Metadata Enrichment with Optional LLM Backends

pysradb can enrich detailed metadata by combining a sentence-transformer embedding model with an LLM backend. The workflow infers standardized columns for age, sex, ethnicity, phenotype, cell_type, tissue, strain, and disease when enough source metadata is available.

This feature is optional. It requires pysradb[enrichment], an embedding model download on first use, and a configured Ollama, LM Studio, or vLLM-compatible endpoint.

# Install pysradb if not already installed
try:
    import pysradb

    print(f"pysradb {pysradb.__version__} is already installed")
except ImportError:
    print("Installing pysradb from GitHub...")
    import sys

    !{sys.executable} -m pip install -q git+https://github.com/saketkc/pysradb
    print("pysradb installed successfully!")

pysradb 3.0.0.dev0 is already installed

/home/runner/work/pysradb/pysradb/pysradb/download.py:15: TqdmExperimentalWarning: Using `tqdm.autonotebook.tqdm` in notebook mode. Use `tqdm.tqdm` instead to force console mode (e.g. in jupyter console)
  from tqdm.autonotebook import tqdm

from pysradb import SRAweb

Quick Start

One line enrichment!

Prerequisites: install the enrichment extra and prepare a local backend, for example Ollama with ollama pull granite4:3b.

The easiest way to enrich metadata is using the enrich=True parameter:

# Optional local setup for LLM enrichment:
#   1. Install pysradb with: python -m pip install "pysradb[enrichment]"
#   2. Install Ollama from https://ollama.com/download
#   3. Run: ollama pull granite4:3b
# The documentation build does not install or start system services.

print("Ollama setup is optional and is intentionally not run in this notebook build.")

Ollama setup is optional and is intentionally not run in this notebook build.

print("Start Ollama locally with: ollama serve")

Start Ollama locally with: ollama serve

print("Pull a model locally with: ollama pull granite4:3b")

Pull a model locally with: ollama pull granite4:3b

db = SRAweb()

df = db.metadata("GSE155673")
display(df.head())

print('To enrich this table with Ollama, run:')
print('db.metadata("GSE155673", enrich=True, enrich_backend="ollama/granite4:3b")')

	study_accession	study_title	study_summary	organism_name	platform_accession	platform_title	experiment_type	bioproject	submission_date	supplementary_files	series_ftp	sample_accession	sample_title
0	GSE155673	Systems biological assessment of immunity to s...	CITE-seq of 7 patients with COVID-19 and 5 hea...	Homo sapiens	GPL20301	NaN	Expression profiling by high throughput sequen...	PRJNA655740	2020/08/07	CSV, MTX, TSV	ftp://ftp.ncbi.nlm.nih.gov/geo/series/GSE155nn...	GSM4712885	cov_01_RNA
1	GSE155673	Systems biological assessment of immunity to s...	CITE-seq of 7 patients with COVID-19 and 5 hea...	Homo sapiens	GPL20301	NaN	Expression profiling by high throughput sequen...	PRJNA655740	2020/08/07	CSV, MTX, TSV	ftp://ftp.ncbi.nlm.nih.gov/geo/series/GSE155nn...	GSM4712886	cov_01_antibody
2	GSE155673	Systems biological assessment of immunity to s...	CITE-seq of 7 patients with COVID-19 and 5 hea...	Homo sapiens	GPL20301	NaN	Expression profiling by high throughput sequen...	PRJNA655740	2020/08/07	CSV, MTX, TSV	ftp://ftp.ncbi.nlm.nih.gov/geo/series/GSE155nn...	GSM4712887	cov_02_RNA
3	GSE155673	Systems biological assessment of immunity to s...	CITE-seq of 7 patients with COVID-19 and 5 hea...	Homo sapiens	GPL20301	NaN	Expression profiling by high throughput sequen...	PRJNA655740	2020/08/07	CSV, MTX, TSV	ftp://ftp.ncbi.nlm.nih.gov/geo/series/GSE155nn...	GSM4712888	cov_02_antibody
4	GSE155673	Systems biological assessment of immunity to s...	CITE-seq of 7 patients with COVID-19 and 5 hea...	Homo sapiens	GPL20301	NaN	Expression profiling by high throughput sequen...	PRJNA655740	2020/08/07	CSV, MTX, TSV	ftp://ftp.ncbi.nlm.nih.gov/geo/series/GSE155nn...	GSM4712889	cov_03_RNA

To enrich this table with Ollama, run:
db.metadata("GSE155673", enrich=True, enrich_backend="ollama/granite4:3b")

---

Backend Configuration

Use enrich_backend to select the LLM endpoint and embedding_model to select the sentence-transformer model used to identify relevant metadata fields.

LLM Backend Example

enrichment_example = '''
from pysradb import SRAweb

db = SRAweb()
df = db.metadata(
    "GSE155673",
    enrich=True,
    enrich_backend="ollama/granite4:3b",
)
'''
print(enrichment_example)

from pysradb import SRAweb

db = SRAweb()
df = db.metadata(
    "GSE155673",
    enrich=True,
    enrich_backend="ollama/granite4:3b",
)

expected_cols = [
    "age",
    "sex",
    "ethnicity",
    "phenotype",
    "cell_type",
    "tissue",
    "strain",
    "disease",
]
print("Enrichment adds standardized columns such as:")
print(", ".join(expected_cols))

Enrichment adds standardized columns such as:
age, sex, ethnicity, phenotype, cell_type, tissue, strain, disease

Embedding Model Example

The embedding model is still required because pysradb uses it to find relevant metadata fields before invoking the LLM backend.

embedding_example = '''
df = db.metadata(
    "GSE155673",
    enrich=True,
    enrich_backend="ollama/granite4:3b",
    embedding_model="abhinand/MedEmbed-large-v0.1",
)
'''
print(embedding_example)

df = db.metadata(
    "GSE155673",
    enrich=True,
    enrich_backend="ollama/granite4:3b",
    embedding_model="abhinand/MedEmbed-large-v0.1",
)