TPCAV: Interpreting deep learning genomics models via concept attribution.

Area of Science:

• Genomics
• Computational Biology
• Machine Learning

Background:

• Interpreting deep learning models in genomics is difficult.
• Current methods struggle with complex genomic features like chromatin accessibility.
• Concept attribution methods offer global interpretation but haven't been widely used in genomics.

Purpose of the Study:

• To introduce the first application of concept attribution for interpreting genomics deep learning models.
• To develop an improved method, TPCAV (Testing with PCA-projected Concept Activation Vectors), addressing limitations of existing techniques.
• To enable robust interpretation of diverse genomic features and model types.

Main Methods:

• Adapted the Testing with Concept Activation Vectors (TCAV) method.
• Introduced TPCAV using PCA-based decorrelation for genomic embedding features.
• Developed a strategy for extracting concept-specific input attribution maps.
• Evaluated TPCAV on various genomics models and input representations.

Main Results:

• TPCAV provides more reliable DNA motif interpretation than TCAV.
• TPCAV performance is comparable to TF-MoDISco for DNA motif prediction.
• TPCAV successfully interprets general concepts like repetitive elements and chromatin accessibility.
• TPCAV generalizes to foundation models and models with chromatin signal inputs.
• TPCAV identifies representative transcription factor binding sites linked to specific concepts.

Conclusions:

• TPCAV offers a flexible and robust approach to interpreting genomics deep learning models.
• This method extends beyond DNA motif analysis to broader genomic features and advanced model architectures.
• TPCAV facilitates deeper biological insights and downstream investigations into distinct binding mechanisms.