Nutrient-response modeling with a single and interpretable artificial neuron

Area of Science:

• Nutritional Sciences
• Machine Learning
• Bioinformatics

Background:

• Classical nonlinear regression models are standard for estimating nutrient requirements but lack flexibility.
• Machine learning (ML) methods are often perceived as "black boxes", hindering biological interpretation in nutrition research.
• There is a need for interpretable ML approaches in nutrient-response modeling.

Purpose of the Study:

• To introduce a minimal and interpretable ML framework for nutrient-response modeling.
• To develop a method that overcomes the limitations of classical models while maintaining biological insight.
• To provide robust, uncertainty-quantified estimates of key nutritional metrics.

Main Methods:

• A single artificial neuron with hyperbolic tangent activation was employed, mathematically resembling a flexible four-parameter sigmoidal function.
• The framework incorporates modern ML best practices: data augmentation, Bayesian regularization, and bootstrap resampling.
• The approach was evaluated on 12 diverse datasets from poultry and fish studies, including amino acid and phosphorus responses.

Main Results:

• The single artificial neuron model demonstrated performance matching or exceeding classical models in nutrient-response modeling.
• The method provided robust, uncertainty-quantified estimates for metrics like asymptotic response, inflection point, and nutrient requirements.
• Full analytical transparency was maintained, addressing the "black box" concern of ML.

Conclusions:

• The proposed interpretable ML framework effectively models nutrient responses, offering greater flexibility than traditional methods.
• The 'NutriCurvist' application provides a user-friendly, no-code tool for precision nutrition.
• This approach supports data-driven decision-making in nutritional sciences by enhancing interpretability and robustness.