Most simple regression models assume that the X variable is fixed by the experimenter and has no error associated with it, only the Y variable has random error, so all residual errors are vertical deviation from the trend line. As a result, the line is tilted to be ‘flatter’ than the main axis of the data.

In reality, the X variable often has just as much error/uncertainty as the Y variable, and the residuals should be perpendicular to the trend line.

One regression that allows for that is Major Axis Regression.

Another is to just take the main axis from PCA.