AUTOENCODIX: a generalized and versatile framework to train and evaluate autoencoders for biological representation...

Area of Science:

• Deep learning
• Bioinformatics
• Computational biology

Background:

• Autoencoders are powerful for representation learning and multimodal data integration in data-driven research.
• Current autoencoder implementations lack standardization, versatility, comparability, and generalizability.

Purpose of the Study:

• To present AUTOENCODIX, an open-source framework for standardized and flexible autoencoder pipeline preprocessing, training, and evaluation.
• To offer user-centric insights and recommendations for navigating autoencoder architectures, hyperparameters, and tradeoffs in representation learning.

Main Methods:

• Developed AUTOENCODIX, an open-source framework for autoencoder pipelines.
• Applied AUTOENCODIX to pan-cancer (The Cancer Genome Atlas), single-cell sequencing, and imaging datasets.
• Evaluated autoencoder architectures including ontology-based and cross-modal variants.

Main Results:

• AUTOENCODIX provides a standardized and flexible pipeline for autoencoder implementation.
• Ontology-based and cross-modal autoencoders offer advantages in explainability and data translation.
• Key tradeoffs in reconstruction capability, embedding quality for machine learning, and explainability reliability were identified.

Conclusions:

• AUTOENCODIX facilitates robust and reproducible autoencoder-based research in complex biological data.
• The framework aids researchers in selecting optimal autoencoder architectures and hyperparameters for specific downstream tasks.
• Findings offer practical guidance for leveraging autoencoders in bioinformatics and computational biology.