v0.1.0

IsoGraph v0.1.0 — Initial Release

Initial release of IsoGraph, a Python library for discovering isoform-switch and splicing co-expression modules from bulk RNA-seq data. Combines compositional transcript-usage modeling with splice-graph-aware latent network inference to recover gene-module structure and trait associations.

Features

Network backends — six selectable inference strategies:

• baseline — Deterministic sparse network (fast reference)
• latent — Factor Analysis + LedoitWolf partial correlation with cross-validation
• graph — Graph-Laplacian-smoothed latent model
• vae (default) — PyTorch nonlinear VAE with early stopping and posterior-collapse diagnostics
• wgcna — R blockwiseModules wrapper for direct WGCNA comparison
• gpu_latent — Woodbury-identity Factor Analysis for memory-efficient large-scale inference

Compositional transforms — CLR and logit normalization for transcript-usage proportions

Benchmarking framework — fixture-driven recovery scoring across synthetic datasets (24–12,000 genes) and real BrainSeq-style inputs with snapshot regression testing

CLI — isograph benchmark, isograph fit, isograph freeze-real, isograph compare, and isograph export; Hydra configuration with -- override syntax

Artifact export — reproducible modules, edges, traits, and calibration reports

Performance (VAE backend on core_v1 fixtures)

Fixture	Genes	Recovery	Runtime
toy_v1	24	1.00	7.5 s
medium_v1	400	1.00	4.2 s
realistic_v1	200	1.00	1.7 s
large_v1	800	0.80	3.8 s

Requirements

Python 3.11–3.14. PyTorch and R are optional dependencies required only for the vae and wgcna backends respectively.

---

Full Changelog: https://github.com/heart-gen/IsoGraph/commits/v0.1.0