v0.1.4

[0.1.4] — 2026-05-08

IsoGraph v0.1.4 introduces multi-channel module discovery, extending IsoGraph beyond isoform-switch-only networks to jointly model isoform-switch coordinates and gene abundance. This release adds multiplex feature graphs, gene channel-role classification, synthetic multiplex benchmarks, calibrated edge policies, WGCNA comparison diagnostics, and multiplex-aware explanation outputs. It also improves repository hygiene by removing bulky generated artifacts from version control while retaining compact benchmark evidence.

Added

• Multi-channel multiplex module discovery — IsoGraph now models gene abundance as a distinct feature channel alongside isoform-switch coordinates. Multi-isoform genes can contribute both switch and abundance features, while single-isoform genes contribute abundance-only features. Module discovery operates on a typed multiplex feature graph that supports switch–switch, switch–abundance, and abundance–abundance edge policies, enabling discovery of switch-dominant, abundance-dominant, coupled, and discordant modules.

• Gene channel-role classification — FitArtifacts.module_gene_roles now classifies module genes as:

  • switch_only
  • abundance_only
  • coupled
  • discordant

  Roles are computed with compute_module_gene_roles() in models/base.py.

• multiplex_v1 synthetic benchmark suite — added four multiplex benchmark fixtures:

  • toy_multiplex_v1
  • medium_multiplex_v1
  • noisy_multiplex_v1
  • large_multiplex_v1

  These include planted modules, switch/abundance truth tables, and channel-role truth tables generated by generate_multiplex_suite() in benchmarks/synthetic.py.

• Multiplex benchmark runner integration — benchmark() now dispatches multiplex_v1 fixtures and reports switch recall, abundance recall, and overall module recovery by fixture and backend.

• Calibrated multiplex edge policies — VAE, graph, and latent configs now support:

  • allow_abundance_abundance
  • alpha_switch
  • alpha_abundance
  • alpha_abundance_grid

  Tuned configs are available as:

  • stage9_multiplex_vae
  • stage9_multiplex_graph
  • stage9_multiplex_latent

• WGCNA multiplex comparison config and diagnostics — added stage9_multiplex_wgcna and WGCNA calibration reports, including selected soft-threshold power, scale-free topology fit R², module count, and unassigned-gene count.

• Role-aware recovery metrics — multiplex benchmark JSON now reports recall for switch_only, abundance_only, coupled, and discordant genes, along with giant-component diagnostics.

• Multiplex stress reporting scripts — added:

  • scripts/stress_multiplex_summary.py for backend-vs-WGCNA summaries
  • scripts/stress_multiplex_explain.py for Stage 8 explanation accuracy checks on multiplex artifacts

• Extra-large multiplex stress fixture — added optional xxlarge_multiplex_v1, which generates 12,000 genes, 240 samples, and 16 planted multiplex modules only when explicitly requested. Added configs:

  • stress_multiplex_xxlarge_vae
  • stress_multiplex_xxlarge_wgcna

• Per-channel attribution in explain-module — gene_driver_table, VAE decoder attribution output (vae_drivers.parquet), and Captum Integrated Gradients output (ig_attributions.parquet) now include feature_id and feature_type. Per-gene deduplication retains the channel with the strongest signal.

• Configurable WGCNA timeout — WgcnaModelConfig.timeout_seconds now supports long-running large multiplex comparisons.

Changed

• Stage 8D/8E accuracy scripts are now multiplex-aware — explanation scoring now detects feature-score sample columns and collapses multiplex feature rows back to gene-level metrics before evaluating decoder Jacobian or Integrated Gradients outputs.

• Generated artifact policy updated — bulky generated benchmark directories and dataset bundles are now ignored by git. Compact benchmark evidence remains tracked as JSON under artifacts/reports/.

Benchmarks

• Standard multiplex stress — VAE recovered the medium, noisy, and large multiplex fixtures with recovery scores of 1.0, 0.9885057471264368, and 0.9977777777777778, respectively. Role-aware recall was 1.0 across switch_only, abundance_only, coupled, and discordant genes for these fixtures. Graph and latent backends cleared the Stage 9 gates without giant-component failures. WGCNA remains available as the comparison backend.

• Explain-module multiplex stress — Stage 8 explanation metrics remain accurate on multiplex artifacts. VAE, graph, and latent achieved switch_strength_auc=1.0 wherever defined, while WGCNA ranged from 0.993174484983867 to 1.0 across stress fixtures.

• 12k multiplex stress — on xxlarge_multiplex_v1, VAE recovered the correct 16 planted modules with recovery 0.9266666666666667. WGCNA achieved similar recovery, 0.9191666666666665, but detected 1,898 modules, indicating severe over-segmentation.

Removed

• Tracked bulky generated artifacts — previously tracked benchmark payload directories and generated dataset bundles were removed from version control. Compact JSON reports remain trackable.

Pull Requests

• [#12] Add multiplex abundance filtering and benchmarks
• [#13] Update for xxlarge multiplex stress benchmark and ignoring artifacts

Full Changelog: v0.1.3...v0.1.4