SIGMA (sigma-omics)

SIGMA resolves pathological metabolic transitions with interface-aware spectral graph learning.

SIGMA takes a spatial metabolomics intensity matrix and tissue coordinates as its primary inputs. Pathological annotations, transferred labels, or weak spatial anchors guide interface inference. Its spectral graph model separates low-frequency tissue organization from interface-associated metabolic signals.

Matched spatial transcriptomic data can optionally provide auxiliary molecular information and support downstream biological validation and interpretation. SIGMA returns an inferred pathological interface, signed-distance coordinates, interface-associated metabolic programs, ranked metabolites, influence ranges, and directional anisotropy.

Install

After the first PyPI release:

pip install sigma-omics

The distribution name is sigma-omics; the Python import remains sigma_spatial.

Install from source

pip install -e .

Quick start

import scanpy as sc
from sigma_spatial import run_sigma

adata = sc.read_h5ad("sample.h5ad")
result = run_sigma(
    adata,
    anchor_key="sigma_anchor",        # 1=tumor, 0=non-tumor, NaN=unknown
    representation_key="X_harmony",  # validated auxiliary representation
    random_state=0,
)

# Main outputs
result.obs[["sigma_region_probability", "sigma_boundary", "sigma_d_signed"]]

For an unfamiliar input, inspect the declared modality before choosing a workflow:

from sigma_spatial import inspect_input

assessment = inspect_input(adata)  # uses var['feature_type']: SM, ST, or both
assessment.print_report()

SM-only data are routed to the explicitly anchored weak-anchor workflow; joint SM+ST data use SM as the target and an ST representation for auxiliary alignment. ST-only data may be used for interface/signature validation, but SIGMA does not label ST-only outputs as metabolic programs.

Reproduce a reference dataset

Reference data releases contain the core AnnData, frozen feature ranking, and frozen program assignment table. These are all required because re-clustering a ranking is not guaranteed to preserve manuscript cluster identities.

import scanpy as sc
from sigma_spatial import run_reference_analysis

adata = sc.read_h5ad("HBC515_SIGMA_core.h5ad")
report = run_reference_analysis(
    adata,
    dataset="HBC515",
    ranking_path="BC515_lambda_ranking.csv",
    assignments_path="BC515_metabolic_program_assignments.csv",
    output_dir="results/HBC515_reference",
)

The preset fixes the manuscript matrix source, program, distance windows, and random seed. A frozen representative table can additionally be supplied with representatives_path= when a manuscript panel used a curated final display set.

Generate the downstream result set

For a new user, the recommended entry point is the combined workflow:

import sigma_spatial as sigma

sigma.assess_input(adata).print_report()
result = sigma.run_analysis(
    adata,
    output_dir="results/sample_01",
    prefix="sample_01",
    workflow="auto",
    evidence={"same_section_pathology": True},
    report_level="standard",
    random_state=0,
)
print(result.workflow)
print(result.downstream.leading_program)

For scientific safety, workflow="auto" reads existing provenance or requires explicit evidence. It does not infer the annotation source from whichever workflow gives the strongest result. Supported evidence flags are same_section_pathology, matched_st, adjacent_section_pathology, and region_defined.

report_level controls plotting without changing numerical analysis:

• none: tables and provenance only;
• standard: program patterns, profiles, lollipop, and representative metabolites;
• complete: standard report plus influence-range and anisotropy panels;
• manuscript: publication-layout report with the complete diagnostics.

After SIGMA has produced a signed-distance field, one call discovers the metabolic programs and writes the standard figures and tables:

from sigma_spatial import run_downstream_analysis

report = run_downstream_analysis(
    result,
    "results/sample_01",
    prefix="sample_01",
    selection_mode="lambda_profile",
    workflow="direct_pathology",   # one of the four evidence workflows
    signature_scores=None,           # optional matched-ST score arrays
)
print(report.leading_program)
print(report.representatives[["mz", "cluster"]])

The report includes spatial program maps, signed-distance profiles, two-sided near-versus-far interface statistics, five representative metabolite maps, and directional anisotropy. When independent ST signature scores are provided, it additionally creates the program-signature heatmap and ST interface-enrichment plot. The downstream call does not refit the SIGMA model.

workflow is selected once from direct_pathology, multiomics_inferred, transferred_pathology, or region_defined_disease. Each workflow has one fixed lambda-profile preset shared by its samples; users do not tune thresholds separately for every section. For sparse signed-distance grids, SIGMA can deterministically reduce the number of bins and records both the requested and effective bin counts. The exact preset is saved beside every result.

For exact reproduction of a frozen manuscript analysis, set selection_mode="reference" and pass its frozen final ranking table. If the manuscript program was selected using dataset-specific biological criteria, also pass leading_program=<frozen cluster>; otherwise the leading program is selected automatically from the two-sided near-versus-far analysis. The JSON summary records both the automatic result and any explicit reference override.

For SM-only data, use the separately validated HCC-derived entry point:

from sigma_spatial import run_sigma_weak_anchor

result = run_sigma_weak_anchor(
    adata,
    anchor_key="sigma_anchor",
    already_log=True,
    random_state=0,
)

This path preserves the executed HCC P1/P4 loss without introducing an RNA target. See docs/input_format.md and docs/output_schema.md.

Scope

The package contains reusable SIGMA computation and standardized downstream reporting. Manuscript-specific simulation, benchmarking, GO enrichment, and sample-specific interpretation remain under examples/ or the analysis repository.