AUTOENCODIX
Summary¶
AUTOENCODIX is an open-source framework addressing the lack of standardization across autoencoder (AE) implementations used for biological representation learning: it provides a single, configurable, end-to-end pipeline — preprocessing, model training (with built-in hyperparameter tuning via Optuna), visualization, and downstream machine-learning evaluation — spanning five AE architectures, applied to TCGA pan-cancer multi-omics and single-cell (plus imaging) datasets. The paper functions as both a software resource and a benchmarking study, providing practical recommendations on architecture and hyperparameter tradeoffs rather than claiming a single best AE variant [1].
Architectures¶
Five AE variants are implemented behind a common YAML-configured pipeline (make data / make model / make visualize / make ml_task):
- Vanillix: a standard (non-variational) autoencoder, fully connected encoder/decoder with a bottleneck.
- Varix: a variational autoencoder (VAE) baseline, same fully connected structure plus a KL-divergence loss term against a prior latent distribution.
- Stackix: a stacked/hierarchical VAE for multi-omics integration — independent VAEs are trained per modality, then their embeddings are concatenated and passed through a further VAE, reducing sensitivity to cross-modality scale/distribution differences.
- Ontix: an ontology-based VAE whose decoder is restricted to connections defined by a user-supplied ontology (e.g., Reactome pathways, chromosomal gene location), so each latent dimension is directly interpretable as a biological process rather than requiring post-hoc feature-importance explanation.
- X-modalix: a cross-modal VAE (inspired by prior modality-translation VAEs) enabling translation between data modalities, including imaging data.
Key Findings¶
- β (KL-loss weight) tradeoff in VAEs: sweeping β from 0.01–10 (annealed over 1,000 epochs) showed reconstruction capability (explained variance R²) highest at very low β, while higher β produced denser, more disentangled (lower total-correlation) latent spaces — a classic reconstruction-vs-disentanglement tradeoff, with the authors settling on β ∈ {0.01, 0.1, 1} for further benchmarking.
- No universally best architecture: across two datasets (TCGA, single-cell), multiple modality combinations, three latent dimensionalities (2, 8, 29 — the last matching the number of Reactome top-level pathways for Ontix), and three downstream ML algorithms, no single AE architecture consistently won on downstream classification/regression task performance; all AE embeddings outperformed random-feature baselines, and results were architecture × algorithm × task dependent.
- Reconstruction quality does not predict embedding usefulness: Ontix showed markedly worse raw reconstruction (expected, given its sparse, ontology-constrained decoder) but this did not translate to worse downstream embedding performance — the paper explicitly warns against using reconstruction loss alone as a proxy for embedding quality.
- Hyperparameter tuning matters most for Ontix: standard (untuned) configurations were fairly robust for most architectures, but ontology-based VAEs at latent dimensions 8 and 29 were the main architecture to show consistent improvement in both reconstruction and downstream embedding quality from tuning (lower dropout, faster learning rate).
- Ontology-based explainability validated on sex chromosomes: using chromosomal location as a non-overlapping ontology benchmark, Ontix showed a strong, biologically expected signal on the X chromosome in TCGA data (and, as expected, no signal on the small Y chromosome), supporting that Ontix's pathway-restricted latent dimensions carry genuine biological signal rather than arbitrary structure.
Availability¶
Framework: github.com/jan-forest/autoencodix. Configuration reference: ConfigParams.md. Reproducibility code for the paper's experiments: github.com/jan-forest/autoencodix-reproducibility (archived on Zenodo).
See Also¶
- scVI — a variational-inference autoencoder purpose-built for single-cell transcriptomics, illustrating the kind of task-specific AE that AUTOENCODIX's general-purpose, multi-architecture pipeline aims to standardize and benchmark against.
- mtVAE — a domain-specific (metabolomics) VAE application; AUTOENCODIX generalizes the same representation-learning approach across omics modalities within one configurable framework.