Enhanced performance of gene expression predictive models with protein-mediated spatial chromatin interactions

Area of Science:

• Genomics
• Computational Biology
• Molecular Biology

Background:

• Gene expression prediction traditionally relies on sequence and epigenetic data.
• Understanding 3D chromatin structure is crucial for accurate gene expression modeling.
• Existing models like ExPecto and Enformer have limitations in capturing spatial genomic information.

Purpose of the Study:

• To enhance gene expression prediction by integrating protein-mediated 3D chromatin interactions.
• To develop a novel algorithm, Spatial Gene Expression (SpEx), incorporating spatial genomic data.
• To evaluate the performance improvement of SpEx compared to baseline models.

Main Methods:

• Utilized the ExPecto algorithm architecture.
• Incorporated ChIA-PET interaction data mediated by cohesin, CTCF, and RNAPOL2 across multiple cell lines.
• Assessed model performance using Spearman's rank correlation coefficient (SCC).

Main Results:

• The developed SpEx algorithm demonstrated statistically significant improvements in gene expression prediction accuracy across most cell lines.
• Compared to the baseline ExPecto (0.82 SCC) and Enformer (0.83 SCC), SpEx achieved an average SCC of 0.83.
• Specific improvements included a 0.04 increase in SCC for RNAPOL2 on GM12878 and an SCC of 0.86 for RNAPOL2 on H1.

Conclusions:

• Integrating 3D protein-mediated chromatin interactions into gene expression prediction models significantly enhances accuracy.
• The SpEx algorithm represents a novel advancement in computational genomics, offering improved predictive power.
• This approach provides a more comprehensive understanding of the spatial organization's role in gene regulation.