UNICORN: Towards universal cellular expression prediction with a multi-task learning framework.

Area of Science:

• Human genetics
• Computational biology
• Genomics

Background:

• Predicting cell-type-specific multi-omic phenotypes from biological sequences like gene expression is a significant challenge in human genetics.
• Existing computational methods have limitations in accurately capturing these complex relationships.

Purpose of the Study:

• To introduce UNICORN, a novel computational method designed to improve the prediction of cell-type-specific multi-omic phenotypes.
• To demonstrate UNICORN's superior performance compared to existing methods in gene expression and phenotype prediction.

Main Methods:

• UNICORN utilizes embeddings from biological sequences and external knowledge from pre-trained foundation models.
• The method employs carefully designed loss functions for predictor optimization.
• It integrates multi-omic information for enhanced prediction capabilities.

Main Results:

• UNICORN significantly outperforms existing methods in both gene expression and multi-omic phenotype prediction at cellular and cell-type levels.
• The method generates uncertainty scores for its predictions.
• UNICORN successfully links personalized gene expression profiles with corresponding genome information.

Conclusions:

• Foundation model embeddings enhance the understanding of biological sequences' roles in prediction tasks.
• Incorporating multi-omic data improves prediction performance.
• UNICORN offers a powerful tool for characterizing complex biological systems, including disease states and perturbations.