Chapter 23 Predicting Molecular Properties

The competition was hosted by a group of UK universities as a featured competition at https://www.kaggle.com/c/champs-scalar-coupling and was subtitled “Can you measure the magnetic interactions between a pair of atoms?”

  • Hosts:
    • CHemistry and Mathematics in Phase Space (CHAMPS) University of Bristol
    • Cardiff University
    • Imperial College
    • University of Leeds

The price money was $30,000, the competition ended 21.08.2019 and 2,749 teams had submitted a solution.

  • Challenge
    • Develop algorithm that can predict the magnetic interaction between two atoms in a molecule
  • Data
    • dipole moments
    • magnetic shielding tensor
    • mulliken charge
    • potential energy
  • Benefit
    • designing molecules to carry out specific cellular tasks
    • designing better drug molecules

The metric was Log of the Mean Absolute Error (MAE), calculated for each scalar coupling type, and then averaged across types, so that a 1% decrease in MAE for one type provides the same improvement in score as a 1% decrease for another type.

\[Log MAE = score=\frac{1}{T} \sum_{t=1}^{T} \log \left(\frac{1}{n_{t}} \sum_{i=1}^{n_{t}}\left|y_{i}-\hat{y}_{i}\right|\right)\]

Where:
  • \(T\) is the number of scalar coupling types
  • \(n_t\): is the number of observations of type \(t\)
  • \(y_i\): is the actual scalar coupling constant for the observation
  • \(\hat{y}_{i}\) is the predicted scalar coupling constant for the observation

The best possible score for perfect predictions is approximately -20.7232. The leader board look as follows:

  1. -3.23968
  2. -3.22349
  3. -3.19498

The winning solution is described in the next chapter