KansformerEPI: a deep learning framework integrating KAN and transformer for predicting enhancer-promoter...

Area of Science:

• Genomics
• Computational Biology
• Molecular Biology

Background:

• Enhancer-promoter interactions (EPIs) are crucial for gene regulation and understanding disease mechanisms.
• Current computational methods for genome-wide EPI prediction often lack a multi-cell line perspective and fail to capture complex nonlinear feature relationships.
• Existing models are typically limited to single cell lines, hindering scalability and broad applicability.

Purpose of the Study:

• To develop a global enhancer-promoter interaction (EPI) prediction model applicable across multiple cell types.
• To improve the accuracy and scalability of computational EPI prediction by effectively modeling nonlinear relationships between epigenetic and sequence features.
• To provide a versatile tool for understanding transcriptional regulation and disease mechanisms across diverse cellular contexts.

Main Methods:

• Developed KansformerEPI, a novel global EPI prediction model integrating KAN and Transformer architectures within an encoder.
• The Kansformer encoder captures nonlinear relationships among epigenetic and sequence features for enhanced prediction.
• Applied KansformerEPI for cross-tissue EPI prediction across various cell types, including HMEC, IMR90, K562, and NHEK.

Main Results:

• KansformerEPI demonstrated superior accuracy and stability in predicting enhancer-promoter interactions compared to existing methods like TransEPI, TargetFinder, and SPEID.
• The model successfully achieved cross-tissue prediction, highlighting its scalability and reduced dependency on tissue-specific datasets.
• Experimental results validated the model's effectiveness across diverse biological datasets.

Conclusions:

• KansformerEPI offers a scalable and accurate solution for predicting enhancer-promoter interactions across multiple cell types.
• The model's ability to capture nonlinear feature relationships advances computational approaches in gene regulation research.
• This work provides valuable insights into transcriptional regulation and disease mechanisms, applicable across various tissues and reducing the need for extensive, cell-type-specific datasets.