Explainability of Protein Deep Learning Models

Area of Science:

• Proteomics
• Computational Biology
• Bioinformatics

Background:

• Protein embeddings are vital for state-of-the-art solutions in proteomics, particularly for protein interaction prediction.
• The black-box nature of these models necessitates transparency to understand underlying mechanisms.
• Explainable AI (XAI) offers methods to investigate the inner workings of these complex models.

Purpose of the Study:

• To investigate the interpretability of protein embedding models using XAI.
• To evaluate the effectiveness of various XAI methods in uncovering essential protein properties and interactions.
• To assess the quality of protein embeddings generated by different methods.

Main Methods:

• Extensive testing of nine established XAI methods on 3.3 TB of data.
• Application of XAI to protein interaction site prediction (Seq-InSite) and protein embedding generation (ProtBERT, ProtT5, Ankh).
• Evaluation based on correlation with amino acid properties, interaction propensity, distant residue impact, and XAI infidelity scores.

Main Results:

• Significant variation in the performance of different XAI methods was observed.
• Simple XAI methods demonstrated comparable effectiveness to advanced ones in extracting key information.
• Protein embeddings capture distinct properties, suggesting potential for enhanced embedding quality.

Conclusions:

• XAI is crucial for understanding protein embeddings and their role in predicting protein interactions.
• The choice of XAI method impacts the interpretability of protein embedding models.
• There is considerable scope for improving protein embedding quality and the insights derived from them.