An intercept is the constant term in a linear model. It represents the predicted output when every input feature is zero. In a solubility model using logP, the prediction is written as ŷ = w₀ + w₁x: w₀ is the intercept, and w₁ determines how the prediction changes with logP. section 4.1.4
For multiple features, the intercept can be represented as a bias parameter. A feature value of 1 is paired with that parameter, allowing the constant contribution to appear inside the same dot-product calculation as the feature contributions. section 4.2.1
The intercept serves several practical roles:
- Sets the model’s baseline prediction before feature effects are added. section 4.1.4
- Shifts a regression line without changing the feature coefficient’s slope. section 4.1.4
- Acts as a learnable bias vector in fully connected neural-network layers, where the affine transformation is xWᵀ + b. section 8.2.2
Solubility from logP
A simple solubility model uses logP as its only descriptor. After fitting the line to training data, the intercept w₀ supplies the predicted solubility at logP = 0, while w₁x supplies the adjustment for a molecule’s logP value. The model can be fitted with scikit-learn’s LinearRegression and evaluated using MSE, MAE, or R². section 4.1.4










