IsoGraph

IsoGraph discovers co-regulated transcript programs from bulk RNA-seq by treating gene abundance and isoform switching as separate channels in a multiplex network. Starting from transcript-level counts, it builds gene-local switch coordinates from compositional transcript usage and a standardized abundance signal per gene, infers sparse gene-module structure via a VAE or linear backend, classifies each module gene by the channel driving its membership (switching, abundance, coupled, or discordant), and links modules to phenotypic traits. All steps are benchmark-validated and reproducible.

Core Capabilities

• Multiplex network inference — builds a typed feature graph with separate abundance and switch channels per gene, calibrated edge thresholds, and gene channel role classification (coupled, switch_only, abundance_only, discordant).
• Five interchangeable backends — baseline, latent, graph, vae (default), and wgcna; all support multiplex mode and the full benchmark suite.
• Benchmark validation — three fixture suites: core_v1 (24–800 genes), scale_v1 (6k–12k genes), and multiplex_v1 (typed abundance + switch ground truth). An optional xxlarge_multiplex_v1 fixture (12k genes, 240 samples) tests at full scale.
• Module explanation — isograph explain-module produces gene driver tables, transcript polarity tables, high-vs-low contrasts, publication-ready PDF plots, optional VAE decoder attribution, and Captum Integrated Gradients encoder attribution.
• Structural annotation — isograph annotate-structure labels switch pairs with GTF-derived exon, CDS/UTR, biotype, and coding-status changes.
• Reproducibility — fixture-driven, seed-controlled, version-locked YAML configs; compact JSON reports as durable evidence; real-data freeze via freeze-real.

Installation

Install the core package from PyPI:

pip install isograph

The core package supports Python 3.11 through 3.14.

Optional backends

IsoGraph installs mpmath, which is required by modern SymPy releases. The vae backend also requires PyTorch, but PyTorch is intentionally not installed by IsoGraph because CPU/GPU/CUDA builds are platform-specific. Install the build that matches your system before using it:

pip install torch

See the PyTorch installation guide for GPU/CUDA builds.

The wgcna backend requires R with the WGCNA package and Rscript on PATH.

Quickstart

Step 1 — Validate on bundled fixtures

Run the multiplex benchmark on the toy fixture to confirm your installation (VAE default, ~2 minutes):

isograph benchmark --config-name stage9_multiplex_vae \
  -- fixture_filter=toy_multiplex_v1 stage_name=readme_smoke

This writes artifacts under artifacts/benchmarks/readme_smoke/toy_multiplex_v1/ and a JSON report under artifacts/reports/, including overall recovery and role-aware recall (switch_only, abundance_only, coupled, discordant).

Run all four backends on the full multiplex suite:

isograph benchmark --config-name stage9_multiplex_vae
isograph benchmark --config-name stage9_multiplex_graph
isograph benchmark --config-name stage9_multiplex_latent
isograph benchmark --config-name stage9_multiplex_wgcna

The optional 12k-gene stress fixture:

isograph benchmark --config-name stress_multiplex_xxlarge_vae
isograph benchmark --config-name stress_multiplex_xxlarge_wgcna

Step 2 — Fit your own data

IsoGraph expects a prepared dataset bundle with a manifest.json, sample metadata, feature tables, and count matrices. Providing a gene_counts matrix activates the abundance channel alongside the isoform-switch channel.

isograph fit \
  --dataset-path path/to/my_dataset_bundle \
  --output-dir artifacts/fits/my_dataset

With tuning overrides (all after --):

isograph fit \
  --dataset-path path/to/my_dataset_bundle \
  --backend vae \
  --output-dir artifacts/fits/my_dataset_vae \
  -- vae.hidden_dim=256 vae.latent_dim=8 vae.n_epochs=500 \
     vae.alpha=0.70 vae.alpha_switch=0.70 \
     vae.alpha_abundance_grid="[0.60,0.65,0.70,0.75,0.80,0.85,0.90]"

What you get:

File	Contents
modules.parquet	Module assignments per gene
edges.parquet	Inferred gene-gene edges with weights
traits.parquet	Module–trait associations and p values
feature_scores.parquet	Per-gene switch and abundance scores with feature_type column
module_gene_roles.parquet	Per-gene role: coupled, switch_only, abundance_only, discordant

Step 3 — Explain modules

isograph explain-module \
  --artifact-dir artifacts/fits/my_dataset \
  --feature-table path/to/feature_scores.parquet \
  --feature-meta path/to/transcript_table.parquet \
  --module-ids M000 M001 \
  --plot --output-format pdf \
  --output-dir artifacts/explain/my_dataset

Produces gene driver tables, transcript polarity tables, high-vs-low contrasts, and publication-ready PDF plots per module.

The detailed walkthroughs live in the Wiki, and the formal data model is documented in the RTD source tree.

Documentation

• Details on all of the functions lives in the Read the Docs API.
• Step-by-step tutorials for installation, data preparation, and own-data workflows live in the GitHub Wiki.

Citation

If you use IsoGraph in research, cite the software repository using the metadata in CITATION.cff. If a manuscript or preprint becomes available later, that can be added as a preferred citation target without changing the software citation path.

Acknowledgements

IsoGraph is supported by the National Institute on Minority Health and Health Disparities award R00 MD0169640 and the Alzheimer's Association award 25AARG-1413315.

Reproducibility and Data Provenance

• The benchmark suite is fixture-driven and designed to preserve regression targets across development stages.
• The bundled real-data workflow freezes a reproducible real_caudate_aa_v1 dataset from BrainSEQ-derived inputs and caches intermediate selections under benchmarks/cache/real_data/.
• Benchmark, calibration, runtime, and snapshot artifacts are written into versioned directories under artifacts/ and snapshots/.
• Bulky generated benchmark directories and dataset bundles are ignored by git. Commit durable benchmark evidence as compact JSON reports under artifacts/reports/.

Limitations

• The benchmark CLI is optimized for the bundled fixture suite rather than arbitrary user-defined suites.
• The VAE backend requires a separate PyTorch installation.
• The WGCNA backend requires R with the WGCNA package installed.
• The freeze-real workflow depends on BrainSEQ-style source files and is not a generic data-ingestion command for arbitrary cohorts.
• VAE decoder attribution (--vae-attribution) and Captum Integrated Gradients (--integrated-gradients) require a VAE checkpoint in the fit artifact directory and, for Integrated Gradients, pip install isograph[torch-explain].