# Pairwise Neural Networks for Ranking Molecular Structures Based on Properties

## TL;DR

This paper introduces a deep learning model that ranks molecular structures based on their properties, improving the efficiency of molecular discovery.

## Contribution

The novel contribution is a Siamese neural network using pairwise learning to rank molecular structures more effectively than traditional regression methods.

## Key findings

- The pairwise ranking model outperforms pointwise regression in predicting absolute energetic properties.
- Traditional regression remains better for derived or non-energy properties like HOMO–LUMO gap and dipole moment.
- The ranking model's performance is robust across different molecular representation models and parameter sizes.

## Abstract

The rapid discovery and design of new molecules drive
innovation
in science and technology, advancing energy storage, catalysis, and
drug development. Traditionally, exploring chemical space involves
costly quantum-chemical calculations or slow experimental screening,
which limits the speed of identifying promising candidates. Machine
learning has emerged as a groundbreaking approach to accelerate molecular
discovery by predicting key properties directly from molecular structures.
Moreover, in many cases, if we can rank molecular structures, it is
not necessary to know the exact value of a molecular property. In
other words, a ranker model can be useful for molecular screening.
In this work, we develop a deep learning model to rank molecular structures
using a siamese network approach and pairwise learning to learn the
ranking. According to different properties of the QM7x and QO2Mol
data sets, the results show that the performance of the learn-to-rank
Siamese architecture outperforms standard pointwise regression for
predicting absolute energetic properties, such as total and orbital
energies, while traditional pointwise regression remains effective
for derived (e.g., HOMO–LUMO gap) or nonenergy properties (e.g.,
dipole moment). To further validate the robustness of the proposed
framework, we extended our evaluation to include the Uni-Mol molecular
representation model. Experiments with Uni-Mol V1
and V2 across various model sizes (84 M to 1.1 B
parameters) confirm that the pairwise learning-to-rank objective consistently
outperforms standard pointwise regression, even when using highly
expressive pretrained Transformer backbones.