Tokenising quantum data for label-efficient training of molecular models
- 1Sirius Quantum Solutions Ltd.
Abstract
A model of molecular energies is limited by how much reference data can be afforded, because every training example is itself an expensive calculation. The practical question is therefore not “how accurate can this get” but “how much accuracy do I get per labelled molecule.” We built a representation of a molecule’s quantum features, which we call a quantum token and which a model consumes alongside whatever representation it already uses. The energy of a molecule in a given orbital basis is a linear functional of its one-and two-particle reduced density matrices, so those matrices, rather than the atomic coordinates, are the object the answer is actually computed from. A cheap calculation already produces an approximate one, and the token is a fixed-length encoding of it. It is not a token in the language-model sense: it is a vector of real numbers, not a discrete symbol drawn from a vocabulary, and the name refers to a unit of quantum information minted once and reused by many downstream models. No quantum hardware is involved: an ordinary classical model is trained on these features to predict the residual of the cheap calculation, the part a fast method gets wrong. Only the features are quantum. Holding the learner, the scaffold-disjoint split and the training-set size fixed and varying only the input features, adding the token to a production descriptor makes it 4.4–5.3× more accurate at equal training data, and a model given 50 labelled molecules is more accurate than the same model given 16,000 without it. At 20 to 52 heavy atoms, the size of drug candidates and the range where these models are applied in drug discovery rather than on public benchmarks, 93.7% of molecules fall within chemical accuracy, against 65.0% for the strongest classical descriptor and 4.4% for the uncorrected cheap method.
Related
- Extending AIMNet2 to Macrocyclic Peptides Through Data-Efficient Continual Training
- Noise Ceilings for Molecular Potency Benchmarks: Measuring Label Noise from Independent ChEMBL Documents
- Embedded Morgan Fingerprints for more efficient molecular property predictions with machine learning
- Efficient Light‐Driven Multi‐Electron Accumulation on Viologens by Molecular Control Over Photochemical Reaction Pathways
- Embedded Morgan Fingerprints for More Efficient Molecular Property Predictions With Machine Learning
BibTeX
@misc{karli2026tokenising,
title = {{Tokenising quantum data for label-efficient training of molecular models}},
author = {Derya Karli},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15009119/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009119/v1},
note = {CC BY-NC-ND},
}
RIS
TY - UNPB
TI - Tokenising quantum data for label-efficient training of molecular models
AU - Derya Karli
PY - 2026
DA - 2026/09/20
PB - ChemRxiv
DO - 10.26434/chemrxiv.15009119/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009119/v1
AB - A model of molecular energies is limited by how much reference data can be afforded, because every training example is itself an expensive calculation. The practical question is therefore not “how accurate can this get” but “how much accuracy do I get per labelled molecule.” We built a representation of a molecule’s quantum features, which we call a quantum token and which a model consumes alongside whatever representation it already uses. The energy of a molecule in a given orbital basis is a linear functional of its one-and two-particle reduced density matrices, so those matrices, rather than the atomic coordinates, are the object the answer is actually computed from. A cheap calculation already produces an approximate one, and the token is a fixed-length encoding of it. It is not a token in the language-model sense: it is a vector of real numbers, not a discrete symbol drawn from a vocabulary, and the name refers to a unit of quantum information minted once and reused by many downstream models. No quantum hardware is involved: an ordinary classical model is trained on these features to predict the residual of the cheap calculation, the part a fast method gets wrong. Only the features are quantum. Holding the learner, the scaffold-disjoint split and the training-set size fixed and varying only the input features, adding the token to a production descriptor makes it 4.4–5.3× more accurate at equal training data, and a model given 50 labelled molecules is more accurate than the same model given 16,000 without it. At 20 to 52 heavy atoms, the size of drug candidates and the range where these models are applied in drug discovery rather than on public benchmarks, 93.7% of molecules fall within chemical accuracy, against 65.0% for the strongest classical descriptor and 4.4% for the uncorrected cheap method.
ER -