Ali Banjafar | Searcharxiv

arXiv2026

Property-Specific Molecular Representations via Feature-Space Transfer Compression

In many machine learning applications, molecules need to be transformed into representations, i.e. mathematical objects. Those representations are typically considered to be property-agnostic and as such are expected to be over-complete: for different physical properties, different parts or the representation may be relevant. In this work, we propose a method to sub-select and re-weight the representation by adapting it to the property in question. We find that in most cases this makes representations shorter and more accurate at the same time. The feature selection itself uses cheap semi-empirical data instead of high-quality labels. We study four properties (total energy, heat capacity, dipole moment, and polarizability) for three representations (cMBDF, FCHL19, and MACE-MP-0 descriptors) on two datasets (QM9 and VQM24). We can reduce the number of dimensions of a representation in the median by 72\,\% (range 36-98\,\%) while retaining the accuracy. Tuning for accuracy instead we can increase the learning efficiency for dipole moments such that the same accuracy can be reached with 19\,\% of the training data. Our approach yields data-driven interpretations of feature importance, lossless compact representations, and increased data efficiency, requiring only expendable surrogate data.

physics.chem-ph↗

arXiv2025

Intrinsic Dimensionality of Molecular Properties

Chemical space which encompasses all stable compounds is unfathomably large and its dimension scales linearly with the number of atoms considered. The success of machine learning methods suggests that many physical quantities exhibit substantial redundancy in that space, lowering their effective dimensionality. A low dimensionality is favorable for machine learning applications, as it reduces the required number of data points. It is unknown however, how far the dimensionality of physical properties can be reduced, how this depends on the exact physical property considered, and how accepting a model error can help further reducing the dimensionality. We show that accepting a modest, nearly negligible error leads to a drastic reduction in independent degrees of freedom. This applies to several properties such as the total energy and frontier orbital energies for a wide range of neutral molecules with up to 20 atoms. We provide a method to quantify an upper bound for the intrinsic dimensionality given a desired accuracy threshold by inclusion of all continuous variables in the molecular Hamiltonian including the nuclear charges. We find the intrinsic dimensionality to be remarkably stable across molecules, i.e. it is a property of the underlying physical quantity and the number of atoms rather than a property of an individual molecular configuration and therefore highly transferable between molecules. The results suggest that the feature space of state-of-the-art molecular representations can be compressed further, leaving room for more data efficient and transferable models.

physics.chem-ph↗