurasian hedgehog (Erinaceus europaeus)
• American mink (Neovison vison)
• Sunda pangolin (Manis javanica)
• Platypus (Ornithorhynchus anatinus)
• Ferret (Mustela putorius)
• Northern tree shrew (Tupaia belangeri)
• Birds
• Chicken (Gallus gallus)
• Zebrafinch (Taeniopygia guttata)
• Goose (Anser cygnoides)
• Duck (Anas platyrhynchos)
• Reptiles
• Turtle (Trachemys scripta)
• Amphibians
• Frog (Xenopus tropicalis)
• Axolotl (Ambystoma mexicanum)
• Fish
• Zebrafish (Danio rerio)
• Salmon (Salmo salar)
• Stickleback (Gasterosteus aculeatus)
• Invertebrates
• Fruit fly (Drosophila melanogaster)
• Roundworm (Caenorhabditis elegans)
• Mosquito (Anopheles gambiae)
• Blood fluke (Schistosoma mansoni)
• Plants
• Thale cress (Arabidopsis thaliana)
• Rice (Oryza sativa)
• Tomato (Solanum lycopersicum)
• Corn (Zea mays)
• Microorganisms
• Metagenome
• Other
• Other

Using an SQL database to store results

Using the test database:

SRAgent --no-summaries \
  find-datasets --use-database --tenant test \
    --max-datasets 1 --organisms rat -- \
    "Obtain recent single cell RNA-seq datasets in the SRA database"

Workflow status

By default, most SRAgent sub-commands (e.g., entrez and sragent) use "fancy" CLI output formatting:

Progress bar

Final results

Parameters for controlling the CLI workflow status output

• --no-progress: Disable progress bar
• --no-summaries: Disable LLM summaries for each workflowstep

Example:

SRAgent --no-summaries --no-progress \
  entrez "Convert GSE121737 to SRX accessions"

Configuring models

The models used by SRAgent are configured in the settings.yml file. Options for updating the settings:

1) Provide a new settings file

• Create a new settings yaml file
• Set the DYNACONF_SETTINGS_PATH environment variable to the path to the new settings file
• e.g., export DYNACONF_SETTINGS_PATH=/path/to/settings.yml
• No need to (re)install the package. The settings will be loaded from the new file.

2) Update and install

• Clone the repository
• Update the settings.yml file
• (Re)install the the package
• e.g., pip install .

Using Claude models

SRAgent supports using Anthropic's Claude models:

• Set the ANTHROPIC_API_KEY environment variable to your Anthropic API key
• Switch to the Claude environment using export DYNACONF=claude
• Update the settings.yml file, as needed
• See Configuring models information on setting model parameters
• Run SRAgent commands as usual

Claude models support different reasoning effort levels: * low: 1024 thinking tokens (best for simple tasks) * medium: 4096 thinking tokens (good balance) * high: 16384 thinking tokens (best for complex reasoning) * Anything else: Disables thinking tokens feature

Example:

export ANTHROPIC_API_KEY=your_api_key
export DYNACONF="claude"
SRAgent entrez "Convert GSE121737 to SRX accessions"

You can also customize the specific Claude model in settings.yml:

claude:
  models:
    default: "claude-sonnet-4-5"  # Or any other Claude model version
  temperature:
    default: 0.1
  reasoning_effort:
    default: "medium"  # Set your preferred reasoning effort; use "" to disable

Setting up the SQL database

• Create a GCP Postgresql database. See the docs.