Transplatformer: translating toxicogenomic profiles between generations of platforms.

Area of Science:

• Toxicogenomics
• Computational Biology
• Bioinformatics

Background:

• Transcriptomic profiling is crucial for analyzing biological and toxicological responses.
• Inconsistencies across platforms limit data integration and reuse of historical toxicogenomics datasets.
• Developing computational methods for cross-platform data translation is essential to leverage legacy resources.

Purpose of the Study:

• To develop computational methods for accurate cross-platform translation of gene expression data.
• To maximize the utility of legacy toxicogenomics datasets through data harmonization.

Main Methods:

• Developed TransPlatformer, a deep learning framework utilizing an attention-based architecture.
• Mapped high-dimensional fold-change vectors from microarray to current platforms.
• Trained and evaluated models using the DrugMatrix dataset across three technological generations, comparing mixed-tissue, single-tissue, and cross-tissue paradigms against baseline methods.

Main Results:

• TransPlatformer significantly reduced mean absolute error by over 50% and nearly doubled Pearson correlation compared to baseline methods in mixed-tissue training.
• The framework effectively preserved rare but biologically significant gene expression signals.
• Single-tissue models further improved accuracy for well-represented organs, highlighting the need for data augmentation in low-sample tissues.

Conclusions:

• TransPlatformer offers an effective and scalable solution for cross-platform transcriptomic data translation.
• The approach enables biologically faithful harmonization of gene expression data, facilitating legacy dataset reuse.
• This enhances downstream biomarker discovery and supports reproducible predictive modeling in toxicology.