In this work, neural network models for molecular property prediction are studied, and an architecture for learning to rank is also presented. That is, using a Siamese network architecture, we develop a model for the learning-to-rank task.

For example, consider a scenario in which, given a list of molecular structures, it is necessary to determine the structure with the lowest energy. One possible approach is to use a neural network model that provides an energy estimate for each element in the given list; based on these estimated energy values, the list can then be ordered from lowest to highest energy.

On the other hand, we are not necessarily interested in the exact energy values themselves. Therefore, this problem can instead be addressed from the perspective of a ranking task. The objective is to identify the structure with the lowest energy, and thus a model trained explicitly to rank molecular structures is more appropriate in this context.

To evaluate the proposed methodology, the QM7-X and QO2Mol datasets are used, where we compare the performance of a model trained for ranking (pairwise) with that of a model trained using the traditional approach, known as pointwise. According to the results, the approach designed for learning to rank achieved superior performance.