Positional SHAP (PoSHAP) for Interpretation of machine learning models trained from biological sequences.

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning

Background:

• Deep learning models, particularly recurrent neural networks, excel at biological predictions from sequential data.
• Interpreting these complex models, especially for sequential inputs, remains a significant challenge.

Purpose of the Study:

• To introduce Positional SHAP (PoSHAP), a novel framework for interpreting machine learning models trained on biological sequences.
• To generate positional model interpretations using SHapely Additive exPlanations (SHAP).

Main Methods:

• Developed and applied the Positional SHAP (PoSHAP) framework.
• Utilized SHapely Additive exPlanations (SHAP) to derive positional interpretations.
• Tested PoSHAP on three long short-term memory (LSTM) regression models predicting peptide properties (MHC binding affinity, collisional cross section).

Main Results:

• PoSHAP successfully reproduced known peptide binding motifs for MHC class I molecules (Mamu-A1*001 and A*11:01).
• The framework accurately reflected established properties of peptide collisional cross section (CCS).
• New insights into interpositional dependencies of amino acid interactions within peptide sequences were uncovered.

Conclusions:

• Positional SHAP (PoSHAP) provides effective positional interpretations for models trained on biological sequences.
• The framework demonstrates broad utility for understanding various sequence-based biological prediction models.
• PoSHAP aids in deciphering complex relationships within biological sequence data.