The SHAP Explainer Model for Binary Classifiers Detecting Functional Groups in Molecules Based on FTIR Spectra

Area of Science:

• Computational Chemistry
• Spectroscopy
• Machine Learning

Background:

• Deep learning models, particularly Convolutional Neural Networks (CNNs), offer powerful capabilities but often function as "black boxes", making their decision-making processes difficult to interpret.
• This lack of transparency can limit the adoption of deep learning in regulated fields due to legal and ethical considerations, especially in scientific applications.
• Interpreting the basis of model decisions is crucial for validating their reliability and ensuring alignment with scientific understanding.

Purpose of the Study:

• To investigate and elucidate the decision-making process of a CNN-KAN (Convolutional Neural Network - Kernel Activation Network) model.
• To determine which specific regions within Fourier-Transform Infrared (FTIR) spectra are utilized by the CNN-KAN model for chemical group recognition.
• To assess whether the model's learned decision criteria align with established chemical principles found in scientific literature.

Main Methods:

• A CNN-KAN model was trained as a binary classifier to recognize chemical groups using FTIR spectral data.
• SHapley Additive exPlanations (SHAP) values were employed to analyze and interpret the model's predictions.
• SHAP values were used to identify the specific spectral regions influencing the model's classification decisions (positive or negative).

Main Results:

• The application of SHAP values successfully traced the decision-making pathway of the CNN-KAN model.
• Specific FTIR spectral regions were identified as critical drivers for both positive and negative classification outcomes.
• The identified spectral regions responsible for the model's decisions correspond to areas known in the literature to be significant for detecting particular functional groups.

Conclusions:

• The CNN-KAN model's decisions in recognizing chemical groups from FTIR spectra are based on interpretable spectral features.
• The model's reliance on scientifically validated spectral regions enhances its credibility and applicability in chemical analysis.
• This study demonstrates the utility of SHAP values in demystifying deep learning models in spectroscopy, bridging the gap between complex algorithms and scientific understanding.