Answers are grounded in this paper and cite their sources.
The energies of the highest occupied and lowest unoccupied molecular orbitals (HOMO, LUMO) and their difference, the HOMO–LUMO gap, are a recurring evaluation target in molecular machine learning. These quantities are sensitive to molecular conformation, their errors can compound when derived indirectly, and they connect to interpretable physical mechanisms (ionization potential, electron affinity, conjugation, reactivity). This combination makes them a uniquely demanding test for representation-learning architectures. Other QM9 targets, such as internal energy or heat capacity, are thermodynamic properties dominated by additive atomic contributions, so invariant models fit them comparatively easily. Frontier orbital energies are different: they depend on long-range electronic delocalization and orbital symmetry, properties that an architecture without explicit equivariance or higher-order interaction terms is not guaranteed to capture. This is why we treat HOMO/LUMO/gap prediction as a focused case study rather than folding it into a generic molecular-property survey. We conducted a systematic review following PRISMA 2020 guidelines, searching Google Scholar, Scopus, and IEEE Xplore with the reproducible query "QM9" AND ("HOMO" OR "LUMO") AND "transformer" , then applied explicit inclusion and exclusion criteria requiring quantitative MAE reporting on QM9 (or its QM7/QM8 relatives) for at least one frontier-orbital-related target. The search returned 544 Google Scholar records, 11 Scopus records, and 0 eligible IEEE Xplore records (555 before deduplication). Of these, 59 studies published between 2022 and early 2026 met all criteria and were retained for structured data extraction across 49 fields covering architecture, physical priors, benchmark protocol, performance, computational cost, generalization, uncertainty quantification, and reproducibility. Pooled across the three target properties, equivariant architectures report a lower median MAE than invariant architectures (25.8 meV vs. 52.0 meV). However, the two distributions overlap substantially and both have long right tails, so we read this as a moderate, representation-dependent tendency rather than a categorical claim of superiority. Three-dimensional and hybrid 2D/3D representations dominate the corpus, reflecting a clear methodological preference, although 2D and 1D approaches remain competitive in a subset of studies; this comparison is also confounded by differences in pretraining data and model scale across studies. The corpus shows substantial, previously undocumented reporting gaps: algorithmic complexity is stated in only 35.6% of studies, inference latency in 23.7%, and predictive uncertainty in just 13.6%. We treat these gaps, rather than any single architectural finding, as the field’s most actionable shortcoming. We provide a consolidated comparative table classifying all 59 studies by representation, architecture family, physical prior, benchmark, and reported performance, and we discuss QM9’s known limitations (restriction to nine heavy atoms, closed-shell neutral species, single-conformer DFT geometries) as a boundary on how far every conclusion in this review can generalize.
No takes yet. Share an insight, caveat, or question.
Cruz et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: