AutoTransOP: translating omics signatures without orthologue requirements using deep learning.

Area of Science:

• Computational biology
• Systems biology
• Bioinformatics

Background:

• Animal and in vitro models often fail to accurately predict human biological responses, leading to high clinical trial failure rates.
• Understanding human biology is crucial for developing effective therapeutics and vaccines.
• Existing methods for cross-species data analysis often rely on homologous gene identification, which is not always feasible.

Purpose of the Study:

• To develop a novel computational framework, AutoTransOP, for mapping omics profiles across different biological contexts (species, cell types).
• To enable identification of relevant biological information without requiring direct ortholog mapping.
• To improve the predictive power of non-human models for human disease and therapeutic responses.

Main Methods:

• Developed AutoTransOP, a neural network autoencoder framework.
• Mapped omics profiles from various species and cellular contexts into a shared latent space.
• Validated the framework on inter-species vaccine serology studies.

Main Results:

• AutoTransOP successfully mapped omics data into a global latent space, enabling cross-contextual analysis.
• The framework performed comparably to existing methods in identifying predictive molecular features.
• Crucially, AutoTransOP did not require homology matching, a significant advantage for inter-species comparisons.
• Successfully applied to challenging inter-species vaccine serology data where direct feature mapping is absent.

Conclusions:

• AutoTransOP provides a powerful new approach for integrating and analyzing biological data across diverse contexts.
• This framework can enhance the utility of preclinical models for predicting human responses to therapeutics and vaccines.
• AutoTransOP offers a significant advancement in computational biology for overcoming species-specific data challenges.