BDH-Based Optical Rotation Prediction
Optical rotation is not a single intrinsic number attached to a molecular graph. The measured value depends on stereochemistry, conformation, wavelength, temperature, solvent, concentration, and the measurement convention.
A conditional property
A useful predictor therefore needs more than a structure input. Wavelength, temperature, solvent, concentration, and stereochemical assignment should travel with the estimate so two measurements are not compared as if they were made under identical conditions.
The reported quantity also depends on the measurement convention. For a common specific-rotation definition, [alpha] is related to the observed angle by the path length and concentration. A model that mixes path length, concentration, and specific rotation into one unlabeled target will learn a physically ambiguous number.
Conformer populations matter as well. A model can help prioritise plausible conformers and experimental conditions, but it should not turn conformational uncertainty into false precision.
What the model should return
The output is an estimate with its conditions and uncertainty attached. Under comparable conditions, enantiomers should reverse the sign of rotation; the magnitude remains dependent on the data, conformers, and measurement setup represented by the model.
That makes the model useful for screening and experiment planning, not a replacement for a polarimeter or a guarantee of absolute configuration.
Useful quality checks are symmetry-aware and condition-aware. Under comparable conditions, an enantiomer pair should show the opposite sign, while a racemic mixture may show a value close to zero even though each enantiomer is optically active. These checks help identify mislabeled or incomparable training measurements.
Where it fits in a real workflow
The practical role is to narrow which candidates and conditions deserve more expensive calculation or measurement. Every retained prediction should remain traceable to the input structure and the conditions that produced it.
The next step might be conformer refinement, a higher-level electronic-structure calculation, or a measurement under a specified wavelength, temperature, solvent, and concentration. The model is valuable when it makes that next experiment more selective, not when it hides the need for one.